Can one size fit all?: Measuring Failure in Multi-Document Summarization Domain Transfer Alexandra DeLucia, Mark Dredze Abstract Abstractive multi-document summarization (MDS) is the task of automatically summarizing information in multiple documents, from news articles to conversations with multiple speakers. The training approaches for current MDS models can be grouped into four approaches: end-to-end with special pre-training (“direct”), chunk-then-summarize, extract-then-summarize, and inference with GPT-style models. In this work, we evaluate MDS models across training approaches, domains, and dimensions (reference similarity, quality, and factuality), to analyze how and why models trained on one domain can fail to summarize documents from another (News, Science, and Conversation) in the zero-shot domain transfer setting. We define domain-transfer “failure” as a decrease in factuality, higher deviation from the target, and a general decrease in summary quality. In addition to exploring domain transfer for MDS models, we examine potential issues with applying popular summarization metrics out-of-the-box. 1 Introduction Figure 1: Sample generations from Summ N trained on Multi-XScience (Summ N -Science) and PRIMERA trained on Multi-News (PRIMERA-News) on the ConvoSumm-Reddit dataset (Conversation). Relevance is measured automatically with UniEval. Multi-document summarization (MDS) is an extended task of summarization, where the goal is to represent the key ideas from multiple documents in a single new document or paragraph. Current MDS systems (models) are trained and evaluated on data from a single domain, such as news, but there have been no robust studies on how models trained on these domains transfer, or perform, on different domains. While the overarching goal of MDS models is to condense information, models trained on one type of summary (e.g., news) often fail to transfer successfully to other domains (e.g., science) due to the unique nature of how summaries across different domains are written. Recent MDS system surveys compare different models either on a single domain (e.g., conversation/dialogue ( Zhang et al. 2021 ; Mullick et al. 2024 ) ) or focus on MDS corpora biases ( Wolhandler et al. 2022 ; Dey et al. 2020 ) . While Zhang et al. 2021 found that retrieve-then-summarize models are best according to ROUGE, Mullick et al. 2024 did not find a clear pattern in model performance across a variety of metrics. Few works explore the zero-shot domain transfer ability of MDS systems, but Zhang et al. 2021 did find that pre-training on news datasets can transfer well to conversation. However, Fabbri et al. 2021a a zero-shot model from news to conversation had lower performance than on conversation alone. An example is in Figure 1 , where two models, one trained on Multi-News ( Fabbri et al. 2019 ) and the other on Multi-XScience ( Lu et al. 2020 ) are evaluated on a sample from ConvoSumm-Reddit ( Fabbri et al. 2021a ) . The difference in writing style is immediately apparent: the news-trained model draws out quotes, the science-trained model inserts citations, and both are completely different from the expected style of the Reddit conversation summary. The difference in style is also captured by similarity metrics ROUGE ( Lin 2004 ) and UniEval-Relevance ( Zhong et al. 2022 ) –while the ROUGE scores, which capture word overlap, only differ by a few points, the semantic-based, trained Relevance scores vary widely. While ConvoSumm-Reddit summaries are written to represent multiple points of view, other ways to represent online discourse is through an overall topic-based summary or detailed utterance-by-utterance summary. All types would “condense” the same conversation but at varying levels of granularity, style, and word choice. Similarly, summaries for other domains also vary in incorporated information and style. Figure 1 highlights the three axes of domain transfer for MDS systems: 1) domain of the training data, 2) type of model, and 3) evaluation metrics. While there are popular benchmarks for summarization, including the notable SummEval ( Fabbri et al. 2021b ) , few are for multi-document summarization, and most datasets focus on the news domain. We choose three high-quality human-written datasets to represent News (Multi-News ( Fabbri et al. 2019 ) ), Science (Multi-XScience ( Lu et al. 2020 ) ), and Conversation (ConvoSumm-Reddit ( Fabbri et al. 2021a ) ) domains. Regarding the type of model, while there are many models for summarization, the types of models for abstract multi-document summarization are end-to-end with special pre-training (“direct”), chunk-then-summarize, extract-then-summarize, and direct with large language model (LLM) decoders. There are many models within each type ( Mullick et al. 2024 ) and for greater in-depth analyses, we choose one, strong representative model of each kind. We focus on neural, transformer models and omit extractive comparisons due to previous work demonstrating the stronger performance of neural models ( Fabbri et al. 2021b ) . Similarly, while there are many evaluation metrics for generated text, which each have their strengths and weaknesses ( He et al. 2023 ; von Däniken et al. 2022 ; Fabbri et al. 2021b ) , we group commonly used metrics into three categories: 1) similarity to summary reference (ground-truth), 2) overall quality (e.g., fluency), and 3) groundedness in the source documents. We find that: 1 1 1 All code and models will be released upon publication. • Smaller language models (BART-size) hold more stylistic bias with their training data; • Within News, Science, and Conversation, there is no clear “best” domain for general transfer; • Lexical and semantic similarity metrics both fail to capture summary quality; • Current non-reference-based metrics do not scale well to the multi-document setting; • Trained evaluators have bias and penalize out-of-domain datasets; Figure 2: Summary of all the domains (datasets), multi-document summarization (MDS) model types, and evaluation metrics analyzed in this work. 2 Domain Transfer of Multi-Document Summarization While there is in-depth work on analyzing summarization and MDS systems and corpora ( Zhang et al. 2021 ; Mullick et al. 2024 ; Fabbri et al. 2021b ; Dey et al. 2020 ) , there is no systematic evaluation of how each model type performs when applied out-of-domain and how the automatic evaluation metrics can over- or under-represent model performance. We draw from existing models and tools to explore the intricacies of domain transfer with MDS models and their evaluation methods. The main dimensions of our analyses are domain (i.e., dataset), model type, and evaluation metrics. See Figure 2 for summary. 2.1 Summarization Dataset Domains News is a common domain for summarization datasets ( Hermann et al. 2015 ; Fabbri et al. 2019 ; Jiang and Dreyer 2024 ; Choi et al. 2024 ; Gholipour Ghalandari et al. 2020 ; Litkowski and Road 2004 ) There are datasets for other domains like scientific articles ( Lu et al. 2020 ) and conversation ( Fabbri et al. 2021a ; Gliwa et al. 2019 ; Chen et al. 2021 ) . We select a single, high-quality multi-document dataset with human-written summaries to represent each domain. All documents and summaries are in English. Conversation. We use the Reddit subset of the released data from ConvoSumm ( Fabbri et al. 2021a ) ( ConvoSumm-Reddit ). Since the summaries were written by paid crowdsource workers, this dataset is significantly smaller than the news and science datasets, with only 500 entries. News. Multi-News ( Fabbri et al. 2019 ) is a news-based MDS dataset featuring 56,216 human-written summaries from Newser. 2 2 2 newser.com Science. Representing the Science domain, the entries in Multi-XScience pair a related works section of an academic paper with 1) its abstract and 2) the abstracts from cited work. The 40,528 papers were collected from arXiv and the Microsoft Academic Graph (MAG) ( Lu et al. 2020 ) . Since the paper dump was from pre-2020, we assume all writing on arXiv is human-generated. 2.2 Multi-Document Summarization Model Types Current high-performing fine-tuned abstractive summarization methods can be grouped into four approaches: summary-specific pre-training ( Zhang et al. 2020a ; Xiao et al. 2022a ; Peper et al. 2024 ) , incremental summarization (“chunk-then-summarize”) ( Zhang et al. 2022 ) , extracting claims or viewpoints (“extract-then-summarize”) ( Fabbri et al. 2021a ; Ernst et al. 2022 ; Ouyang et al. 2023 ) , and fine-tuning or few-shot summarization with large pre-trained models ( Mullick et al. 2024 ) . These groupings have been used in prior work ( Mullick et al. 2024 ) . Modified Pretraining. Prior work has shown that fine-tuning a language model specifically pre-trained with summary-focused objectives improves downstream performance over non-specialized models ( Xiao et al. 2022b ; Zhang et al. 2020a ; Peper et al. 2024 ) . Zhang et al. 2020a first introduced this concept of summary-specific pre-training, and it was later augmented to the multi-document case ( Xiao et al. 2022b ; Peper et al. 2024 ) . We select PRIMERA to represent this group. PRIMERA is Longformer-Encoder-Decoder model initialized from BART-Large ( Beltagy et al. 2020 ) and pre-trained with an Entity Pyramid objective on a large news corpora ( Xiao et al. 2022b ) . While PELMS ( Peper et al. 2024 ) adjusts the pre-training strategy to improve upon PRIMERA, we omit comparison to this model because it also uses a data filtering step, which overlaps with the extract-then-summarize approach. Chunk-then-summarize. Another approach to MDS is incremental summarization of documents, as done with Summ N ( Zhang et al. 2022 ) . Summ N first segments the source documents and target summary and creates mappings between the source and target segments (via ROUGE max similarity), which and trains a separate BART-Large-CNN model ( Zhang et al. 2021 ) each “coarse” stages before the final “fine-grained” stage which produces the end summary. Extract-then-summarize. A different approach to the long-context problem of MDS systems is to first extract the salient portions of the documents before summarizing them. The main models in this category are COMPO ( Ouyang et al. 2023 ) , ProCluster ( Ernst et al. 2022 ) , and ConvoSumm ( Fabbri et al. 2021a ) ,. COMPO is specifically tailored toward conversational summarization and provides summaries based on extracted sub-components (topics and actions) within the dialogue. ProCluster extracts “propositions” in documents, clusters them, and then summarizes the clusters. ConvoSumm first filters sentences from documents to only the “claims” and then summarizes the claims. A variation of the extract-then-summarize pipeline is attribute-then-generate ( Slobodkin et al. 2024 ) which requires multiple steps of content selection, sentence planning, and sentence-by-sentence generation. We chose ConvoSumm over other extract-then-summarize models due to its applicability to many domains and non-overlap with different approaches with clustering or sentence-planning steps. Pre-trained LLM. While the above models are strong MDS systems, they are all the size of BART-Large (336M params.). A growing trend across natural language processing research has been to fine-tune or adapt large, pre-trained decoder language models (LLMs) through in-context learning. We select Llama-3.1-8B-Instruct to represent this approach due to its strong performance on a variety of tasks ( Dubey et al. 2024 ) . 2.3 Summarization Evaluation Most MDS papers and their surveys use off-the-shelf automatic evaluation metrics without discussing drawbacks. Summarization metrics usually measure one or more of the following attributes: similarity to a ground-truth summary, overall quality, and factuality. While many MDS systems are evaluated with manual annotations, we focus on automatic evaluation in this work. Similarity to Ground Truth. Most text evaluation metrics compare the model output to a known, quality summary or ground truth (i.e., reference). The most common metric for evaluating summary quality is ROUGE ( Lin 2004 ) but is better viewed as a similarity metric since it is the token recall of the generated summary based on the reference (i.e., ground-truth). The complement of ROUGE is BLEU ( Papineni et al. 2002 ) , which measures precision instead of recall of generated summaries. Other metrics in the spirit of ROUGE/BLEU are semantic-based instead of lexical, which do not penalize generations with similar meanings but different word choices. The most popular version is BERTScore ( Zhang et al. 2020b ) , which maps tokens from the generation to the reference before calculating overlap. An alternative is MoverScore, which “optimized” version of BERTScore ( Zhang et al. 2020b ) which uses soft token alignments from the hypothesis to the reference instead of a hard 1:1 mapping, allowing for a greater reward for semantically similar generated text. Other similarity metrics are sub-“dimensions” of a trained multi-dimensional evaluator, like UniEval. UniEval is a T5-Large model fine-tuned on synthetically augmented datasets for positive and negative examples of different “dimensions” of summary attributes (fluency, relevance, coherence, consistency) ( Zhong et al. 2022 ) . The UniEval-Relevance dimension is a similarity metric that compares the reference and generation as a whole instead of per token. Overall Quality. Not all desired qualities of a summary depend on a reference, some are intrinsic or rely on the source documents (e.g., fluency). UniEval measures both fluency and coherence. UniEval-Fluency is based solely on the generation, and UniEval-Coherence measures the coherence of the generation with respect to the source. Factuality. A growing concern with generated text is the possibility of hallucinations and non-faithfulness to the source material ( Min et al. 2023 ) . While objective factuality is important for the safety of LLMs (e.g., knowing that boiling water is hot), for MDS we typically assume that the source documents are factually correct and that a summary is only factual if its claims match the claims from the source. We measure summary factuality with UniEval-Consistency , a single-dimensional evaluator trained to measure the factual consistency of generated text to the source (as opposed to continuously trained as with UniEval-Fluency and UniEval-Coherence). A popular framework for factuality analysis is decompose-then-verify ( Min et al. 2023 ; Jiang et al. 2024 ) , and we leave this analysis to future work. We include ROUGE , MoverScore , and multiple dimensions of UniEval in our analysis as representations for similarity to ground truth, factuality, and overall quality metrics. While BARTScore ( Yuan et al. 2021 ) measures many of the same dimensions as UniEval, we chose UniEval because it is better correlated with human evaluation ( Zhong et al. 2022 ) and the potential bias of BARTScore to model generations, since most of the models are BART-based (“self bias”) ( He et al. 2023 ) . Another evaluation for summaries is source coverage, or how much of the source documents are represented in the summaries ( Wolhandler et al. 2022 ) . We leave this evaluation to future work. 3 Experimental Setup Our goal is to explore the intricacies of domain transfer with multi-document summarization models and their evaluation methods. The main dimensions of our analyses are different datasets (i.e., domains), models, and evaluation metrics ( Figure 2 ). We chose 3 dataset domains (News, Science, Conversation), 4 model types (end-to-end/direct), chunk-then-summarize, extract-then-summarize, and pre-trained LLMs), and 7 metrics for a total of 84 points of analysis ( Section 2 ). 3.1 Model Training The hyperparameters for each model and dataset combination are in Appendix B . Each model was trained on each dataset, and we refer to them as ModelName - Domain , e.g., ConvoSumm fine-tuned on Multi-News is ConvoSumm-News. PRIMERA. As input, the source documents are delineated with a special token, and the model’s global attention is set on those tokens. To handle examples with source documents that are too long for the context window, we follow the author’s code and evenly truncate the documents to fit. We used the authors’ released models of PRIMERA fine-tuned on Multi-XScience and MultiNews and trained PRIMERA-ConvoSumm-Reddit with the released author code. Summ N . We use a 3 course-stage process for all datasets. A separate model is trained for each coarse stage, leading to N+1 , or 4 fine-tuned BART models for each training dataset. An important note is due to the setup of summarizing chunks, there are no too-long context issues as with direct models like PRIMERA and all source documents are considered by the model. ConvoSumm. There are multiple versions of ConvoSumm but we select ConvoSumm-argument-filtered (“arg-filtered”) due to its comparable performance to the more complex argument-graph version. The argument filtration model is a BERT model fine-tuned on annotated data from the Change My View community on Reddit (i.e., conversation, AMPERSAND ( Chakrabarty et al. 2019 ) ) and annotated persuasive essays on a variety of topics ( Stab and Gurevych 2017 ) ). A BART-Large model with an expanded context window (2048) is fine-tuned on the extracted claims. Llama 3.1. Unlike the other models, we do not fine-tune Llama 3.1 8B-Instruct. 3 3 3 https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct To compare to the domain-transfer experiments, we provide an example from the training set to the model, i.e., for News-Convo, an example from Multi-News is provided to Llama and then the model is prompted to complete an example from ConvoSumm. We also include a zero-shot version which does not provide a complete example to Llama. 3.2 Evaluation Details ROUGE scores have been known to vary depending on the implementation used. We use the HuggingFace wrapper for the Google Research implementation. 4 4 4 https://huggingface.co/spaces/evaluate-metric/rouge We report ROUGE-1 (R1) and ROUGE-L-Sum (RL-Sum) to follow prior work. For the other evaluation metrics, MoverScore and multiple dimensions from UniEval , we use the trained models as they are without modification. Factuality Ground-Truth Similarity Quality Cons. (UE) Rel. (UE) R1 RLSum MoverScore Fluency (UE) Coh. (UE) Conversation 0.68 0.75 0.30 0.20 0.07 0.91 0.72 Avg. \Delta +18% -13% -9% -19% -31% -1% +4% Ref. 0.71 0.99 1.00 1.00 1.00 0.90 0.73 News 0.72 0.78 0.43 0.22 0.20 0.91 0.79 Avg. \Delta -7% +2% -51% -41% -92% -2% +1% Ref. 0.72 0.99 1.00 1.00 1.00 0.92 0.80 Science 0.78 0.63 0.24 0.15 0.03 0.78 0.64 Avg. \Delta -7% -12% +6% +1% -7% -1% +4% Ref. 0.55 0.99 1.00 1.00 1.00 0.66 0.43 Table 1: Multi-document summarization (MDS) model performance and relative performance aggregated by domain when trained and tested on the same domain (top) and different domains ( \Delta ), respectively. Results are aggregated across all the test splits of each dataset. A positive \Delta shows a performance gain and a negative \Delta is a performance loss as compared to the model’s scores on the in-domain data. A relative change of over 1 is possible with MoverScore because of its [-1,1] range. Detailed results are in Table 7 . “Ref” refers to the results when running the metrics on the ground-truth data. 4 Results We analyze the results in the framework of our three axes: dataset domain, model type, and evaluation metric. For ease of analysis, we show the relative performance of each model when evaluated on a new domain in Tables 2 and 1 . The relative performance is the percentage gain or loss of a model’s performance when compared to the performance on the in-domain data. The zero-shot baseline with Llama 3.1 is compared to the in-context Llama 3.1 performance for each dataset (i.e., no examples instead of one example). See Appendix C for the detailed per-model and dataset results. 4.1 Domain Generalizability When aggregated by dataset over the different model types, clear patterns emerge from Table 1 . Science summaries have the lowest general quality. The quality is lowest in the Science domain compared to Conversation and News according to Fluency and Coherence metrics, even before evaluating Science-trained MDS models on other domains, 0.78 and 0.64 , respectively. These scores are 10+ points lower than those of News ( 0.91 / 0.79 ) and Conversation ( 0.91 / 0.72 ). Fluency is affected the least by domain transfer. The fluency scores are the most stable, i.e., have the least change in performance (Avg \Delta ) compared to the other metrics. While on average there is a slight decline in fluency, as indicated by the negative change, the decline is negligible in the (-0.02,-0.01)\% range. Fluency does not appear to be an issue with these models, mostly likely due their size (BART-Large and Llama 8B). Factuality Ground-Truth Similarity Quality Cons. (UE) Rel. (UE) R1 RLSum MoverScore Fluency (UE) Coh. (UE) ConvoSumm 0.65 0.62 0.32 0.18 0.08 0.83 0.57 Avg. \Delta -2% -11% -28% -26% -152% +3% -4% PRIMERA 0.78 0.74 0.35 0.20 0.12 0.89 0.78 Avg. \Delta +3% -10% -19% -24% -16% -14% -1% Summ N 0.66 0.60 0.33 0.19 0.08 0.81 0.56 Avg. \Delta +4% -9% -32% -33% -31% +7% +18% Llama 3.1 0.82 0.91 0.29 0.18 0.11 0.94 0.94 Avg. \Delta +1% -1% +7% +4% +25% -0.00% +0.00 Llama 3.1 (0-shot) +1% -4% -5% +0% -7% -1% -1% Table 2: Multi-document summarization (MDS) model performance and relative performance aggregated by model when trained and tested on the same domain (top) and different domains ( \Delta ), respectively. Results are aggregated across all the test splits of each dataset. A positive \Delta shows a performance gain and a negative \Delta is a performance loss as compared to the model’s scores on the in-domain data. A relative change of over 1 is possible with MoverScore because of its [-1,1] range. LLama zero-shot is compared to Llama with in-domain examples. Detailed results are in Tables 5 and 6 . There is no clear general-purpose domain for transferring. Each domain has different benefits and cons to train on for transferring, as indicated by the varied relative performance scores ( \Delta ). Transferring from Conversation to News and Science can lead to greatly increased factuality, as measured by UniEval-Consistency ( +18\% ), but also a decline in Ground-Truth Similarity ( -31\% to -9\% ). Transferring from News appears to be the worst according to the similarity metrics, over -92\% according to MoverScore. The Science-trained models appear to be the most neutral overall, ranging from -12\% to +6\% across all metrics, but claiming that it is the best domain for general MDS model training is difficult because the model scores started lower than the other domains to begin with. News-trained models have the worst similarity when transferring to non-News data. There is a steep decline in performance when News-trained models are evaluated on Conversation and Science as indicated by ROUGE ( -51\% R1, -41\% RL-Sum) and MoverScore ( -92\% ). However, the Relevance score is high ( +2\% ). The discrepancy in word-mapping metrics (lexical with ROUGE and semantic with MoverScore) and overall similarity (Relevance) show that a summary can still incorporate information from the ground-truth without having an overlap in words. Conversation and News data have the most unique styles. The ConvoSumm-Reddit example from Figure 1 shows the stark difference in style between the types of summaries in ConvoSumm-Reddit and those from Multi-News. This difference in style explains the performance loss from Conversation and News domains, as the summary wording would be completely different and penalized by the ground-truth similarity metrics ( -31 to -9\% and -92 to +2\% , respectively). 4.2 MDS Model Transfer Grouping the domain transfer results by model type across all dataset domains reveals even more performance trends ( Table 2 ). Pre-trained LLM outperforms all fine-tuned smaller models. For all metrics besides similarity, Llama 3.1 outperforms the other models concerning factuality and quality. This trend holds during domain transfer, with a performance difference ranging from -1 to +25\% across all metrics. The second-best performing model across all metrics, including accuracy, is PRIMERA, indicating an advantage of directly fine-tuning a long-context model (4K context window). Zero-shot performance is the same as in-context learning with Llama 3.1. Surprisingly, Llama has a similar performance at generating summaries with (in-context learning, ICL) or without being shown an example (zero-shot). This indicates a strong bias to its pre-training versus “learning” from the example. These results could differ with the addition of more examples (at the expense of a smaller context window available for the source documents) or a larger model. ConvoSumm claim extraction is ineffective on Science. The first step in the extract-then-summarize model ConvoSumm is to filter out sentences that are not identified as claims by a small, fine-tuned model. This step removes 35\% of sentences on average from each ConvoSumm-Reddit example, 51\% from Multi-News, but less than 1\% from Multi-XScience. This is intuitive since there would be more filler sentences in conversations and news articles, but less so in scientific publications. Chunk-then-summarize is the only approach guaranteed to fit all source documents. With all models, including Llama 3.1 with a 131K context window, there were issues fitting all source documents for every example comfortably in the context windows. As mentioned in Section 3 , PRIMERA “handles” this issue by evenly dividing the context window among the source documents and truncating the ends. The assumption that the beginning of documents is more important than the end does not hold for all domains and has only been identified in News ( Dey et al. 2020 ) . While ConvoSumm was able to fit all source documents from ConvoSumm-Reddit, it did have to truncate documents from 25 samples in Multi-XScience and 699 samples from MultiNews. PRIMERA, with a context window double the size of ConvoSumm (2K vs 4K), was able to fit all Conversation and Science samples, but truncated 835 Multi-News examples. Llama only had issues with 3 samples from Multi-News. 4.3 Evaluation of Evaluation Word-based similarity scores are low across all models. For all models, the generated summaries are not very similar to the ground-truth reference according to the lexical and semantic similarity metrics ROUGE and MoverScore. The model with the highest accuracy ( 0.20 ROUGE_L and 0.12 MoverScore) was PRIMERA, higher than the much larger LM, Llama-8B. For lexical-based metrics like ROUGE, the low similarity can be explained through the model’s use of synonyms, which would not be captured through token-based recall. However, the low MoverScores can be explained by (1) the penalization of generations longer or more detailed than the reference and (2) the lack of a normalizing constant as provided by BERTScore ( Zhang et al. 2020b ) . 5 5 5 Zhang et al. 2020b use a learned constant to normalize the scores to [0,1] for easier readability. Still, low scores across both accuracy metrics indicate that the generated summaries differ from their human-written counterparts. Despite overall low accuracy scores, all models have the highest ROUGE/MoverScore performance on Multi-News (News), indicating the summaries might not be as abstract as the other datasets. Surprisingly, the Relevance score which compares the reference and generation as a whole instead of token mappings, shows higher similarity, especially for Llama ( 0.91 average). Accuracy is affected the most. The reference-based metric, accuracy, shows the most extreme gains and losses compared to the other metrics. This is intuitive because a model’s summary can still be good but not be written in the same manner, or style , as the reference, which can happen if a model is biased towards the writing style of its training data. The zero-shot baseline, which is only prompted and not given any styles, shows the least variance in accuracy compared to the other models. All models are fluent. Unsurprisingly, all models have high fluency scores ( 0.80+ ) since all models are of size BART-Large and above. While fluency is still an important attribute to check, it might not be as important as it once was. Non-reference metrics do not scale well to the multi-document setting. Evaluating for attributes that depend on the source documents (Consistency and Coherence) is most affected by the long-document nature of multi-document summarization. The options are either to 1) truncate the documents, which can greatly alter the scores depending on which documents are included, or 2) calculate the metric for every source document and aggregate the scores. More work is needed to determine whether truncating the source documents for UniEval over- or under-estimates model performance. Another option is to turn to large-context models like GPT-4 to be a “judge” for specific criteria ( Kim et al. 2024 ; Vu et al. 2024 ) . This use of LLMs, which typically can fit 10K+ tokens in their context windows, is a potential solution to this scaling issue. The reference-based metrics, i.e., similarity metrics that only compare the ground-truth and generated summaries, are not affected by this issue. Trained metrics can penalize out-of-training data. Just as summarization models can have lower performance on out-of-domain data, so can trained evaluation models like UniEval. As noted before, the fluency and coherence scores on Science are 0.10 points lower than on News and Conversation. This could be an issue with the summarization model training, but the fluency and coherence scores are also lower on the ground-truth summaries, 0.30+ points lower (“Ref.” in Table 1 ), which should not be the case as all summaries are human-written and should have high scores. An explanation is that UniEval was trained on a synthetically augmented version of news data, CNN/DailyMail, which is very different from Scientific text. 5 Discussion and Conclusion In most work that surveys or proposes multi-document summarization (MDS) systems, there is little overlap between the analysis of model performance and the shortcomings of popularly used summarization evaluation metrics. We bridged this gap through experiments and analysis of not only the zero-shot domain transfer ability of MDS systems but also their evaluation. We analyzed three components of a multi-document summarization (MDS) pipeline for domain transfer: 1) the domain of the training data, 2) the model type, and 3) the common evaluation metrics. We found important facets to consider in each component. For domain transfer, the training and test domains are important for model performance, regardless of the model type. While we did not find a clear trend between Conversation, News, and Science, it was apparent that the transfer was highly affected by the style of the summaries, albeit more than the domain itself. Summaries from different domains but written in the same style (e.g., high-level vs low-level) could transfer better with higher similarity to the reference. For different approaches to MDS, a model that didn’t hold a strong bias towards the style of its training (or example) data is ideal, as is the case with a large pre-trained decoder model, like Llama 3.1. If the domain styles differ greatly, a zero- or few-shot approach with less bias is ideal. Notable issues we uncovered with evaluation metrics that prevented the fair evaluation of summaries were domain-specific penalizations, mostly for Science, questions with scaling to the MDS case, and the inability of ground-truth similarity metrics to adjust for style change (ROUGE and MoverScore). When evaluating an MDS model on a new domain, word-similarity metrics judge style more than content similarity and trained evaluators could over- or under-judge model performance due to limited information in the context window. Limitations Despite the rise in popularity of large, pre-trained decoder models, most fine-tuned multi-document summarization models are on the scale of a few hundred million parameters, e.g., BART-Large. We did not analyze the effect of model size or scale concerning summarization performance or domain transfer abilities. Another limitation is our focus on English, as all of the datasets and learned metrics (i.e., UniEval), were in English or trained on English data. Ethical Considerations We did not evaluate the model output and training datasets for absolute factuality (i.e., correctness with respect to the world) or harmful language. We leave these tasks to future work or the incorporation of an additional step for model safety. Appendix A Dataset Details We include three MDS datasets: Multi-News ( Fabbri et al. 2019 ) , Multi-XScience ( Lu et al. 2020 ) ), and ConvoSumm-Reddit ( Fabbri et al. 2021a ) . We keep the original train, validation, and test splits as the original papers for all datasets. See Table 3 for an example from each dataset and Table 4 for summary statistics. Multi-News. For Multi-News we used the ‘‘cleaned’’ version provided by the authors 6 6 6 https://github.com/Alex-Fabbri/Multi-News/issues/36 with 44964/5621/5621 examples for train, validation, and test sets, respectively. We keep the NEWLINE_CHAR symbol in the documents. Multi-XScience. The Multi-XScience dataset had many examples with missing source documents. In addition to skipping those with no source documents, we also removed non “multi” examples which had less than two source documents, resulting in 11523/1937/1901 examples for train, validation, and test sets, respectively. We skipped 18846/3129/3192 examples from train/val/test sets with less than two source documents (i.e., abstracts) and 20023/3383/3403 examples from train/val/test with completely missing source documents. We also replace all document-specific citations cite_5 with a more generalizable cite . ConvoSumm. While ConvoSumm features multiple conversation datasets, we focus on Reddit. Similar to the original work, we flatten comment threads as individual source documents, resulting in 201/50/250 examples for train, validation, and test sets, respectively. We keep the |NEWLINE| symbols in the documents. Appendix B Training Details Unless specified otherwise, each model was trained on a single Nvidia A100 GPU with 80GB of memory. The training hyperparameters for PRIMERA, Summ N , and ConvoSumm-arg-filtered are in Tables 10 , 9 and 11 , respectively. Below are the prompts and details for running Llama for the zero-shot and in-context 1-shot comparisons. PRIMERA The authors’ released models were from HuggingFace fine-tuned on Multi-XScience ( https://huggingface.co/allenai/PRIMERA-multixscience ) and MultiNews ( https://huggingface.co/allenai/PRIMERA-multinews ). The authors’ code is on GitHub ( https://github.com/allenai/PRIMER ). Llama 3.1 8B-Instruct. Due to the large size of Llama 8B, we hosted it as a service with vLLM on 4 Nvidia A100 80GB GPUs ( Kwon et al. 2023 ) . 7 7 7 https://docs.vllm.ai We followed model examples on how to interact for the summarization task ( Table 12 ). Appendix C Raw Results The domain-transfer results in Section 4.2 show the summarized relative performance of the models. The detailed performance across domains and models are in Tables 7 and 6 , respectively. The raw, un-normalized numbers are shown in Table 8 . Dataset Source Documents Summary Multi-News Whether a sign of a good read; or a comment on the ’pulp’ nature of some genres of fiction, the Oxfam second-hand book charts have remained in The Da Vinci Code author’s favour for the past four years. NEWLINE_CHAR NEWLINE_CHAR Dan Brown has topped Oxfam’s ’most donated’ list again, his fourth consecutive year … A charity shop is urging people to stop donating The Da Vinci Code after becoming overwhelmed with copies. NEWLINE_CHAR NEWLINE_CHAR The Oxfam shop in Swansea has been receiving an average of one copy of the Dan Brown novel a week for months, leaving them with little room for any other books. NEWLINE_CHAR NEWLINE_CHAR Staff who are struggling to sell copies of the book have put a note up in the store saying they would rather donors hand in their vinyl instead. –The Da Vinci Code has sold so many copies—that would be at least 80 million—that it’s bound to turn up in book donation piles. But at one charity shop in the UK, it’s been donated so heavily that the shop has posted a sign propped up on a tower of Da Vinci Code copies that reads: Ÿou could give us another Da Vinci Code… but we would rather have your vinyl!T̈he manager of the Oxfam shop in Swansea tells the Telegraph that people are laughing and taking pictures of the sizable display: Ï would say that we get one copy of the book every day.Ḧe says people buy them öccasionally,b̈ut with vinyl sales up 25% in the past year, they’d rather take records. Dan Brown’s book isn’t the only one that shops like Oxfam struggle to re-sell. Last year, Oxfam was hit with a large and steady supply of Fifty Shades of Grey, and it similarly begged donors: P̈lease—no more.B̈ut Brown has a particular kind of staying power. The Da Vinci Code was published in 2003, and within six years Brown had booted John Grisham from the No. 1 slot on the list of writers whose books were most often donated to Oxfam’s 700 shops, reported the Guardian at the time. The Independent in 2012 reported Brown’s best-seller was the most-donated book for the fourth year running. (See why Dan Brown took heat from the Philippines.) Multi-XScience Despite the huge amount of recent research efforts on entity resolution (matching) there has not yet been a comparative evaluation on the relative effectiveness and efficiency of alternate approaches. We therefore present such an evaluation of existing implementations on challenging real-world match tasks. We consider approaches both with and without using machine learning to find suitable parameterization and combination of similarity functions. … This tutorial provides a comprehensive and cohesive overview of the key research results in the area of record linkage methodologies and algorithms for identifying approximate duplicate records, and available tools for this purpose. It encompasses techniques introduced in several communities including databases, information retrieval, statistics and machine learning. It aims to identify similarities and differences across the techniques as well as their merits and limitations. For accurate entity linking and deduplication, early works on the subject examined many methods such as cosine similarity match, distance based match, TF IDF, and Soundex. The well known similarity measures for entity linking are summarized and reviewed in cite ; and the work in cite presents a comparative evaluation of some existing works. ConvoSumm Title: Hand from tonight. Subreddit: poker Playing at my local pub league. Villian is loose aggressive. Sorry if this is bad formatting just quickly whipped it up for your opinions. |NEWLINE| Button: ( 30000) |NEWLINE| SB: Villian: (50000) … … Shove pre but I imagine this is the sort of game where you get told off for doing that A lot of commenters debate the semantics and word choice of another comment. Some commenters say that the poster should shove even though other players will not like that move very much. Other commenters discuss the differences of some poker terminology. Table 3: Examples from each of the three datasets: Multi-News, Multi-XScience, and ConvoSumm-Reddit. # Examples # Source Docs. Summary Length Avg. Doc. Length ConvoSumm (Reddit) Test 250 14.42 (5.55) 57.78 (12.04) 53.02 (26.60) Train 201 14.98 (6.28) 58.32 (12.95) 49.46 (26.38) Val 50 16.74 (7.21) 59.18 (14.99) 52.88 (27.29) MultiNews Test 5621 2.73 (1.08) 217.00 (67.20) 982.98 (1779.55) Train 44964 2.76 (1.12) 218.27 (68.04) 982.95 (1851.20) Val 5621 2.70 (1.04) 216.72 (67.67) 980.06 (1163.94) MultiXScience Test 1901 3.55 (1.95) 106.35 (41.60) 195.32 (48.01) Train 11523 3.58 (2.05) 106.82 (42.14) 197.03 (47.91) Val 1937 3.55 (1.99) 104.03 (41.38) 196.98 (48.84) Table 4: Statistics for each dataset used in this work: Multi-News, Multi-XScience, and ConvoSumm-Reddit. The number of tokens in the (ground truth) summary and documents are measured with the BART tokenizer. Since we remove examples with missing information, our numbers might not match those in the original dataset papers ( Appendix A ). Domain Factuality Accuracy Quality Train/Test Cons. (UE) Rel. (UE) RL-Sum MoverScore Fluency (UE) Coh. (UE) ConvoSumm Convo 0.64 0.70 0.21 0.06 0.91 0.60 News 0.64 0.62 0.19 0.17 0.89 0.66 Science 0.67 0.55 0.15 0.01 0.69 0.46 All 0.65 0.62 0.18 0.08 0.83 0.57 PRIMERA Convo 0.65 0.73 0.21 0.09 0.92 0.71 News 0.74 0.82 0.25 0.21 0.93 0.78 Science 0.96 0.68 0.14 0.05 0.83 0.86 All 0.78 0.74 0.20 0.12 0.89 0.78 Summ N Convo 0.65 0.63 0.21 0.07 0.89 0.62 News 0.69 0.78 0.21 0.20 0.89 0.80 Science 0.65 0.39 0.15 -0.02 0.66 0.27 All 0.66 0.60 0.19 0.08 0.81 0.56 Llama 3.1 Convo 0.80 0.92 0.14 0.07 0.93 0.94 News 0.82 0.90 0.19 0.19 0.94 0.93 Science 0.85 0.90 0.12 0.08 0.94 0.95 All 0.83 0.91 0.13 0.08 0.93 0.95 Table 5: Multi-document summarization (MDS) model performance when trained and tested on the same domain . The results for “All” are averaged across all domains for each model. Results are on the test splits of each dataset. Metrics followed by (UE) are from UniEval. For all metrics a greater score is better. Domain Factuality Accuracy Quality Train Test Fact. (UE) Rel. (UE) R1 RLSum MoverScore Fluency (UE) Coh. (UE) ConvoSumm Convo News 0.05 -0.41 -0.37 -0.44 -1.67 0.00 -0.25 Convo Science 0.06 -0.09 -0.25 -0.33 -0.72 -0.01 0.02 News Convo -0.11 0.30 -0.46 -0.31 -1.12 0.02 0.19 News Science -0.07 0.10 -0.41 -0.28 -0.90 0.05 0.02 Science Convo -0.07 -0.30 -0.16 -0.10 -2.79 0.03 -0.13 Science News -0.01 -0.28 -0.01 -0.08 -1.90 0.07 -0.07 All 0.01 -0.02 -0.11 -0.28 -0.26 -1.52 0.03 -0.04 PRIMERA Convo News 0.20 -0.31 -0.05 -0.22 -0.22 -0.09 -0.08 Convo Science 0.48 -0.07 -0.24 -0.36 -0.47 -0.10 0.20 News Convo -0.22 0.00 -0.61 -0.57 -1.03 -0.07 0.00 News Science 0.29 -0.17 -0.48 -0.45 -0.78 -0.11 0.09 Science Convo -0.47 -0.12 -0.20 -0.13 -0.74 -0.27 -0.25 Science News -0.11 0.04 0.43 0.30 2.25 -0.21 -0.03 All 0.03 -0.10 -0.19 -0.24 -0.16 -0.14 -0.01 Summ N Convo News 0.31 -0.08 -0.22 -0.30 -0.67 0.04 0.14 Convo Science 0.26 -0.02 -0.31 -0.35 -0.79 0.04 0.31 News Convo -0.32 0.05 -0.69 -0.60 -1.22 -0.00 -0.00 News Science -0.17 -0.21 -0.52 -0.41 -1.00 -0.01 -0.25 Science Convo -0.00 -0.34 -0.59 -0.51 5.56 0.05 0.14 Science News 0.17 0.03 0.37 0.17 -3.74 0.26 0.72 All -0.06 0.04 -0.09 -0.32 -0.33 -0.31 0.07 0.18 Llama 3.1 Convo News 0.04 -0.04 0.69 0.42 1.74 0.01 -0.01 Convo Science 0.07 -0.03 0.05 0.03 0.29 -0.00 0.03 News Convo -0.02 0.04 -0.48 -0.36 -0.72 -0.01 0.03 News Science 0.05 -0.00 -0.42 -0.30 -0.56 -0.01 0.04 Science Convo -0.06 0.02 -0.10 -0.05 -0.34 -0.01 -0.02 Science News -0.03 -0.05 0.70 0.49 1.11 0.01 -0.05 All 0.01 -0.01 0.07 0.04 0.25 -0.00 0.00 Llama 3.1 (0-shot) zero-shot Convo -0.01 -0.03 -0.16 -0.07 -0.23 -0.01 -0.02 zero-shot News 0.02 -0.07 -0.01 0.06 -0.01 -0.00 -0.03 zero-shot Science 0.01 -0.01 0.01 0.02 0.01 -0.01 0.01 All 0.01 -0.04 -0.05 0.00 -0.07 -0.01 -0.01 Table 6: Multi-document summarization (MDS) model relative performance when trained and tested on different domains . Results are on the test splits of each dataset. A positive number shows a performance gain and a negative number is a performance loss as compared to the model’s scores on the in-domain data. A relative change of over 1 is possible with MoverScore because of its [-1,1] range. Domain Factuality Accuracy Quality Train Test Fact. (UE) Rel. (UE) R1 RLSum MoverScore Fluency (UE) Coh. (UE) Convo News 0.15 -0.21 0.01 -0.14 -0.20 -0.01 -0.05 Science 0.22 -0.05 -0.19 -0.25 -0.42 -0.02 0.14 All 0.18 -0.13 -0.09 -0.19 -0.31 -0.01 0.04 News Convo -0.17 0.10 -0.56 -0.46 -1.02 -0.02 0.05 Science 0.03 -0.07 -0.46 -0.36 -0.81 -0.02 -0.03 All -0.07 0.02 -0.51 -0.41 -0.92 -0.02 0.01 Science Convo -0.15 -0.18 -0.26 -0.20 0.42 -0.05 -0.06 News 0.01 -0.06 0.37 0.22 -0.57 0.03 0.14 All -0.07 -0.12 0.06 0.01 -0.07 -0.01 0.04 zero-shot Convo -0.01 -0.03 -0.16 -0.07 -0.23 -0.01 -0.02 News 0.02 -0.07 -0.01 0.06 -0.01 -0.00 -0.03 Science 0.01 -0.01 0.01 0.02 0.01 -0.01 0.01 All 0.01 -0.04 -0.05 0.00 -0.07 -0.01 -0.01 Table 7: Multi-document summarization (MDS) model relative performance when trained and tested on different domains . The results are averaged across all model types. Results are on the test splits of each dataset. Domain Factuality Accuracy Quality Train Test Cons. (UE) Rel. (UE) R1 RLSum MoverScore Fluency (UE) Coh. (UE) ConvoSumm Convo News 0.68 0.41 0.20 0.12 -0.04 0.92 0.45 Convo Science 0.68 0.63 0.23 0.14 0.02 0.91 0.61 News Convo 0.57 0.81 0.21 0.13 -0.02 0.90 0.78 News Science 0.60 0.69 0.24 0.13 0.02 0.93 0.67 Science Convo 0.63 0.38 0.21 0.14 -0.02 0.71 0.40 Science News 0.67 0.39 0.25 0.14 -0.01 0.73 0.43 PRIMERA Convo News 0.78 0.50 0.31 0.16 0.07 0.84 0.65 Convo Science 0.96 0.68 0.25 0.14 0.05 0.83 0.86 News Convo 0.58 0.82 0.18 0.11 -0.01 0.86 0.79 News Science 0.96 0.68 0.25 0.14 0.05 0.83 0.86 Science Convo 0.51 0.60 0.20 0.12 0.01 0.60 0.64 Science News 0.86 0.71 0.35 0.18 0.15 0.65 0.83 Summ N Convo News 0.85 0.58 0.24 0.15 0.02 0.93 0.71 Convo Science 0.82 0.62 0.22 0.14 0.01 0.92 0.81 News Convo 0.47 0.82 0.13 0.08 -0.05 0.89 0.80 News Science 0.57 0.62 0.21 0.12 -0.00 0.89 0.60 Science Convo 0.64 0.26 0.10 0.07 -0.16 0.69 0.31 Science News 0.76 0.40 0.33 0.18 0.07 0.83 0.47 Llama 3.1 Convo News 0.83 0.89 0.41 0.22 0.18 0.94 0.93 Convo Science 0.85 0.90 0.26 0.16 0.09 0.93 0.97 News Convo 0.80 0.94 0.21 0.14 0.05 0.93 0.95 News Science 0.86 0.90 0.24 0.15 0.08 0.94 0.96 Science Convo 0.80 0.93 0.21 0.14 0.05 0.93 0.93 Science News 0.82 0.86 0.39 0.23 0.18 0.94 0.91 Llama 3.1 (0-shot) zero-shot Convo 0.79 0.90 0.21 0.15 0.05 0.93 0.92 zero-shot News 0.83 0.84 0.40 0.23 0.19 0.94 0.90 zero-shot Science 0.86 0.90 0.23 0.16 0.08 0.92 0.96 Table 8: Multi-document summarization (MDS) model performance when trained and tested on the different domains. Results are on the test splits of each dataset. Metrics followed by (UE) are from UniEval. For all metrics a greater score is better. Stage Setting Multi-News ConvoSumm-Reddit Multi-XScience General total number of updates 30000 5000 - warmup updates 500 200 - learning rate 3e-5 - - max input sequence 1024 - - max tokens (in batch) 2048 - - update frequency 4 - - patience 2 - - label smoothing 0.1 - - dropout 0.1 - - attention dropout 0.1 - - weight decay 0.01 - - optimizer Adam betas=(0.9, 0.999), eps=1e-8 - - LR scheduler polynomial decay - - Gradient clipping 0.1 - - Stage 1 max input sequence 1024 - - beam 1 - - length penalty 2 1 - max length 450 - - min length 150 0 - no-repeat-ngram-size 3 0 - Stage 2 max input sequence 1024 - - beam 1 - - length penalty 2 1 - max length 550 - - min length 150 0 - no-repeat-ngram-size 3 0 - Stage 3 max input sequence 1024 - - beam 1 - - length penalty 2 1 - max length 900 1024 - min length 400 0 - no-repeat-ngram-size 3 0 - Table 9: Training details for Summ N on all datasets. Weights initialized from BART-Large-CNN. An entry of “-” indicates the same setting as the previous column. Setting Multi-News* ConvoSumm-Reddit Multi-XScience* max input sequence 4096 4096 4096 total number of steps 25K 5K 20K warmup steps 2.5K 200 2K optimizer Adam Adam Adam learning rate 3e-5 3e-5 3e-5 batch size 16 16 16 seed 0 42 0 label smoothing 0.1 0.1 0.1 max length 1024 1024 1024 min length 0 0 0 beam 1 1 1 Table 10: Training details for PRIMERA on all datasets. Models marked with (*) were trained by ( Xiao et al. 2022b ) and only available training information is filled in. Weights initialized from BART-Large. Setting Multi-News ConvoSumm-Reddit* Multi-XScience max input sequence 2048 2048 2048 max tokens (in batch) 2048 - - update frequency 1 1 1 total number of updates 30000 200 200 warmup updates 500 20 20 learning rate 3e-5 3e-5 3e-5 label smoothing 0.1 0.1 0.1 dropout 0.1 - - attention dropout 0.1 - - weight decay 0.01 - - optimizer Adam betas=(0.9, 0.999), eps=1e-8 - - LR scheduler polynomial decay - - Gradient clipping 0.1 - - max length 1024 1024 1024 min length 0 0 0 beam 1 1 1 length penalty 1 1 1 Table 11: Training details for ConvoSumm on all datasets. Models marked with (*) were trained by ( Fabbri et al. 2021a ) and only available training information is filled in. Weights initialized from BART-Large. An entry of “-” indicates the same setting as the previous column. Zero-shot <|begin_of_text|><|start_header_id|> user <|end_header_id|> Write a summary that incorporates all of the following documents: Document: {source document 1} … Document: {source document n } <|eot_id|><|start_header_id|> assistant <|end_header_id|> In-context Learning <|begin_of_text|><|start_header_id|> user <|end_header_id|> Write a summary that incorporates all of the following documents: Document: {source document 1} … Document: {source document n } <|eot_id|><|start_header_id|> assistant <|end_header_id|> {target summary} <|begin_of_text|><|start_header_id|> user <|end_header_id|> Write a summary that incorporates all of the following documents: Document: {source document 1} … Document: {source document n } <|eot_id|><|start_header_id|> assistant <|end_header_id|> Table 12: Prompt templates for Llama 3.1 8B Instruct for the zero-shot (top) and single-shot in-context learning tasks (bottom).