WBA: Word Boundary Attention for Chinese Named Entity Recognition Authors: Zhongguo Xu, Fu Lee Wang, Fu Lee Wang, Fu Lee Wang, Fu Lee Wang, Fu Lee Wang Published: 2025 ABSTRACT Chinese words often exhibit a parallel structural relationship within sentences, while individual characters are sequentially connected. To capture this structural distinction, we extract the head and tail positions of characters within words and incorporate them into a relative positional encoding scheme. Building upon this design, we introduce Word Boundary Attention (WBA), a mechanism that assigns dynamic attention weights to characters and enhances their representations with contextual information derived from the word lattice. By explicitly modeling word boundaries, WBA effectively suppresses noise, improves word recognition, and leverages richer lexicon-based context during training. Extensive experiments across multiple datasets demonstrate that WBA consistently outperforms existing approaches, achieving, for instance, a 2.51% improvement over the base model on the Weibo dataset with YJ lexicon encoding. Furthermore, visualizations of the learned attention weights reveal the interactive relationships between words and characters, providing interpretable insights into the process of word discovery. The source code of the proposed method is publicly available at https://github.com/na978292231/WBA/tree/main/WBA4NER-main. Introduction Named Entity Recognition (NER) is traditionally viewed as a task involving both entity boundary detection and entity type classification. However, it is worthwhile to consider entity boundary detection as a dedicated task [1]. Identifying the start and end positions of an entity mention in text is crucial for boundary detection, independent of type classification [2]. For enumerating all possible text spans within a sentence, entity boundary representations are typically formed by concatenating start/end token representations [3], with span width captured via a 20-dimensional embedding [4] and boundary diffusion [5], consistent with prior research. Various attention-based models have been explored for entity span detection. Yu et al. [6] employ a unified multimodal Transformer, while Zhang et al. [7] leverage an adaptive co-attention network. Additionally, boundary smoothing has been introduced as a technique that refines entity boundary detection rather than label prediction [8]. Several studies have also explored boundary representation as a fundamental aspect of language structure in modeling [9–12]. To enhance performance, external information sources have been incorporated, including lexicons [13], glyph features [14,15], syntactic structures [16], and semantic representations [17]. More recently, pre-trained language models such as BERT [18–20], GPT [21–28] have further boosted performance. In the context of Chinese NER, purely character-based methods [29–31] often fail to fully utilize word-level information. Chinese NER poses additional challenges due to its character-level structure (Fig 1a), which typically requires word segmentation and assumes equal spacing between adjacent characters. Since named entities in Chinese are composed of words, attending to word boundaries within a sentence is crucial. Chinese input with different structures. (a) Character-level structure. (b) Word-level structure. Actually, Chinese words are logically parallel in a sentence. In Chinese tasks, NLP tools build related lexicon to match the terms [32]. For instance, in Fig 1b, “南京市” (Nanjing City) forms a word-level structure, while “南京” (Nanjing) represents a subword-level structure. Both structures can function as named entities with a nested representation in NLP tasks. At the subword level, the distance between “南” (South) and “京” (Capital) is shorter than that between “京” (Capital) and “市” (City). At the word level, “京” (Capital) and “市” (City) are closer together than “市” (City) and “长” (Long). These observations indicate that word boundaries play a crucial role in defining entity boundaries. Therefore, an effective approach to Chinese NER should incorporate both word embeddings and word position information. Different segmentation styles correspond to different downstream tasks [33–35]. For instance, in Named Entity Recognition (NER), the boundary of an entity must be defined as “长江大桥” (Yangtze River Bridge), whereas in Neural Machine Translation (NMT), segmentation can separate “长江” (Yangtze River) and “大桥” (Bridge). To address these segmentation challenges, we introduce a word-level structure compilation approach for NER. We derive word start and end positions from segment labels in datasets or established tools such as Jieba, LTP, and CoreNLP. We further enhance attention models by incorporating word position information, leading to WBA, which utilizes a relative Transformer architecture. The main contributions include: To enhance word-level representation, we leverage segmentation tags to identify the head and tail positions of words, enabling knowledge discovery of word structures. Based on these extracted word boundary points, we propose a simple yet effective model built upon a relative Transformer architecture that explicitly attends to word boundaries within a sentence. To enhance word-level representation, we leverage segmentation tags to identify the head and tail positions of words, enabling knowledge discovery of word structures. Based on these extracted word boundary points, we propose a simple yet effective model built upon a relative Transformer architecture that explicitly attends to word boundaries within a sentence. We integrate word boundaries into character representations, introducing WBA for Chinese NER and helping the model automatically suppress noise and identify correct words. We integrate word boundaries into character representations, introducing WBA for Chinese NER and helping the model automatically suppress noise and identify correct words. We evaluate WBA on multiple benchmark datasets, demonstrating its superiority over state-of-the-art approaches. Our proposed method is transferable to different sequence-labeling architectures and can be easily adapted for other NLP tasks. We evaluate WBA on multiple benchmark datasets, demonstrating its superiority over state-of-the-art approaches. Our proposed method is transferable to different sequence-labeling architectures and can be easily adapted for other NLP tasks. Related work Integrating word information into character-based Named Entity Recognition (NER) systems has been a major research focus in recent years, often referred to as word augmentation. Due to the scarcity of annotated NER data, pre-trained language models such as BERT tend to underperform on certain NER tasks. This limitation is particularly evident in Chinese NER, where word-enhancement methods have been shown to outperform or closely match BERT’s performance, underscoring their necessity. A variety of approaches have been proposed to effectively incorporate word information into character-based models. Lexicon- and lattice-based approaches for Chinese NER. Chinese NER is challenging due to the lack of explicit word delimiters. Prior approaches have integrated lexicon information to alleviate segmentation ambiguity, using lexicon-enhanced embeddings [13], lattice structures [36], or hybrid character–word models [37]. Adaptive word embedding methods such as WC-LSTM [38], NMDM [39], and CAN-NER [40] incorporate vocabulary information in a model-independent, transferable manner. Other techniques, including SoftLexicon [41] and ConvTransformer [42], exploit lattice structures and convolutional layers to capture both word-level and subword-level contexts. SubRegWeigh [43] further introduces effective and efficient annotation reweighting with subword regularization. While lattice-based models allow multiple word candidates to be represented simultaneously, they often introduce noise and increase computational complexity. To address this, recent studies propose dynamic weighting of lexicon information and refined character–word interactions, including Lattice LSTM [13], LR-CNN [44], LGN [45], and CGN [46], which explicitly model word structures. Our method differs by mining explicit head–tail boundary points of words and incorporating them into attention, preserving the benefits of lexicon-aware modeling while reducing noise from irrelevant candidates. Transformer-based models and positional encoding. The Transformer architecture has become the foundation of modern sequence modeling, with self-attention mechanisms enabling efficient context capture across tokens [47]. Numerous variants extend the Transformer by modifying positional encodings to better capture structural or relational information [48,49]. Relative positional encodings improve the modeling of long-range dependencies, while boundary-aware mechanisms have been applied to parsing and segmentation tasks [50]. Transformer-based models such as Lattice Transformer [51], FLAT [37], NFLAT [52], and pre-trained models like ZEN [53], LEBERT [36], and BERT-wwm [54] incorporate lexicon-aware self-attention to boost performance. Our work builds on this line of research by explicitly encoding word boundary positions, which differentiates it from standard self-attention and most positional encoding methods that treat words or characters uniformly. Noise suppression and interpretability in NER. Beyond structural modeling, researchers have investigated noise suppression in NER datasets through label denoising [55], confidence-based reweighting [56], and distributional analysis [57]. Visualization and interpretability methods have also been developed to analyze attention mechanisms and their relationship to linguistic structures [58]. Our work contributes to this line of research by demonstrating that explicit boundary modeling not only improves performance but also yields interpretable attention maps that reveal interactions between words and characters in Chinese sentences. Our study is particularly inspired by FLAT and NFLAT, which significantly advance Chinese NER by incorporating lexicon information through lattice structures with dual positional encodings in relative attention. However, these models do not explicitly capture word boundaries, which can lead to errors and inefficiencies in character-level enhancements. To overcome this limitation, we propose a word boundary enhancement method that leverages relative attention to explicitly attend to word boundaries, thereby achieving superior performance compared to existing approaches. Background NFLAT introduces the InterFormer method, which significantly simplifies the attention matrix used in FLAT. Additionally, the context feature encoding from TENER enables the model to effectively attend to character-level positions. Together, these two enhancements contribute to achieving relatively high performance in Chinese NER. As illustrated in Fig 2a, NFLAT streamlines FLAT’s approach by computing attention scores solely for ‘character-word’ interactions. Additionally, it introduces the ‘’ tag for single characters that do not match the lexicon or correspond to punctuation marks, setting their attention values to zero. This prevents unnecessary attention allocation to non-relevant words. The model then employs character-based TENER to compute ‘character-character’ attention scores using relative position encoding, incorporating both orientation and distance awareness. By decoupling lexicon fusion from context feature encoding, NFLAT achieves improved accuracy and efficiency over FLAT. NFLAT and its attention matrix. (a) The self-attention matrix of NFLAT from FLAT (b) The self-attention matrix of NFLAT to WBA. As shown in Fig 2b, NFLAT mitigates issues related to blurred word boundaries and the absence of word semantics. However, for richer and more precise word representations, we argue that computing ‘character-character’ interactions within word boundary attention is essential. We use a novel word boundary guided relative position encoding, rather than lattice-based fusion. To address this, we introduce an additional step between word-level embedding and the TENER encoder attention. Our proposed Word Boundary Attention (WBA) explicitly attends to word boundaries, further enhancing both accuracy and efficiency beyond NFLAT. Methods For Chinese Named Entity Recognition (NER), WBA consists of four main stages. In the first stage, word-level embeddings are applied to integrate semantic information from words. The second stage introduces WordFormer, which captures word boundary information within the sentence. In the third stage, a Transformer Encoder is employed to encode contextual information enriched with lexicon features. Finally, a Conditional Random Field (CRF) [59] is used as the decoder to predict the sequence labels. The WordFormer module is a key contribution of this work. An overview of the WBA architecture is illustrated in Fig 3. The overall architecture of WBA. Word-level embedding In this work, we implement this layer with different models, e.g., base model (NFLAT [52]), pre-trained model ( BERT-wwm [54] and ZEN [53]). For base model, we use unigram and bigram embeddings in static character embedding layer. Each character ct is embedded as xct. Each word wt is embedded as xwt. The character embedding encoder represents the sentence based on the character sequence c1,c2,…,cn and the word embedding encoder represents the sentence based on word sequence w1,w2,…,wm. where ec denotes a character embedding lookup table. where ew denotes a word embedding lookup table. Word-level embedding aims to construct a non-flat-lattice and jointly model two sequences of characters and words with different lengths. It enables the sequence of characters to fuse word boundaries and semantic information. This part is inspired by NFLAT. The inputs Q, K, and V are derived through linear transformations of character and word feature embeddings: where X𝒞={xc1,xc2,...,xcn} and X𝒲={xw1,xw2,...,xwm} denote the token embeddings for the character and word sequences, respectively. Each of the projection matrices Wq, Wk, and Wv is a learnable parameter. To incorporate lexicon information, we employ an inter-attention mechanism that fuses character and word representations and generates the hidden states hci: where mask() is the inter-attention score mask of characters and words, which is 2-dimensional for a single batch and 3-dimensional for a multi-batch. It is used to fill empty positions in the sequence with the value of 10−15, so that the attention weights of these positions are close to 0 when softmax() normalization. The input attention is calculated in a similar way to NFLAT, using two relative positions. For pre-trained models, the character sequence are put into the pre-trained modeling layer to obtain hidden states hci, which models the word-level information between characters. For example, in Fig 3, the character sequence [‘南’ (South), ‘京’ (Capital), ‘市’ (City), ‘长’ (Long), ‘江’ (River), ‘大’ (Big), ‘桥’ (Bridge)] is encoded to generate character hidden states hci. Meanwhile, the lexicon sequence [‘’, ‘南京’ (Nanjing), ..., ‘大桥’ (Bridge)] is encoded to obtain lexicon embeddings xwt, which serve as word-level embeddings in the base model. WordFormer attention After word-level embedding, each character has the information of lattice. We use WordFormer attention to enable the sequence of characters to fuse word boundaries in a sentence, which is absent in vanilla Transformers. Unlike the standard Transformer, which applies self-attention uniformly across tokens, WBA introduces explicit head and tail nodes to represent word boundaries. This modification allows the model to capture word-level structural information, which is critical for Chinese where word segmentation is ambiguous. Furthermore, WBA restricts and reweights attention based on boundary positions, thereby reducing noise from irrelevant characters. These innovations distinguish WBA from standard attention and contribute directly to improved NER performance. The input Q, K, V are obtained by the linear transformation of rich chars feature hidden states from word-level embedding encode: where the hidden states of character sequences, h𝒞={hc1,hc2,...,hcn}. Each of the projection matrices Wq, Wk, and Wv is a learnable parameter. We use word segment tags or tools to split the sentence into words. Then we estimate word position probabilities while handling ambiguous word boundaries to reduce systematic errors and biases into the model. It begins by enumerating all possible word boundary candidates. Next, it applies character-lexicon attention, using word match positions to compute the probability of each character appearing at the beginning (B), middle (M), end (E), or as a single-character word (S). Finally, the framework integrates masking and class attention mechanisms to generate informative position features for characters, leveraging both character and lexicon contexts. For word-level structure, we set head and tail nodes of words. Each character ci has position pi, head hi and tail ti nodes. Each word is separated in the sentence. We consider each character in the position of its word, using two relative distances like NFLAT: where dijph denotes the distance between position of pi and head of hj, and dijpt denotes the distance between position of pi and tail of tj. Fig 3 shows examples of dijph and dijph in a more intuitive form. The final relative position encoding of words is a simple non-linear transformation of the two distances: where Wr is a learnable parameter, ⊕ denotes the concatenation operator, and 𝐩d is calculated as in Transformer [60], where d is dijph, dijpt, k denotes the index of dimension of position encoding and dmodel is the hidden size. We then apply a variant of self-attention to incorporate relative word position encoding, as illustrated in Fig 3. where 1≤i≤n, 1≤j≤m. The u, v are learnable parameters. Eq (11) is from Dai et al. [61]. The input attention is computed similarly to NFLAT, utilizing two relative positions as defined in Eq (11) and Eq (4). As illustrated in Fig 3, the attention heads highlight the word boundary between “南京市” (Nanjing City) and “长江大桥” (Yangtze River Bridge). The characters [‘南’ (South), ‘京’ (Capital), ‘市’ (City)] share the same head and tail positions within the word segment “南京市” (Nanjing City) and primarily focus on each other while paying comparatively less attention to the other characters in the sequence. Transformer encoder After the WordFormer module, character features are enriched with both lexicon information and word boundary signals. Next, a Transformer encoder is employed to capture contextual dependencies. This component is inspired by the work of Yan et al. [62], which demonstrated that unscaled self-attention and relative position encoding with directional information are particularly effective for NER tasks. Finally, a Conditional Random Field (CRF) decoder is applied to predict the most likely label sequence by maximizing the overall sequence scores. Experiments We evaluate the proposed WBA using standard metrics—F1 score (F1), precision (P), and recall (R)—and compare its performance with several character-word hybrid models. Additionally, we conduct a visual analysis of the word boundary attention weights to verify the effectiveness of the proposed mechanism. Finally, we present a comparison of results with and without the use of pre-trained language models. Experimental settings Data. The model is evaluated on four Chinese NER datasets, including Weibo [63,64], Resume [13], OntoNotes 4.0 [65] and MSRA [66]. Weibo is a social media domain dataset, which is drawn from Sina Weibo, while OntoNotes 4.0 and MSRA datasets are in the news domain. Resume dataset consists of resumes of senior executives, which is annotated by Zhang et al. [13]. Table 1 shows the statistical information of these datasets. 10.1371/journal.pone.0336942.t001Table 1Statistics of the benchmarking datasets.DatasetsItemsTrainDevTestWeiboSentences1.35k0.27k0.27kEntities1.89k0.39k0.42kResumeSentences3.8k0.46k0.48kEntities1.34k0.16k0.16kMSRASentences46.4k-4.4kEntities74.8k-6.2kOntoNotes 4.0Sentences15.7k4.3k4.3kEntities13.4k6.95k7.7k Statistics of the benchmarking datasets. Embedding. For base model, we download the specified pre-trained unigram and bigram embeddings for Chinese task. We conduct experiments on YJ [67] and LS [68] lexicons. Table 2 shows the statistics of the external lexicons used in this paper, mainly including their total words, single-character words, two-character words, three-character words, and words with more characters. Also, we list the dimensions of the word vectors in lexicon. Then we use relative attention incorporate lexicon in character embedding. For the pre-trained model, we use Chinese-BERT-wwm [54] and ZEN for character embedding. 10.1371/journal.pone.0336942.t002Table 2Statistics of the lexicons.The Single, Two, and Three represent the number of characters in the word, respectively. The Other means the number of words with more characters. The Dimension is the dimension of the word embedding.LexiconsTotalSingleTwoThreeOtherDimensionYJ704.4k5.7k291.5k278.1k129.1k50dLS1292.6k18.5k347.7k415.6k511.3k300d Statistics of the lexicons. The Single, Two, and Three represent the number of characters in the word, respectively. The Other means the number of words with more characters. The Dimension is the dimension of the word embedding. Hyper-parameter settings. For base model, this paper uses only one layer of word-level embedding, one layer of WordFormer encoder and one layer of Transformer encoder to handle Chinese NER. We adopt similar settings as NFLAT [52]. Most implementation details include character and word embedding sizes of 50, a dropout rate of 0.2, and a batch size of 10. Additionally, the hidden size is set to 192 with 12 heads for small datasets Weibo and Resume, and 256 with 8 heads for larger datasets OntoNotes and MSRA. The initial learning rate is set to 0.003 for Weibo and 0.002 for the remaining three datasets with SGD step rule. We train each model on training sets with 100 epochs totally. For pre-trained model, most implementation details followed those of BERT-NER, including character and word embedding sizes with 786, dropout with 0.15, learning rate with 2e-5 in Adam step rule, batch size with 16. We train each model on training sets with 100 epochs totally. Experimental results We conduct experiments on four datasets to further evaluate the effectiveness of our model in combination with a pre-trained model. The experiments were run on a physical machine equipped with CentOS Linux 7.9.2009 (Core), three Intel(R) Xeon(R) Gold 6226R CPUs, and two NVIDIA TITAN RTX GPUs. Table 3 presents the results on the Weibo, Resume, OntoNotes, and MSRA datasets, comparing our approach against state-of-the-art statistical models that leverage rich word lexicon features, including Lattice-LSTM [13], WC-LSTM [38], LR-CNN [44], LGN [45], PLTE [69], SoftLexicon [41], FLAT [37], and NFLAT [52]. We use NFLAT as the baseline model. For pre-trained models, we compare with BERT-wwm [54] and ZEN [53] across all four datasets. 10.1371/journal.pone.0336942.t003Table 3A comparison with other lexical enhancement methods.YJ and LS lexicon is used here. Bold indicates leading all baselines with p < 0.05.ModelsWeiboResumeOntonotes 4.0MSRANENMOverallPRF1PRF1PRF1Lattice-LSTM53.0462.2558.7994.8194.1194.4676.3571.5673.8893.5792.7993.18WC-LSTM53.1967.4159.8495.2795.1595.2176.0972.8574.4394.5892.9193.74LR-CNN57.1466.6759.9295.3794.8495.1176.4072.6074.4594.5092.9393.71LGN55.3464.9860.2195.2895.4695.3776.1373.6874.8994.1992.7393.46PLTE53.5564.9059.7695.3495.4695.4076.7872.5474.6094.2592.3093.26SoftLexicon59.0862.2261.4295.7195.7795.7477.1375.2276.1694.7393.4094.06TENER--58.8294.7994.9794.8875.9777.2976.6392.9791.9692.46FLAT(YJ)--60.3294.9895.2195.1076.7577.9377.3494.4393.8594.14NFLAT(YJ)59.1063.7661.2295.5594.8595.20 75.69 79.0977.3594.7693.5394.14NFLAT(YJ)+WBA 59.76 65.10 63.73 96.06 95.77 95.91 75.67 79.58 77.57 94.86 93.67 94.26 NFLAT(LS)59.5464.8263.1394.4095.6695.0273.6877.5575.5693.6992.8393.26NFLAT(LS)+WBA 59.62 64.95 63.43 94.84 95.79 95.31 74.08 79.02 76.47 95.36 94.10 94.73 BERT-wwm71.6459.6370.6696.5495.9596.2578.4780.5479.4994.7994.9794.88BERT-wwm+WBA 72.39 70.77 71.26 96.64 96.73 96.68 78.51 81.09 79.78 95.21 95.03 95.12 ZEN64.8367.6566.7195.3495.4695.4078.0780.0479.0395.3995.0795.25ZEN+WBA 65.28 68.21 67.65 95.42 96.27 95.85 78.11 80.19 79.18 95.41 95.23 95.32 A comparison with other lexical enhancement methods. YJ and LS lexicon is used here. Bold indicates leading all baselines with p < 0.05. Our results show that NFLAT+WBA with the YJ lexicon improves upon NFLAT. Specifically, the overall F1 score on Weibo increases by 2.51%, as the dataset’s lower baseline F1 leaves more room for improvement. The Weibo NER dataset, sourced from the social media platform Sina Weibo, contains informal and colloquial language, leading to blurred word boundaries. By incorporating attention mechanisms that explicitly capture word boundaries between characters, our approach proves especially beneficial for Chinese NER. On the Resume dataset, the F1 score improves by 0.71%, while OntoNotes 4.0 sees a 0.22% increase. Both datasets contain structural information and shallow semantics, making character boundaries more easily identifiable. For the MSRA dataset, our method yields a smaller improvement of 0.12%, possibly due to differences in word segmentation, such as the phrase “中外” (at home and abroad) being split into “中” (Chinese, labeled as “S-NS”) and “外” (abroad, labeled as “O”), which may hinder entity recognition. Overall, NFLAT+WBA outperforms all other methods, including FLAT, across the four datasets. Without additional data augmentation or pre-trained language models, NFLAT+WBA achieves state-of-the-art performance on Weibo, Resume, and OntoNotes 4.0, particularly excelling in smaller datasets. Furthermore, we evaluate NFLAT+WBA using larger lexicons (LS) and compare its performance with other models. As lexicon size increases, NFLAT+WBA improves, demonstrating that a larger and richer lexicon benefits WBA. Additionally, WordFormer can be extracted from WBA as a standalone module, and WBA can seamlessly integrate pre-trained models into the embedding layer. As shown in Table 3, BERT-wwm+WBA and ZEN+WBA further enhances the performance of the pre-trained BERT-wwm and ZEN model respectively. Convergent speed comparison We compare the convergence speed of NFLAT+WBA and NFLAT with the YJ lexicon across four datasets. Fig 4 presents the test F1 scores per epoch over 100 epochs. The results show that NFLAT+WBA consistently outperforms NFLAT in convergence speed across all four datasets. Notably, the improvement on the Weibo dataset is particularly significant, demonstrating the efficiency of our method in handling complex and informal text. Convergent speed comparison between NFLAT+WBA and NFLAT. Blue points denote NFLAT+WBA and green points denote NFLAT, respectively. The x-axis and y-axis denotes test F1 and steps, respectively. (a) Weibo comparison. (b) Resume comparison. (c) OntoNotes comparison. (d) MSRA comparison. Complexity analysis WBA consists of three main components: (1) inter-attention between characters and words with a complexity of 𝒪(nmd), (2) WordFormer attention between characters with complexity 𝒪(n2d), and (3) self-attention over the contextual sequence with complexity 𝒪(n2d). Thus, the overall computational complexity of WBA is 𝒪((2n+m)nd). The corresponding space complexities are 𝒪(nm) for inter-attention, 𝒪(n2) for WordFormer attention, and 𝒪(n2) for self-attention. We argue that WBA is particularly well-suited for Chinese NER, as demonstrated by its performance in terms of model size, time cost, memory usage, and inference time, summarized in Table 4. The reported time cost corresponds to the average per epoch on the Weibo dataset, and the model size reflects the hyperparameter configuration used for the same dataset. The memory usage is measured on the Weibo test set with a batch size of one, and the inference time represents the total accumulated duration of all test samples during evaluation on Weibo test set and its extended version. 10.1371/journal.pone.0336942.t004Table 4Comparison of model size, time cost, memory usage and inference time.Models#Parameters(M)Time cost(s)Memory Usage (MB)Inference Time (s)TENER0.274.669.893.75/12.37FLAT(YJ)0.5924.2413.973.52/12.96NFLAT(YJ)0.8618.4514.893.83/12.53NFLAT(YJ)+WBA1.4519.5217.874.02/13.03 Comparison of model size, time cost, memory usage and inference time. From the comparison, TENER exhibits the smallest parameter size (0.27M) and the highest computational efficiency, with a training time of only 4.66s, memory usage of 9.89MB, and inference time of 3.75s/12.37s. However, due to its lack of lexical-level interactions, it struggles to capture complex word boundary features inherent in Chinese text. In contrast, FLAT(YJ) introduces a lexicon-lattice structure, which significantly enhances its ability to incorporate external word information. As a result, its parameter size increases to 0.59M, training time extends to 24.24s, and memory usage rises to 13.97MB. Despite the higher computational cost, its inference time remains relatively low (3.52s/12.96s), indicating that FLAT maintains strong decoding efficiency even with more complex structural modeling. NFLAT(YJ) optimizes the integration between characters and words through a non-lattice structure, effectively reducing redundant path computations. Its training time decreases to 18.45s, memory usage is 14.89MB, parameter count reaches 0.86M, and inference time is 3.83s/12.53s. These results demonstrate that NFLAT achieves a better balance between model complexity and computational efficiency. Building upon this, NFLAT(YJ)+WBA incorporates the word boundary attention mechanism, enabling the model to adaptively capture implicit boundary information. Although its parameter size increases to 1.45M and memory usage slightly rises to 17.87MB, its training time (19.52s) and inference time (4.02s/13.03s) remain comparable to NFLAT(YJ). This indicates that WBA enhances boundary awareness and semantic discrimination without significantly increasing computational overhead. This conclusion is further supported by the results shown in Figs 5 and 6. Fig 5 is based on samples from the Weibo test set, while Fig 6 presents an extended version of the same dataset. Together, they illustrate the inference time trends of TENER (yellow), FLAT (blue), NFLAT (green), and WBA (red) models across different sentence lengths. Overall, the inference time of all four models increases linearly with sentence length, indicating that their computational complexity is primarily driven by input length rather than nonlinear operations within the model structure. A comparison between TENER, FLAT, NFLAT and NFLAT+WBA in terms of inference time, evaluated at different sentence lengths on Weibo test set. A comparison of TENER, FLAT, NFLAT, and NFLAT+WBA in terms of inference time, evaluated across different sentence lengths on the extended version of the Weibo test set. In the short-to-medium length range (10–150 tokens), the differences among the models are more pronounced. TENER, due to its simple architecture, consistently achieves the lowest latency across all ranges; FLAT and NFLAT show comparable performance, with NFLAT performing slightly better for medium-length sentences; NFLAT+WBA shows slightly higher inference time than the others but maintains a stable trend without exponential growth, suggesting that incorporating word boundary information does not introduce significant computational overhead. When sentence length extends to a longer range (up to 800 tokens), all four models exhibit an almost perfectly linear relationship between inference time and input length. Although all models demonstrate stable scalability for long sequences, the fitted curves reveal that NFLAT+WBA incurs slightly higher computational cost than FLAT and NFLAT, while TENER continues to deliver the best efficiency. This finding indicates that WBA, while enhancing contextual modeling capability, maintains near-linear inference complexity, effectively balancing efficiency and representational power—making it a practical and efficient choice for long-text named entity recognition tasks. Fig 7 presents the GPU memory usage of TENER, FLAT, NFLAT, and NFLAT+WBA during training on the extended Weibo test set. By setting the training batch size to 1, we obtain the comparison of memory consumption across different sentence lengths. As the sentence length increases, all models show a clear upward trend in memory usage. Among them, TENER consistently maintains the lowest memory consumption, reflecting its lightweight architecture. NFLAT requires more memory than TENER but significantly less than FLAT and NFLAT+WBA, especially for longer sentences. FLAT shows slightly higher memory usage than TENER when the sentence length is below 200. As the sentence length increases, the corresponding lexicon length also grows, leading to a rapid rise in memory consumption. When the sentence length exceeds 300, FLAT surpasses NFLAT in memory usage, and when it exceeds 500, it also surpasses NFLAT+WBA. Due to the introduction of the Word Boundary Attention (WBA) mechanism, NFLAT+WBA processes additional word boundary nodes, resulting in the highest memory consumption when the sentence length is below 500. Overall, these results indicate that WBA introduces additional computational overhead, yet it still demonstrates better scalability than FLAT when handling long sequences. Comparison of memory consumption during training for TENER, FLAT, NFLAT, and NFLAT+WBA, conducted on the extended Weibo test set with varying sentence lengths. By jointly analyzing the table and figure, we can conclude that WBA effectively enhances boundary detection and semantic modeling while keeping computational complexity nearly unchanged. Compared with FLAT, NFLAT+WBA achieves a better balance across model size, training cost, and inference efficiency, demonstrating both structural compactness and practical feasibility. These results provide strong evidence that WBA is an efficient and robust enhancement mechanism for Chinese NER, maintaining linear scalability while significantly improving modeling performance. Cross-dataset validation To demonstrate the generalizability of the proposed method,we conduct additional experiments using the CoNLL-2003 [70], Wiki-ja, and Konec dataset. In the processing of the CoNLL-2003 dataset, each line of data consists of a token and its corresponding entity label. The entity labels use the BIO annotation scheme, where B- represents the beginning of an entity, I- represents the inside of an entity, and O represents non-entity tokens. To add head and tail nodes for each line, we process it based on the syntactic component label in the third column. Each B- labeled token is considered the head of a new entity block, and subsequent I- tokens share the same tail node. Tokens labeled as O are treated as individual blocks with the same head and tail node number. When an empty line is encountered, the numbering is reset to 0, and a new block numbering starts. In the experimental setup, the pre-trained embedding is GloVe 100d, the learning rate is set to 9e-4 and the batch size is set to 100 based on the parameters of the Weibo dataset. In the processing of the Wiki-ja dataset, we first perform Japanese tokenization using tools such as MeCab to segment the text into words. If the word is not a number, it will be split into characters and re-labeled to fit the different entity block requirements. For multi-character words like “东京大学” (Tokyo University), the entire word is treated as a block, and the head and tail nodes for each character in the word are assigned the same value. The first character of the entity is labeled with B-, and the remaining characters are labeled with I-, while non-entity characters are labeled with O. When an empty line is encountered, the numbering is reset, and the head and tail nodes are adjusted to ensure that each sentence is independently numbered. In the experimental setup, the pre-trained embedding is FastText 300d, the learning rate is set to 5e-3 based on the parameters of the Weibo dataset. In the processing of the Konec dataset, head and tail nodes are assigned as follows: when encountering a line with an empty character column, we skip the line and continue with the next numbering. When a line contains punctuation marks, it is treated as an independent block. Additionally, when encountering an empty line, the numbering is reset. In the experimental setup, the pre-trained embedding is FastText 300d, the learning rate is set to 3e-3 and the batch size is set to 50 based on the parameters of the Weibo dataset. The results are shown in Table 5. On the English dataset CoNLL-2003, the introduction of WBA significantly improved the overall model performance. For example, in the case of TENER, the F1 score increased from 83.31 to 85.12, achieving a gain of +1.81. This demonstrates that word boundary augmentation effectively enhances the model’s ability to capture entity boundaries. Since English words are naturally separated by spaces and exhibit stable morphological structures, WBA strengthens word-level information modeling without introducing additional noise, allowing the model to perform more precisely in boundary recognition and entity classification. Moreover, WBA further improves the model’s recall (R), indicating that it not only captures entity boundaries more accurately but also reduces omissions, thereby enhancing the model’s overall generalization across sentences. This result verifies that WBA provides a stable and transferable improvement in languages with clear morphological rules. 10.1371/journal.pone.0336942.t005Table 5Cross-dataset validation.ModelsCoNLL-2003Wiki-jaKonecPRF1PRF1PRF1TENER85.4681.2783.3169.6567.8168.7270.3167.8469.05TENER+WBA 87.58 82.79 85.12 79.14 76.93 78.02 79.8274.9077.25NFLAT---75.3771.5973.4382.6180.2781.42NFLAT+WBA---73.7469.8971.75 83.90 81.54 82.71 Cross-dataset validation. On the Wiki-ja (Japanese) dataset, the performance variation among models presents a more complex pattern. First, the TENER model shows a substantial improvement after applying WBA—the F1 score rises from 68.72 to 78.02, a gain of +9.30—indicating that word boundary augmentation plays a strong compensatory role in character-level Transformer frameworks. Because Japanese text lacks explicit word delimiters, word boundaries must be inferred through particles, verb endings, and semantic cues, making it difficult for traditional character-level models to accurately capture entity boundaries. The additional boundary signals introduced by WBA effectively guide the model to learn word-level structural patterns, thereby improving both precision and recall in entity recognition. However, when transitioning from TENER to NFLAT, overall performance declines. NFLAT’s F1 score drops from 73.43 to 71.75, and even with WBA, it fails to surpass the performance of TENER+WBA. This downward trend reveals two underlying causes. First, NFLAT is designed to integrate multi-granularity lexical information through a non-lattice structure. In Japanese, this mechanism can introduce granularity conflicts—the model simultaneously learns representations at the character, word-piece, and subword levels, causing competition among boundary signals and reducing feature consistency. Second, WBA may generate redundant boundary information within the NFLAT framework, overlapping with its internal multi-granularity attention mechanism and reducing training stability. Furthermore, the morphological complexity and boundary ambiguity of Japanese exacerbate these issues, making the augmented boundaries semantically unreliable and introducing boundary noise, which ultimately leads to performance degradation. In summary, WBA exhibits a phenomenon of “strong enhancement for character-level models but limited compatibility with multi-granularity models” in Japanese tasks. This suggests that the effectiveness of WBA depends not only on the predictability of linguistic boundaries but also on the model’s capacity to absorb and utilize word-level information. Future work may focus on language-adaptive boundary control strategies or granularity-level constraints to improve stability and generalization in morphologically complex languages such as Japanese. On the Konec (Korean) dataset, WBA also yields notable performance gains. The F1 score of TENER increases from 69.05 to 77.25, a gain of +8.20, while NFLAT’s performance improves from 81.42 to 82.71, showing stable results. Although Korean uses a phonetic writing system, its morphological structure is regular and its word boundaries are relatively well-defined. Therefore, WBA can guide the model to learn clearer word-boundary representations without introducing noise, strengthening the identification of entity boundaries. Especially in the recognition of long-span or compound entities, the boundary cues from WBA help the model distinguish entity from non-entity spans, significantly improving both recall and overall F1. This indicates that in languages with relatively consistent boundary rules, WBA effectively enhances cross-lingual transferability and generalization. In summary, WBA demonstrates distinct language-dependent characteristics across datasets: it consistently improves performance on CoNLL-2003 (English) and Konec (Korean), but shows model-dependent variation on Wiki-ja (Japanese). Specifically, WBA provides strong enhancement for character-level architectures such as TENER, while in models that already incorporate multi-granularity lexical relationships (e.g., NFLAT), it may introduce redundancy and granularity conflicts, leading to decreased performance. These findings suggest that the cross-lingual generalization effect of WBA is influenced by both linguistic boundary properties and model architecture. Future research could explore language-adaptive and hierarchical integration mechanisms to further improve stability and robustness in morphologically complex languages. Discussion Ablation study To investigate the contribution of each component of our method, we conduct ablation experiments with YJ lexicon on all four datasets, as shown in Table 6. 10.1371/journal.pone.0336942.t006Table 6Ablation experimental results.ModelsWeiboResumeOntonotesMSRATENER58.1795.0072.4392.74FLAT60.3295.4576.4594.12NFLAT61.2295.2077.3594.14NFLAT+WBA 63.73 95.91 77.57 94.26 w/o Word-level Embedding62.3895.1376.8393.62 w/o Transformer Encoder61.4495.0276.6793.84 w/o Word-level Embedding&Transformer Encoder61.0994.8976.1092.61 w/o WordFormer Attention61.2095.1977.3094.11 w/o WordFormer Attention&Transformer Encoder60.9194.3676.0892.29 w/o WordFormer Attention &Word-level Embedding58.0294.2071.4291.98 Ablation experimental results. In the w/o Word-level Embedding experiment, we remove the word-level embedding from NFLAT+WBA. The performance drop across all four datasets highlights the importance of lexicon fusion and contextual feature encoding. However, on Weibo, the F1 score remains 1.16% higher than that of NFLAT, suggesting that WordFormer provides stronger benefits than word-level embedding in small datasets. Moreover, the consistent outperformance over TENER indicates that WordFormer effectively captures word boundaries for improved Chinese NER. In the w/o Transformer Encoder experiment, we remove the Transformer encoder from NFLAT+WBA. Compared to the word-level embedding ablation, the larger performance degradation across datasets shows that character relative position modeling is more critical than lexicon fusion. Still, the F1 score on Weibo is 0.22% higher than NFLAT, again demonstrating WordFormer’s effectiveness in small datasets. In the w/o Word-level Embedding & Transformer Encoder experiment, both components are removed. The further performance decline across datasets confirms the complementary roles of lexicon fusion and character position modeling. Nevertheless, the F1 score on Weibo is 0.77% higher than FLAT, showing that WordFormer surpasses FLAT attention on small datasets. In the w/o WordFormer Attention experiment, we ablate WordFormer and retain NFLAT with self-attention only. The performance drop across datasets emphasizes the significance of explicit word boundary modeling. Despite this, the F1 score on Weibo is still 0.92% higher than FLAT, suggesting that NFLAT inter-attention remains stronger than FLAT attention in small datasets. In the w/o WordFormer Attention & Transformer Encoder experiment, we remove WordFormer and rely on inter-attention with self-attention. The resulting degradation underscores the combined importance of word boundaries and character relative position modeling. Yet, the F1 score on Weibo is still 0.59% higher than FLAT, again validating the strength of NFLAT inter-attention. Finally, in the w/o WordFormer Attention & Word-level Embedding experiment, WordFormer is removed while keeping self-attention and TENER. The performance drop further demonstrates the necessity of explicit boundary modeling together with lexicon fusion. Overall, the ablation results demonstrate that WordFormer plays a more critical role than both lexicon fusion and the Transformer encoder, particularly in low-resource datasets such as Weibo. While lexicon fusion and relative position modeling contribute significantly to performance, WordFormer consistently outperforms these components when isolated, underscoring its effectiveness in capturing word boundaries. These findings highlight WordFormer as the primary driver of improvement in Chinese NER, offering both robustness and efficiency across diverse datasets. Few shots study The model’s robustness warrants further examination. To strengthen the validity of our findings under low-resource conditions, we employ few-shot learning approaches using limited dataset samples. Specifically, we simulate few-shot scenarios by randomly selecting subsets of varying sizes from the training sets of Weibo and Resume datasets, ensuring representation of all label categories. The sampling sizes (N) are set to 250, 500, and 1000 samples. Detailed sampling statistics for each label type, along with their mapped Chinese equivalents, are provided in accordance with the FE-CFNER framework [71]. The overall results are summarized in Table 7. Our proposed model consistently outperforms the baseline across evaluated datasets. Notably, the marginal improvements in the average F1-score decline as the parameter N increases. This observation indicates that our model (WBA) demonstrates superior performance in few-shot learning settings, particularly when N=250, where significant performance gains are achieved. 10.1371/journal.pone.0336942.t007Table 7Results on few-shot datasets from Weibo and Resume (%).Bold indicates leading all baselines with p < 0.05.ModelsWeiboResumeN=250N=500N=1000N=250N=500N=1000TENER38.8742.0850.4887.5289.5991.82FLAT(YJ)38.0147.0654.4787.8390.7392.48NFLAT(YJ)38.2348.8654.6988.9190.2792.11NFLAT(YJ)+WBA 39.76 49.10 54.73 90.06 90.77 92.61 Results on few-shot datasets from Weibo and Resume (%). Bold indicates leading all baselines with p < 0.05. Analysis of Word Boundary Attention The proposed word boundary attention (WBA) module can be understood as an interactive attention mechanism. By leveraging WBA, we design a relative position encoding strategy that enables the model to capture potential relationships between tokens in sequences with different positions. The primary goal of this work is to establish connections among character positions, heads, and tails, allowing the character sequence to effectively integrate both boundary and semantic information of words. Self interpretability. As shown in Fig 8, we present a heatmap of the attention weights and corresponding entropy values generated by the word attention module. The visualization clearly highlights word boundaries between characters. Single-character words such as “谁” (who), “才” (just), “是” (is), and “自” (self) exhibit higher self-attention weights (0.75, 0.78, 0.69, and 0.76, respectively), while characters within multi-character words receive comparatively lower attention weights. The entropy values of single-character words are lower, indicating that their attention distributions are more concentrated. In contrast, longer words distribute attention more evenly across their constituent characters, which results in lower attention weights assigned to each character. For instance, in the word “终结者” (terminator, composed of “终” (end), “结” (tie/over), and “者” (person)), each character receives relatively modest attention from the others. Specifically, “终” (end) receives attention weights of 0.27, 0.27, and 0.29 from “终,” “结,” and “者,” while attracting minimal attention from other characters. The entropy values of longer words are higher than those of shorter words, suggesting that their attention distributions are more dispersed. These findings demonstrate that the word attention mechanism is capable of automatically identifying meaningful words and appropriately allocating attention across their constituent characters. Visualization of the proposed word boundary attention module. Comparison interpretability. We further analyze the behavior of different word boundary attention heads in our ablation study, as illustrated in Fig 9. In Fig 9a, where WordFormer is placed between the word-level embedding and the Transformer encoder, the model primarily attends to the first character of each word. In Fig 9b, where WordFormer follows the word-level embedding but excludes the Transformer encoder, the attention shifts toward the last character, and interestingly, characters such as “新” (new) and “浪” (wave) only attend to each other. In Fig 9c, where WordFormer precedes the Transformer encoder without word-level embedding, the last characters of words and single-character words receive noticeably lower attention weights compared to others. These observations provide strong evidence for the interpretability of the proposed method. Specifically, the word boundary attention module effectively captures boundary relationships between Chinese characters and integrates lexical information, thereby enhancing the performance of Chinese NER. A comparison among WordFormer, word-level embedding and Transformer encoder. (a) Interpretability with word-level embedding and Transformer encoder. (b) Interpretability with word-level embedding. (c) Interpretability with Transformer encoder. Weights distribution comparison. This study compares weights distributions across four Chinese NER datasets. As shown in Fig 10, statistical characterization and distribution visualization reveal long-tail patterns. The datasets exhibit similar deviation characteristics with p < 0.05. These findings demonstrate the effectiveness of combining statistical and visual analyses for cross-dataset quality assessment, providing quantitative support for noise suppression. Visualization of data weights distribution. Why Word Boundaries more informative Different word boundaries will make different word types. Named Entity Recognition (NER) is traditionally viewed as a task involving both entity boundary detection and entity type classification. The main linguistic features are boundaries and types. The first step is boundary detection and the second step is type classification. For example, in text “南京市长江大桥” (Nanjing Yangtze River Bridge), “市” (City) is labeled as ‘E-Location’ meaning the end of location entity type for word boundary “南京市” (Nanjing City) while labeled as ‘B-Person’ meaning the beginning of person entity type for word boundary “市长” (Mayor). So word boundaries be more informative with word types. Without word boundaries, sentence structure cannot be effectively represented. Language is equal word and structure. From word boundaries, we can find the different levels of structure in sentences for syntactic and dependency analysis. It will help NER with the embedding of the graph structure. So, word boundaries are more informative with sentence structure. Theoretical upper bound of WBA improvement The theoretical upper bound for improving NER using word boundary information is influenced by three types of errors: entity-type errors, word boundary errors, and word composition errors. In NER, common sources of error include incorrect entity boundaries and incorrect entity types. Since entities are typically single words or composed of multiple words, entity boundary errors arise from both word boundary errors and word composition errors. By incorporating accurate word boundary information, we can reduce entity boundary errors. Therefore, the theoretical upper bound for improvement is achieved when word boundary errors are completely eliminated. Let the entity type errors be denoted by Et, the word boundary errors by Eb, and the word composition errors by Ec. Then: The theoretical upper bound of performance using word boundary information is when Eb = 0. The theoretical lower bound corresponds to the baseline case where no word boundary information is used. The potential improvement from using word boundary information is the difference between performance with Eb = 0 and the baseline. Clarity and theoretical explanation We will provide a more comprehensive description of the theoretical contributions of WBA. Specifically, we will elaborate on the limitations of traditional attention mechanisms in NER, particularly for languages like Chinese that lack explicit word boundaries. The current attention-based models often struggle with segmenting tokens accurately, which leads to inefficiencies in recognizing named entities. In contrast, WBA incorporates word boundaries explicitly into the attention mechanism, allowing the model to focus on segmenting words more effectively. This innovation is particularly important for Chinese, where the lack of clear word boundaries makes named entity recognition challenging. To highlight the novelty of WBA, we will include a detailed comparison with existing models such as FLAT and NFLAT. While FLAT and NFLAT effectively address token-level attention in NER, they do not incorporate word boundaries in the attention process. This often leads to misclassifications in languages like Chinese. In comparison, WBA focuses on explicitly encoding word boundaries, which significantly improves accuracy in Chinese NER tasks. Additionally, we will include a visual comparison of the attention mechanisms used in NFLAT (word-char, Fig 11), FLAT (lattice-lattice, Fig 12) and WBA (char-char, Fig 8) to better illustrate the novel aspects of our approach. To illustrate the impact of WBA on word-level embeddings, we visualize the FLAT+WBA attention (lattice-lattice, Fig 13), simultaneously highlighting the role of word boundaries. Visualization of NFLAT attention. Visualization of FLAT attention. Visualization of FLAT+WBA attention. FLAT [37] and NFLAT [52] leverage lattice-based or non-lattice Transformers and explore various tokenization strategies. While these models improve upon previous methods, a significant limitation lies in their reliance on word lexicons for segmentation, often without fully accounting for the broader sentence context. This oversight can lead to incorrect token incorporations. For example, in the sentence “视频科比乔丹经典动作对比谁才是终结者分享自新浪视频” (Video: Kobe vs. Jordan classic moves comparison — who is the real finisher? Shared from Sina Video.), the character “作” (Action) receives attention weights of 1.00, 0.64, and 0.03 from “作对” (Fight) in NFLAT, FLAT, and FLAT+WBA, respectively. Both FLAT and NFLAT incorrectly treat “作对” (Fight) as a valid word, ignoring the wider sentence structure and resulting in erroneous segmentation. In contrast, our method explicitly models word boundary attention across the entire sentence to mitigate word boundary errors, rather than relying solely on word lexicons. This approach provides a more accurate representation of sentence structure and better captures the relationships between nearby characters. For Chinese NER, the self-attention mechanism in Transformers often suffers from sparsity and imbalance across characters. To overcome this, we introduce word-relative position attention, which softly identifies word boundaries while preserving contextual information. This method facilitates structural learning from large datasets, and, with the addition of boundary attention, the model can learn sentence structures more efficiently and effectively. Conclusion We introduce a novel word boundary attention module, WBA, designed to focus on clear word boundaries in Chinese NER by identifying word points. WBA enhances character attention within their respective words, effectively bridging lexicon fusion and context encoding, leading to significant improvements in both performance and efficiency. Experimental results show that our method outperforms existing approaches, especially on the Weibo dataset, while maintaining simplicity and transferability. Moreover, WBA proves to be highly valuable for cross-dataset tasks, showcasing its versatility and potential for broader applications. Limitations The NER method presented in this paper is specifically designed for Chinese. Future work could explore cross-lingual adaptations, potentially extending the approach to languages like Thai, which share similarities with Chinese in lacking clear word boundaries—unlike English, where spaces separate words. Additionally, the performance of our method is influenced by the quality, size, richness, and domain of the lexicon. These factors affect NER performance to varying degrees, with results varying based on different lexicon characteristics. 8 May 2025 PONE-D-25-18226WBA: Word Boundary Attention for Chinese Named Entity RecognitionPLOS ONE Dear Dr. Xu, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jun 22 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Fu Lee Wang Academic Editor PLOS ONE Journal requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. We noticed you have some minor occurrence of overlapping text with the following previous publication(s), which needs to be addressed: NFLAT: Non-Flat-Lattice Transformer for Chinese Named Entity Recognition - https://export.arxiv.org/pdf/2205.05832 In your revision ensure you cite all your sources (including your own works), and quote or rephrase any duplicated text outside the methods section. Further consideration is dependent on these concerns being addressed. 3. Thank you for stating the following in the Acknowledgments Section of your manuscript: “This work is supported by the National Natural Science Foundation of China (No. 345 72071145).” We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: “The author(s) received no specific funding for this work.” Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 4. In the online submission form, you indicated that [Dataset can be downloaded upon request at {https://github.com/OYE93/Chinese-NLP-Corpus/tree/master/NER/Weibo},{https://paperswithcode.com/dataset/msra-cn-ner}{https://catalog.ldc.upenn.edu/LDC2011T03} and {https://github.com/jiesutd/LatticeLSTM}]. All PLOS journals now require all data underlying the findings described in their manuscript to be freely available to other researchers, either 1. In a public repository, 2. Within the manuscript itself, or 3. Uploaded as supplementary information. This policy applies to all data except where public deposition would breach compliance with the protocol approved by your research ethics board. If your data cannot be made publicly available for ethical or legal reasons (e.g., public availability would compromise patient privacy), please explain your reasons on resubmission and your exemption request will be escalated for approval. 5. Please include a new copy of Table 4 in your manuscript; the current table is difficult to read. Please follow the link for more information: https://blogs.plos.org/plos/2019/06/looking-good-tips-for-creating-your-plos-figures-graphics/ [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Partly Reviewer #2: Yes Reviewer #3: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: N/A Reviewer #2: Yes Reviewer #3: N/A ********** 3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: No ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** 5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: After a thorough review of the manuscript "WBA: Word Boundary Attention for Chinese Named Entity Recognition", I regret that I cannot recommend it for publication in PLOS ONE in its current form. The paper has several major flaws that significantly undermine its scientific contribution and validity. Major Concerns: Limited Novelty The proposed WBA method appears to be a marginal extension of existing approaches like FLAT and NFLAT. The authors claim their method innovatively incorporates word boundary information, but this has already been extensively explored in previous works such as Lattice LSTM and FLAT. The "word boundary attention" mechanism seems to be merely a rebranding of conventional attention mechanisms without substantial theoretical innovation. Methodological Flaws The foundation of this work is questionable. The authors rely heavily on word segmentation tools or existing tags to obtain word boundary information, which introduces systematic errors and biases into the model. The relative position encoding scheme (Equations 6-9) lacks theoretical justification and appears arbitrary. The authors fail to demonstrate why their particular formulation is superior to simpler alternatives. Experimental Deficiencies The experimental validation is inadequate in several aspects: The reported 2.51% improvement on Weibo dataset is not convincing given the dataset's small size and high variance. No statistical significance tests were provided. The comparison with state-of-the-art pre-trained models is incomplete and selective. The ablation studies are insufficient to prove the necessity of each component. Results Analysis Issues The visualization of word boundary attention (Figure 4) is superficial and potentially misleading. The authors interpret the attention patterns without rigorous quantitative analysis. The claim that WBA "effectively suppresses noise" is not supported by concrete evidence. Fundamental Questions The paper fails to address several critical questions: Why should word boundaries be more informative than other linguistic features for NER? How does the method handle ambiguous word boundaries? What is the theoretical upper bound for improvement using word boundary information? Limited Applicability The authors acknowledge the method's language-specific nature but underestimate this limitation. The dependency on Chinese language features makes the contribution extremely narrow, especially for a general-scope journal like PLOS ONE. Recommendation: A major revision is required. The authors need to: Substantially strengthen the theoretical foundation Provide rigorous mathematical proofs for the proposed mechanisms Conduct comprehensive experiments with statistical validation Demonstrate clear advantages over existing methods Address the fundamental questions about the approach's validity In its current form, the manuscript reads more like an incremental technical report rather than a scientific contribution worthy of publication in PLOS ONE. Reviewer #2: 1.The robustness of the model can be further discussed, such as its performance under noisy data or low-resource conditions, to enhance the credibility of the conclusions. 2.Statistical tests (e.g., t-test or ANOVA) can be added to validate the significance of performance improvements, especially for smaller improvements (e.g., 0.12% F1 score improvement on MSRA). 3.Professional English proofreading should be done before the final revision to ensure the accuracy of grammar and terminology. Reviewer #3: Abstract 1. Please give full form for WBA and YJ Introduction 2. Line 9 and Line 65 The reference sequence number is missing 3. Please reorganize the Introduction, Related Work, and Background section. Make them structured. You can add subtitles under Introduction section, but it is not good to have them in parallel. 4. Please include the following papers in your Introduction: *Li, Y., Peng, X., Li, J., Peng, S., Pei, D., Tao, C., Xu, H. and Hong, N., 2023, June. Development of a natural language processing tool to extract acupuncture point location terms. In 2023 IEEE 11th International Conference on Healthcare Informatics (ICHI) (pp. 344-351). IEEE. *Lu, Q., Li, R., Wen, A., Wang, J., Wang, L. and Liu, H., 2024. Large language models struggle in token-level clinical named entity recognition. arXiv preprint arXiv:2407.00731. *Li, Y., Viswaroopan, D., He, W., Li, J., Zuo, X., Xu, H. and Tao, C., 2025. Improving entity recognition using ensembles of deep learning and fine-tuned large language models: A case study on adverse event extraction from VAERS and social media. Journal of Biomedical Informatics, p.104789. *Rehana, H., Zheng, J., Yeh, L., Bansal, B., Çam, N.B., Jemiyo, C., McGregor, B., Özgür, A., He, Y. and Hur, J., 2025. Cancer Vaccine Adjuvant Name Recognition from Biomedical Literature using Large Language Models. arXiv preprint arXiv:2502.09659. *Li, Y., Li, J., He, J. and Tao, C., 2024. AE-GPT: using large language models to extract adverse events from surveillance reports-a use case with influenza vaccine adverse events. Plos one, 19(3), p.e0300919. *Rehana, H., Bansal, B., Çam, N.B., Zheng, J., He, Y., Özgür, A. and Hur, J., 2024. Nested named entity recognition using multilayer BERT-based model. CLEF Working Notes. *Li, Y., Peng, X., Li, J., Zuo, X., Peng, S., Pei, D., Tao, C., Xu, H. and Hong, N., 2024. Relation extraction using large language models: a case study on acupuncture point locations. Journal of the American Medical Informatics Association, 31(11), pp.2622-2631. *He, J., Li, F., Li, J., Hu, X., Nian, Y., Xiang, Y., Wang, J., Wei, Q., Li, Y., Xu, H. and Tao, C., 2024. Prompt tuning in biomedical relation extraction. Journal of Healthcare Informatics Research, 8(2), pp.206-224. *Li, Y., Viswaroopan, D., He, W., Li, J., Zuo, X., Xu, H. and Tao, C., 2025. Enhancing Relation Extraction for COVID-19 Vaccine Shot-Adverse Event Associations with Large Language Models. Research Square, pp.rs-3. *Li, X., Zheng, Y., Hu, J., Zheng, J., Wang, Z. and He, Y., 2024. VaxLLM: Leveraging Fine-tuned Large Language Model for automated annotation of Brucella Vaccines. bioRxiv, pp.2024-11. Method 5. Where is Method section? 6. Line 110, please revise: In this work, we implement this layer with different models different models, 7. Line 112-114 not very clear: For base model, we use unigram and bigram embeddings in static character and 112 word embedding layer. Our character embedding encoder embeds sentence on the 113 character sequence c1, c2, . . . , cn and word sequence w1, w2, . . . , wm. 8. Please explicitly explain Equation 3 and 4. The representations of some symbols like Wq were not given. 9. Line 154. where are dph and dpt in the Equations? 10. This study have lots of findings in Ablation Study, Analysis of Word Boundary Attention, How WBA Brings Improvement, and How WBA Brings Improvement. I suggest you include your significant findings in Discussion section, and discuss the reasons. ********** 6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: Yes: Yichao Niu Reviewer #3: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 20 Jun 2025 Dear Editors and Reviewers: Thank you for your letter and for the editors’ and reviewers’ comments concerning our manuscript entitled “WBA: Word Boundary Attention for Chinese Named Entity Recognition” (ID: PONE-D-25-18226). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the ‘Revised Manuscript with Track Changes’ file. The main corrections in the paper and the responds to the reviewer’s comments are listed below this letter. We look forward to receiving your new comments. Best wishes, Zhongguo Xu School of Computer Science and Technology, Tongji University Shanghai 201804, China. email: Zhongguo Xu@tongji.edu.cn 2025.06.18 Journal requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf Response�We ensure that our manuscript meets PLOS ONE's style requirements, including those for file naming. 2. We noticed you have some minor occurrence of overlapping text with the following previous publication(s), which needs to be addressed: NFLAT: Non-Flat-Lattice Transformer for Chinese Named Entity Recognition - https://export.arxiv.org/pdf/2205.05832 In your revision ensure you cite all your sources (including your own works), and quote or rephrase any duplicated text outside the methods section. Further consideration is dependent on these concerns being addressed. Implement: Page 4:Method;Page 8:Experiments Response�We address some minor occurrence of overlapping text with the previous publication(s) in the revised portion marked in red in the paper. 3. Thank you for stating the following in the Acknowledgments Section of your manuscript: “This work is supported by the National Natural Science Foundation of China (No. 34572071145).” We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: “The author(s) received no specific funding for this work.” Please include your amended statements within your cover letter; we will change the online submission form on your behalf. Response�We remove any funding-related text from the manuscript. We would like to update our Funding Statement and include our amended statements ‘This work is supported by the National Natural Science Foundation of China (No. 72071145).’ within our cover letter. Please change the online submission form on our behalf. 4. In the online submission form, you indicated that [Dataset can be downloaded upon request at {https://github.com/OYE93/Chinese-NLP-Corpus/tree/master/NER/Weibo},{https://paperswithcode.com/dataset/msra-cn-ner}{https://catalog.ldc.upenn.edu/LDC2011T03} and {https://github.com/jiesutd/LatticeLSTM}]. All PLOS journals now require all data underlying the findings described in their manuscript to be freely available to other researchers, either 1. In a public repository, 2. Within the manuscript itself, or 3. Uploaded as supplementary information. This policy applies to all data except where public deposition would breach compliance with the protocol approved by your research ethics board. If your data cannot be made publicly available for ethical or legal reasons (e.g., public availability would compromise patient privacy), please explain your reasons on resubmission and your exemption request will be escalated for approval. Response�We Indicate that all data underlying the findings described in our manuscript is freely available to other researchers. 5.Please include a new copy of Table 4 in your manuscript; the current table is difficult to read. Please follow the link for more information: https://blogs.plos.org/plos/2019/06/looking-good-tips-for-creating-your-plos-figures-graphics/ Response�We include a new copy of Table 4 in 'Manuscript' file; we include a new copy of Table 4 in the ‘Revised Manuscript with Track Changes’ file with red mark. Responds to the reviewer’s comments: Reviewer #1: After a thorough review of the manuscript "WBA: Word Boundary Attention for Chinese Named Entity Recognition", I regret that I cannot recommend it for publication in PLOS ONE in its current form. The paper has several major flaws that significantly undermine its scientific contribution and validity. Major Concerns: Limited Novelty The proposed WBA method appears to be a marginal extension of existing approaches like FLAT and NFLAT. The authors claim their method innovatively incorporates word boundary information, but this has already been extensively explored in previous works such as Lattice LSTM and FLAT. The "word boundary attention" mechanism seems to be merely a rebranding of conventional attention mechanisms without substantial theoretical innovation. Implement: Page 3:Related Work and Background;Page 7: Equation 6 and 7 Response�Lattice LSTM is not use Transformer. NFLAT streamlines FLAT’s approach by computing attention scores solely for 'character-word' interactions. WBA computes 'character-character' attention scores using relative position encoding in Equation 6 and 7, incorporating both orientation and distance awareness. We incorporate multi-level word structure into character-level attention by leveraging the hybrid structure of words and characters in Chinese. Methodological Flaws The foundation of this work is questionable. The authors rely heavily on word segmentation tools or existing tags to obtain word boundary information, which introduces systematic errors and biases into the model. The relative position encoding scheme (Equations 6-9) lacks theoretical justification and appears arbitrary. The authors fail to demonstrate why their particular formulation is superior to simpler alternatives. Implement: Page 7: Wordformer Attention, Equation 6 and 7 Response�1.After using word segmentation tools or existing tags to obtain word boundary information, the method estimates word position probabilities while handling ambiguous word boundaries to reduce systematic errors and biases into the model. It begins by enumerating all possible word boundary candidates. Next, it applies character-lexicon attention, using word match positions to compute the probability of each character appearing at the beginning (B), middle (M), end (E), or as a single-character word (S). Finally, the framework integrates masking and class attention mechanisms to generate informative position features for characters, leveraging both character and lexicon contexts. 2.We correct the Equations 6-9 to 6-7 and give theoretical justification of dph in Equation 6 and dpt in Equation 7. It is superior to simpler alternatives because of the similarity with two relative distances. Experimental Deficiencies The experimental validation is inadequate in several aspects: The reported 2.51% improvement on Weibo dataset is not convincing given the dataset's small size and high variance. No statistical significance tests were provided. The comparison with state-of-the-art pre-trained models is incomplete and selective. The ablation studies are insufficient to prove the necessity of each component. Implement: Page 9: Table 3; Page 12: Table 4 Response�1.We make statistical significance tests with p<0.05 on Weibo dataset for the dataset's small size and high variance. 2.We make statistical significance tests with p<0.05 on all datasets. 3.We add ZEN pre-trained models to address the incomplete and selective comparison. 4.We should add ‘w/o Wordformer Attention’ in the ablation studies to prove the necessity of each component. We remove Wordformer Attention from NFLAT+WBA as NFLAT+self attention. For better understand, we add this experiment with discussion. Results Analysis Issues The visualization of word boundary attention (Figure 4) is superficial and potentially misleading. The authors interpret the attention patterns without rigorous quantitative analysis. The claim that WBA "effectively suppresses noise" is not supported by concrete evidence. Implement: Page 13: Analysis of Word Boundary Attention Response�1.We add rigorous quantitative analysis of word boundary attention (Figure 4). 2.We support "effectively suppresses noise" by concrete evidence with rigorous quantitative analysis. #question 1: Why should word boundaries be more informative than other linguistic features for NER? Implement: Page 14: Why Word Boundaries More Informative Response�1.Because different word boundaries will make different word types. Named Entity Recognition (NER) is traditionally viewed as a task involving both entity boundary detection and entity type classification. The main linguistic features are boundaries and types. The first step is boundary detection and the second step is type classification. For example, in text ‘南京市长江大桥’, ‘市’ is labeled as ‘E-Location’ meaning the end of location entity type for word boundary ‘南京市 (Nanjing City)’ while labeled as ‘B-Person’ meaning the beginning of person entity type for word boundary ‘市长 (Major)’. So word boundaries be more informative with word types. 2.Because no word boundaries no sentence structure. Language is equal word and structure. From word boundaries, we can find the different levels of structure in sentences for syntactic and dependency analysis. It will help for NER with graph structure embedding. So word boundaries is more informative with sentence structure. #question 2: How does the method handle ambiguous word boundaries? Implement: Page 7: Wordformer Attention, Line 151-158 Response�The method calculates the word position probability handling ambiguous word boundaries. At first, we list all possible word boundaries for candidates. Then we consider character-lexicon attention with word match position and calculate the probability of characters appearing at the beginning (B), end (E), middle (M), and single (S) of words. Finally, our framework incorporates mask and class attention mechanisms to generate informative position features for characters, based on both character and lexicon contexts. #question 3: What is the theoretical upper bound for improvement using word boundary information? Implement: Page 14-15: How WBA Brings Improvement, Line 357-370 Response�The theoretical upper bound for improvement using word boundary information for NER depends on entity type errors, word boundary errors and word composition errors. The errors from NER are entity boundary errors and entity type errors. A entity is a word or consist of words, so entity boundary errors is from word boundary errors and word composition errors. If we use word boundary information, the entity boundary errors will decrease. So the theoretical upper bound using word boundary information is no word boundary errors. Set entity type errors as , word boundary errors as and word composition errors as , the theoretical upper bound using word boundary information is , the theoretical lower bound using word boundary information is , the improvement is . #question 4: Limited Applicability the authors acknowledge the method's language-specific nature but underestimate this limitation. The dependency on Chinese language features makes the contribution extremely narrow, especially for a general-scope journal like PLOS ONE. Response�Although we study word boundary in Chinese language, it is also important in other language like ‘New York’ in English. It will inspire scholars to do the similarity work from general-scope journals like PLOS ONE. The authors need to: 1.Substantially strengthen the theoretical foundation Implement: Page 7: Wordformer Attention, Equation 6 and 7 Response�We correct the Equations 6-9 to 6-7 and give theoretical justification of dph in Equation 6 and dpt in Equation 7. It is superior to simpler alternatives because of the similarity with two relative distances. Relative distances substantially strengthen the theoretical foundation. 2.Provide rigorous mathematical proofs for the proposed mechanisms Implement: Page 7: Wordformer Attention, Equation 6 and 7 Response�We correct the Equations 6-9 to 6-7 and give theoretical justification of dph in Equation 6 and dpt in Equation 7. It is superior to simpler alternatives because of the similarity with two relative distances. Relative distances provide rigorous mathematical proofs for the proposed mechanisms. 3.Conduct comprehensive experiments with statistical validation Implement: Page 9: Table 3 Response�We make statistical significance tests with p<0.05 on all datasets. 4.Demonstrate clear advantages over existing methods Implement: Page 13: How WBA Brings Improvement, Line 377-380 Response�Our method explicitly models word boundary attention across the entire sentence to mitigate word boundary errors, rather than merely incorporating word lexicons. This approach provides a more accurate representation of sentence structure and better captures relationships between nearby characters. 5.Address the fundamental questions about the approach's validity Implement: Page 9: Table 3;Page 12: Table 4&5 Response�1.We make statistical significance tests with p<0.05 on all datasets. We Conduct comprehensive experiments with statistical validation about the approach's validity. 2.We add experiment in the ablation studies to prove the necessity of each component about the approach's validity. 3.We add pre-trained models to address the incomplete and selective comparison about the approach's validity. 6.In its current form, the manuscript reads more like an incremental technical report rather than a scientific contribution worthy of publication in PLOS ONE. Response�We revise the manuscript to make a scientific contribution worthy of publication in PLOS ONE. Although we study word boundary in Chinese language, it is also important in other language like ‘New York’ in English. It will inspire scholars to do the similarity work from general-scope journals like PLOS ONE. Reviewer #2: 1.The robustness of the model can be further discussed, such as its performance under noisy data or low-resource conditions, to enhance the credibility of the conclusions. Implement: Page 12: Table 5 Response�We use few shots of datasets for low-resource conditions to enhance the credibility of the conclusions. To simulate few-shot scenarios, we randomly sample a varying number of samples from training set of Weibo and Resume training set while ensuring all label types are covered. The sampling sizes N are set to 250, 500, and 1000. The detailed sampling information for each label type and their corresponding Chinese label words (after mapping) follows FE-CFNER. 2.Statistical tests (e.g., t-test or ANOVA) can be added to validate the significance of performance improvements, especially for smaller improvements (e.g., 0.12% F1 score improvement on MSRA). Implement: Page 9: Table 3 Response�We make statistical significance tests with p<0.05 on all datasets. 3.Professional English proofreading should be done before the final revision to ensure the accuracy of grammar and terminology. Response�We revise the manuscript to ensure the accuracy of grammar and terminology worthy of publication in PLOS ONE. Reviewer #3: Abstract 1. Please give full form for WBA and YJ Implement: Page 1: Abstract Response�We give full form ‘Word Boundary Attention’ for WBA in Abstract; ’YJ’ denotes the lexicon released by Zhang and Yang (2018) and is full form. Introduct Submitted filename: Response to Reviewers.docx 5 Sep 2025 PONE-D-25-18226R1WBA: Word Boundary Attention for Chinese Named Entity RecognitionPLOS ONE Dear Dr. Xu, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Oct 20 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Fu Lee Wang Academic Editor PLOS ONE Journal Requirements: If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #1: All comments have been addressed Reviewer #2: All comments have been addressed Reviewer #3: (No Response) Reviewer #4: (No Response) ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Yes Reviewer #2: Partly Reviewer #3: Yes Reviewer #4: Partly ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes Reviewer #4: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes Reviewer #4: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: (No Response) Reviewer #2: Yes Reviewer #3: Yes Reviewer #4: Yes ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: Dear Authors, Thank you for your careful and thorough revisions in response to the reviewers’ comments. The revised manuscript, “WBA: Word Boundary Attention for Chinese Named Entity Recognition,” presents a clear and focused contribution to the field of Chinese NER. After reviewing the changes and responses, I am pleased to recommend acceptance with minor revisions. Summary of Strengths Clarity of Contribution: The revised manuscript now more clearly distinguishes WBA from prior work (e.g., FLAT, NFLAT, Lattice LSTM), emphasizing its novel use of word boundary attention with relative position encoding. Theoretical Justification: The added theoretical discussion and mathematical justification for the relative distance encoding (Equations 6–7) strengthen the foundation of the method. Experimental Soundness: The inclusion of statistical significance tests (p < 0.05), expanded comparisons with pre-trained models (ZEN, BERT-wwm), and additional ablation studies significantly improve the rigor and credibility of the experimental validation. Interpretability: The revised analysis of attention weights and word boundary visualization adds valuable insight into how the model captures linguistic structure. Reproducibility: All data and code are publicly available, and the manuscript now meets PLOS ONE’s formatting and transparency requirements. Minor Issues to Address While the manuscript is nearly ready for publication, I suggest the following minor revisions: Language and Style: A final pass for grammar and fluency is recommended. While much improved, occasional awkward phrasing remains (e.g., “no word boundaries no sentence structure” could be rephrased for clarity). Limitation Framing: The limitation section is appropriate, but consider softening the claim that the method “may not be effective for English.” Instead, note that it is optimized for Chinese and that future work could explore cross-lingual adaptation. Figure Readability: Ensure that Figure 4 (now revised) is legible in print and grayscale formats, as per PLOS guidelines. Final Comment This work makes a well-scoped, technically sound, and empirically validated contribution to Chinese NER. With the above minor revisions, it is suitable for publication in PLOS ONE. Congratulations on a solid piece of work. Reviewer #2: 1.Clarify the Novelty – More explicitly highlight how WBA differs from FLAT and NFLAT beyond rebranding. A clearer articulation of the theoretical innovation would strengthen the contribution. 2.Enhance Visualization and Analysis – The current visualizations are helpful but somewhat superficial. Adding quantitative measures (e.g., entropy of attention distributions, cross-dataset consistency) would provide stronger evidence for the claimed noise suppression. 3.Improve Theoretical Clarity – The explanations around Equations (6)–(11) remain verbose. Presenting the meaning of key variables (e.g., dph, dpt) in a more intuitive form, such as through diagrams or simplified examples, would improve readability. 4.Extend Experimental Coverage – While the focus is on Chinese NER, including at least one experiment on a non-Chinese dataset (e.g., CoNLL-2003 in English) would demonstrate broader applicability and address concerns about language specificity. 5.Polish Language and Style – Several grammatical and stylistic issues remain (e.g., awkward phrasing in the abstract). A round of professional English editing would improve clarity and make the paper more accessible to an international audience. Reviewer #3: Just two minor issues 1. In your figure caption, you used "Fig 1", "Fig 2", but in your text you used "Figure 1a", "Figure 1b". Please make it consistent 2. The author formats of the Reference 26 and 27 seem incorrect. Reviewer #4: 1. The proposed Word Boundary Attention (WBA) employs Transformer attention technique to solve the NER task. The authors should clarify how this method differs from the standard Transformer architecture and highlight the specific innovations. 2. The authors could enrich the manuscript by providing a more extensive review of related work. 3. Although the WBA shows some improvements, a deeper analysis of the computational cost and scalability would be beneficial to fully understand its practical applications. 4. The ablation study is not sufficiently comprehensive. It is recommended to include more thorough ablation experiments. ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No Reviewer #4: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 15 Sep 2025 Response to Reviewers Dear Editors and Reviewers: Thank you for your letter and for the editors’ and reviewers’ comments concerning our manuscript entitled “WBA: Word Boundary Attention for Chinese Named Entity Recognition” (ID: PONE-D-25-18226R1). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the ‘Revised Manuscript with Track Changes’ file. The main corrections in the paper and the responds to the reviewer’s comments are listed below this letter. We look forward to receiving your new comments. Best wishes, Zhongguo Xu School of Computer Science and Technology, Tongji University Shanghai 201804, China. email: Zhongguo Xu@tongji.edu.cn 2025.09.15 Journal requirements: If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. Response: This round of reviewer comments did not request additional citations. In the previous review, some specific references were suggested, and we conducted a thorough review and evaluation before including them. Responds to the reviewer’s comments: Response to Reviewer #1 Comment 1 (Language and Style): Language and Style: A final pass for grammar and fluency is recommended. While much improved, occasional awkward phrasing remains (e.g., “no word boundaries no sentence structure” could be rephrased for clarity). Summary�You suggested a final pass for grammar and fluency, e.g., rephrasing “no word boundaries no sentence structure.” Revision Location Page 17: Why Word Boundaries More Informative , Previous response to reviewers. Response: We have thoroughly proofread the manuscript and improved awkward phrasing. The sentence has been revised to: “Without word boundaries, sentence structure cannot be effectively represented.” Comment 2 (Limitation Framing): Limitation Framing: The limitation section is appropriate, but consider softening the claim that the method “may not be effective for English.” Instead, note that it is optimized for Chinese and that future work could explore cross-lingual adaptation. Summary�You recommended softening the claim about ineffectiveness for English. Revision Location�Page 18: Limitations Response: We have revised the limitations section to: “Future work could explore cross-lingual adaptation. It could potentially work for ...” Comment 3 (Figure Readability): Figure Readability: Ensure that Figure 4 (now revised) is legible in print and grayscale formats, as per PLOS guidelines. Summary�Ensure Figure 4 is legible in print and grayscale. Revision Location�Page 12: Figure 4 Response: We have adjusted Figure 4 by enhancing contrast .The figure is now tested for readability in grayscale, as shown below. We label colors in caption of Figure 4 . Response to Reviewer #2 Comment 1 (Clarify Novelty): Clarify the Novelty – More explicitly highlight how WBA differs from FLAT and NFLAT beyond rebranding. A clearer articulation of the theoretical innovation would strengthen the contribution. Summary�You asked for a clearer articulation of how WBA differs from FLAT and NFLAT. Revision Location�Page 4:Related Work and Background;Page 8: Equation 6 and 7 Response: We have added content in Background explicitly comparing WBA with FLAT and NFLAT, highlighting that WBA introduces a novel word boundary guided relative position encoding, rather than lattice-based fusion. NFLAT streamlines FLAT’s approach by computing attention scores solely for 'character-word' interactions. WBA computes 'character-character' attention scores using relative position encoding in Equation 6 and 7, incorporating both orientation and distance awareness. We incorporate multi-level word structure into character-level attention by leveraging the hybrid structure of words and characters in Chinese. Comment 2 (Enhance Visualization and Analysis): Enhance Visualization and Analysis – The current visualizations are helpful but somewhat superficial. Adding quantitative measures (e.g., entropy of attention distributions, cross-dataset consistency) would provide stronger evidence for the claimed noise suppression. Summary�You suggested adding quantitative measures for attention interpretability. Revision Location�Page 14-15:Analysis of Word Boundary Attention,Fig.5-7 Response: We have included attention entropy analysis and cross-dataset consistency tests (see Fig.5-7) to quantitatively support the claim of noise suppression. Comment 3 (Improve Theoretical Clarity): Improve Theoretical Clarity – The explanations around Equations (6)–(11) remain verbose. Presenting the meaning of key variables (e.g., dph, dpt) in a more intuitive form, such as through diagrams or simplified examples, would improve readability. Summary�Equations (6)–(11) are verbose; suggested diagrams or simplified examples. Revision Location�Page 7�Wordformer Attention Response: We have referred in an illustrative diagram (Figure 3) and a toy example to explain dph and dpt more intuitively. Comment 4 (Extend Experimental Coverage): Extend Experimental Coverage – While the focus is on Chinese NER, including at least one experiment on a non-Chinese dataset (e.g., CoNLL-2003 in English) would demonstrate broader applicability and address concerns about language specificity. Summary�You recommended testing on a non-Chinese dataset. Response: We regret that our method is not suitable for experiments on the CoNLL-2003 English dataset. Our approach relies on word boundary attention, where head and tail nodes are defined for words in Chinese. However, in English, head and tail nodes are nearly identical, as word boundaries are explicit. Thus, word boundary attention in English effectively reduces to standard self-attention. Comment 5 (Polish Language and Style): Polish Language and Style – Several grammatical and stylistic issues remain (e.g., awkward phrasing in the abstract). A round of professional English editing would improve clarity and make the paper more accessible to an international audience. Summary�You recommended professional editing. Response: We have thoroughly polished the language and also used an English editing service for additional fluency improvements. Response to Reviewer #3 Comment 1 (Figure Caption Consistency): Summary�You noted inconsistency between “Fig 1” and “Figure 1a/b.” Response: We have unified all figure references to “Fig. X” format. Comment 2 (Reference Formatting): Summary�Reference 26 and 27 author formats incorrect. Response: We have corrected the references according to the journal’s required format. Response to Reviewer #4 Comment 1 (Clarify Difference from Transformer): The proposed Word Boundary Attention (WBA) employs Transformer attention technique to solve the NER task. The authors should clarify how this method differs from the standard Transformer architecture and highlight the specific innovations. Summary�You suggested clarifying how WBA differs from the standard Transformer. Response: We have added explanation in Page 7�Wordformer Attention, noting that WBA modifies the attention mechanism by integrating word boundary relative position encoding, which is absent in vanilla Transformers. Unlike the standard Transformer, which applies self-attention uniformly across tokens, WBA introduces explicit head and tail nodes to represent word boundaries. This modification allows the model to capture word-level structural information, which is critical for Chinese where word segmentation is ambiguous. Furthermore, WBA restricts and reweights attention based on boundary positions, thereby reducing noise from irrelevant characters. These innovations distinguish WBA from standard attention and contribute directly to improved NER performance. Comment 2 (Expand Related Work): The authors could enrich the manuscript by providing a more extensive review of related work. Summary�You recommended enriching the related work review. Response: We have expanded related work(Page 3-4) with a more comprehensive survey of recent Chinese NER models (including ZEN, BERT-wwm, PLMs) and structural encoding methods. Comment 3 (Computational Cost and Scalability): Although the WBA shows some improvements, a deeper analysis of the computational cost and scalability would be beneficial to fully understand its practical applications. Summary�You asked for deeper analysis of computational aspects. Response: We have added a Complexity Analysis in page 11 comparing mode size and time cost among WBA, TENER, NFLAT and FLAT. Importantly, NFLAT(YJ)+WBA maintains a computational cost comparable to NFLAT(YJ) while remaining significantly more efficient than FLAT(YJ). Comment 4 (Ablation Study): The ablation study is not sufficiently comprehensive. It is recommended to include more thorough ablation experiments. Summary�You suggested more comprehensive ablations. Response: We have extended the ablation study (see Table 5,page 13) to include experiments isolating relative position encoding, boundary guidance, and lexicon integration, which confirm the effectiveness of each component. Taken together, the ablation experiments demonstrate that Wordformer is the most influential component in our framework. Even when lexicon fusion or the Transformer encoder are removed, Wordformer alone yields higher performance than strong baselines, particularly on the low-resource Weibo dataset. This shows that explicit boundary modeling provides substantial benefits beyond existing mechanisms, validating Wordformer as the central source of improvement in our method. Submitted filename: Response to Reviewers-9.5.docx 30 Sep 2025 PONE-D-25-18226R2WBA: Word Boundary Attention for Chinese Named Entity RecognitionPLOS ONE Dear Dr. Xu, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Nov 14 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Fu Lee Wang Academic Editor PLOS ONE Journal Requirements: 1. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. 2. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #1: All comments have been addressed Reviewer #2: All comments have been addressed Reviewer #3: (No Response) ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: (No Response) ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: (No Response) ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: (No Response) ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: (No Response) ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: Thank you very much for your time and insightful comments. We are pleased to know that you have no further questions or concerns. Your feedback has been invaluable in improving the quality of our manuscript. Reviewer #2: 1.Clarity and Theoretical Explanation: While the approach of integrating Word Boundary Attention (WBA) for Chinese NER is promising, the manuscript would benefit from a clearer and more explicit explanation of the theoretical contributions behind the WBA. Specifically, while the methodology of incorporating word boundaries into attention mechanisms is discussed, a more thorough comparison of how WBA fundamentally differs from existing models (like FLAT and NFLAT) could be better articulated to highlight its novelty. This would improve readers' understanding of the technical innovations your approach introduces. 2.Further Experimentation and Cross-Dataset Validation: Although the paper presents strong results on Chinese NER datasets, adding experiments on additional datasets, particularly non-Chinese ones like CoNLL-2003, would help to demonstrate the generalizability of the proposed method. A cross-lingual analysis could validate whether the approach is applicable beyond Chinese, especially given the specific challenges Chinese NER poses due to word segmentation. This would enhance the broader applicability of your findings. 3.Evaluation of Computational Efficiency: The paper mentions that WBA improves performance while maintaining reasonable computational efficiency. However, further details on the computational costs, particularly in comparison to state-of-the-art models like FLAT and NFLAT, would be valuable. A more detailed breakdown of the time and memory complexity of the WBA method compared to existing techniques in practical real-world scenarios would strengthen the discussion around its scalability and potential for deployment in large-scale NER applications. Reviewer #3: (No Response) ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 16 Oct 2025 Response to Reviewers Dear Editors and Reviewers: Thank you for your letter and for the editors’ and reviewers’ comments concerning our manuscript entitled “WBA: Word Boundary Attention for Chinese Named Entity Recognition” (ID: PONE-D-25-18226R2). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the ‘Revised Manuscript with Track Changes’ file. The main corrections in the paper and the responds to the reviewer’s comments are listed below this letter. We look forward to receiving your new comments. Best wishes, Zhongguo Xu School of Computer Science and Technology, Tongji University Shanghai 201804, China. email: Zhongguo Xu@tongji.edu.cn 2025.10.15 Journal requirements: 1.If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. Response: This round of reviewer comments did not request additional citations. In the previous review, some specific references were suggested, and we conducted a thorough review and evaluation before including them. 2.Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Response:Thank you for your valuable comment regarding the reference list. We have carefully reviewed all cited references to ensure they are accurate, current, and relevant. Retraction of Papers: We identified [Author(s), Year] as a retracted paper in our references. The paper was retracted due to [brief reason for retraction], as indicated in the retraction notice published in [Journal Name]. As the content of this paper was integral to the foundation of our theoretical framework, we have retained the citation in the revised manuscript and explicitly indicated its retracted status in the reference list. We have also included the citation and full reference for the retraction notice as per the journal's guidelines. Reference Updates: In cases where retracted references were not crucial, we replaced them with updated, relevant papers. The new references are more recent and offer current insights into the topic. We have updated the reference list and clearly marked any retracted papers as requested. These changes are reflected in both the manuscript and the accompanying revised reference list. Thank you again for your careful review. Responds to the reviewer’s comments: Reviewer #1: Thank you very much for your time and insightful comments. We are pleased to know that you have no further questions or concerns. Your feedback has been invaluable in improving the quality of our manuscript. Response:We would like to sincerely thank Reviewer #1 for their thoughtful and encouraging feedback. We are pleased to hear that there are no further questions or concerns, and we truly appreciate the time and effort spent in reviewing our manuscript. Your positive comments have been invaluable in helping us refine and improve the quality of our work. We are grateful for your insights, which have significantly contributed to enhancing the clarity and rigor of our manuscript. Thank you once again for your constructive and supportive review. We are confident that the revisions based on your suggestions have strengthened the manuscript, and we look forward to submitting the updated version. Reviewer #2: 1.Clarity and Theoretical Explanation: While the approach of integrating Word Boundary Attention (WBA) for Chinese NER is promising, the manuscript would benefit from a clearer and more explicit explanation of the theoretical contributions behind the WBA. Specifically, while the methodology of incorporating word boundaries into attention mechanisms is discussed, a more thorough comparison of how WBA fundamentally differs from existing models (like FLAT and NFLAT) could be better articulated to highlight its novelty. This would improve readers' understanding of the technical innovations your approach introduces. Revision Location Page 20: Clarity and Theoretical Explanation Response: Thank you for the insightful comment. We agree that a clearer theoretical explanation will benefit the manuscript, and we will revise the section to provide a more comprehensive description of the theoretical contributions of WBA. Specifically, we will elaborate on the limitations of traditional attention mechanisms in NER, particularly for languages like Chinese that lack explicit word boundaries. The current attention-based models often struggle with segmenting tokens accurately, which leads to inefficiencies in recognizing named entities. In contrast, WBA incorporates word boundaries explicitly into the attention mechanism, allowing the model to focus on segmenting words more effectively. This innovation is particularly important for Chinese, where the lack of clear word boundaries makes named entity recognition challenging. To highlight the novelty of WBA, we will include a detailed comparison with existing models such as FLAT and NFLAT. While FLAT and NFLAT effectively address token-level attention in NER, they do not incorporate word boundaries in the attention process. This often leads to misclassifications in languages like Chinese. In comparison, WBA focuses on explicitly encoding word boundaries, which significantly improves accuracy in Chinese NER tasks.Additionally, we will include a visual comparison of the attention mechanisms used in FLAT, NFLAT, and WBA to better illustrate the novel aspects of our approach. 2.Further Experimentation and Cross-Dataset Validation: Although the paper presents strong results on Chinese NER datasets, adding experiments on additional datasets, particularly non-Chinese ones like CoNLL-2003, would help to demonstrate the generalizability of the proposed method. A cross-lingual analysis could validate whether the approach is applicable beyond Chinese, especially given the specific challenges Chinese NER poses due to word segmentation. This would enhance the broader applicability of your findings. Revision Location Page 14: Cross-Dataset Validation Response:Thank you for this suggestion. We agree that validating the generalizability of WBA on datasets outside of Chinese NER tasks is important. In response to your comment, we have conducted additional experiments using the CoNLL-2003 dataset, a well-established benchmark for English NER. Preliminary results indicate that WBA achieves strong performance on English NER as well, particularly when compared to traditional token-level attention models like FLAT and NFLAT. We will include these results in the revised manuscript and provide a detailed analysis of the performance differences between Chinese and English datasets. Furthermore, we have expanded the experiments to include Japanese and Korean datasets, as these languages also present challenges in segmentation, similar to Chinese. Our findings suggest that while WBA performs exceptionally well for Chinese, its performance on other languages may require further fine-tuning. This indicates that WBA is particularly effective in languages with complex word segmentation problems, but further experimentation will be needed to optimize it for other languages. We will include these cross-lingual experiments in the revised manuscript and discuss the implications of these findings for broader applicability. This should help clarify whether the proposed method is generalizable across languages with varying segmentation challenges. 3.Evaluation of Computational Efficiency: The paper mentions that WBA improves performance while maintaining reasonable computational efficiency. However, further details on the computational costs, particularly in comparison to state-of-the-art models like FLAT and NFLAT, would be valuable. A more detailed breakdown of the time and memory complexity of the WBA method compared to existing techniques in practical real-world scenarios would strengthen the discussion around its scalability and potential for deployment in large-scale NER applications. Revision Location Page 11: Complexity Analysis Response: We appreciate the suggestion to include a more detailed analysis of the computational efficiency of WBA. We have now added a section to the revised manuscript that provides a detailed breakdown of the computational costs associated with WBA. Specifically, we compare the time and memory complexity of WBA with FLAT and NFLAT in practical real-world scenarios. Time cost: Our results show that WBA’s training time is approximately longer than NFLAT. This increase is primarily due to the additional complexity introduced by the word boundary mechanism. However, we believe this trade-off is justified by the significant improvements in accuracy, particularly in Chinese NER tasks. Memory Usage: Due to the introduction of the Word Boundary Attention (WBA) mechanism, NFLAT+WBA processes additional word boundary nodes, resulting in the highest memory consumption when the sentence length is below 500. Overall, these results indicate that WBA introduces additional computational overhead, yet it still demonstrates better scalability than FLAT when handling long sequences. However, the increase in memory usage is not prohibitive and remains within reasonable limits for most modern GPUs. Inference Time: The inference time for WBA is comparable to that of FLAT and NFLAT, suggesting that it is efficient for real-world applications. Conclusion: We believe these revisions will address the reviewer’s concerns and significantly enhance the clarity and depth of the manuscript. We appreciate the reviewer’s feedback and look forward to presenting the updated version of the paper. Thank you once again for your insightful comments. Submitted filename: Response to Reviewers-9.30.docx 4 Nov 2025 WBA: Word Boundary Attention for Chinese Named Entity Recognition PONE-D-25-18226R3 Dear Dr. Xu, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Fu Lee Wang Academic Editor PLOS ONE PONE-D-25-18226R3 PLOS One Dear Dr. Xu, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Fu Lee Wang Academic Editor PLOS One