Abstractive Text Summarization for Bangla Language Using NLP and Machine Learning Approaches

Asif Ahammad Miazee, Tonmoy Roy, Md Robiul Islam, Yeamin Safat · 2025

Read Abstractive Text Summarization for Bangla Language Using NLP and Machine Learning Approaches by Asif Ahammad Miazee, Tonmoy Roy, Md Robiul Islam, Yeamin Safat (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-15051.txt.

Introduction

Asif Miazee and three co-authors build a neural network that sums up text in the Bangla language, in a 2025 paper. Summarizing means shrinking a long text to its main ideas in short form. People spend much time reading news to stay informed. Not all of that reading is relevant or needed, the authors note.

Their model aims to produce brief, plain paragraphs from Bangla text. The goal is greater stability and efficiency than past methods. The paper frames the work within language processing and machine learning. Such tools can help readers scan news and documents faster. They may also serve other Bangla language tasks where short forms help. The authors present the model as a step toward practical Bangla text tools.

Catalog record

CREATORAsif Ahammad Miazee, Tonmoy Roy, Md Robiul Islam, Yeamin Safat
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE11 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-15051

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Abstractive Text Summarization for Bangla Language Using NLP and Machine Learning Approaches Asif Ahammad Miazee, Tonmoy Roy, Md Robiul Islam, Yeamin Safat Abstract Text summarization involves reducing extensive documents to short sentences that encapsulate the essential ideas. The goal is to create a brief summary that effectively conveys the main points of the original text. We spend a significant amount of time each day reading the newspaper to stay informed about current events both domestically and internationally. While reading newspapers enriches our knowledge, we sometimes come across unnecessary content that isn’t particularly relevant to our lives. In this paper, we introduce a neural network model designed to summarize Bangla text into concise and straightforward paragraphs, aiming for greater stability and efficiency. I Introduction Text or document summarization refers to the process of reducing a long document into concise sentences that capture the main ideas and essential information.Algorithmic summarization has become essential in our daily lives, as it reduces the effort and time required to obtain concise and relevant summaries that capture the essential information of documents. Summarization methods can be divided into two main types according to how they select and organize content: extractive and abstractive approaches. Extractive summarization entails selecting the key sentences from a text using certain features and combining them to form a summary, akin to highlighting text with a highlighter. Conversely, abstractive summarization generates new sentences that encapsulate the most crucial information, similar to a person writing a summary from their thoughts. Currently, there are machine learning-based summarization tools available. However, finding language-specific models poses a significant challenge. Although significant efforts have been made in Bengali extractive summarization, there are few instances of Bengali abstractive summarization [ 1 ] . Most existing works rely on basic machine learning techniques and face limitations such as small datasets. The absence of established datasets has impeded significant progress in encoder-decoder based summarization systems. Therefore, the most difficult aspect for Bengali Abstractive Summarization (BANS) is the preparation of a well-organized and sanitized dataset. II Related Works Numerous methodologies for abstractive text summarization have been identified. Yeasmin et al. [ 2 ] have provided a comprehensive review of various abstractive techniques. Consequently, Our focus was directed towards our research regarding methods for generating abstractive text summaries specifically for the Bengali language. Haque et al. (2020) provide a comprehensive overview of 14 methods for summarizing Bengali text encompass both extractive and abstractive approaches [ 3 ] . In 2004, Islam et al. [ 4 ] were trailblazers in Bengali extractive summarization, using document indexing combined with keyword-based information retrieval methods. Following this, Uddin et al. [ 5 ] adapted techniques from Extractive text summarization for Bengali texts using English as the intermediate language. In 2010, Das et al. [ 6 ] Das et al. (2010) utilized methods such as theme identification and PageRank algorithms for extractive summarization. Kamal Sarkar first proposed Bengali extractive summarization based on sentence ranking and stemming processes [ 7 ] , with further improvements by Efat et al. [ 8 ] . and Haque et al. suggested an extractive approach based on key phrases [ 9 ] and a method for ranking sentences through pronoun replacement [ 10 ] . In 2017, several notable techniques for Bengali [ 11 ] extractive summarization emerged, including a Abujar et al. [ 12 ] proposed a heuristic approach, Akther et al. [ 13 ] employed a K-means clustering method, and Chowdhury et al. utilized the Latent Semantic Analysis (LSA) technique. [ 14 ] . Ghosh et al. [ 15 ] were the first to use a graph-based sentence scoring feature for Bengali summarization. Additionally, Sarkar et al. [ 16 ] and Ullah et al. [ 17 ] have put forward extractive approaches based on term frequency and cosine similarity, respectively. Recently, Munzir et al. [ 18 ] initiated a deep neural network-based approach for Bengali extractive summarization. Abujar et al. [ 19 ] introduced Word2Vec-based word embedding for Bengali text summarization. Talukder et al. [ 20 ] then proposed an abstractive approach using bi-directional RNNs with LSTM at the encoder and attention mechanisms at the decoder. Our study also employs an LSTM-RNN based attention model similar to Talukder et al. [ 20 ] , but we apply attention to both the encoder and decoder layers and conduct a comparative study with the corresponding results and datasets. Additionally, Abujar et al. [ 21 ] introduced an LSTM-RNN-based text generation method for creating abstractive text summaries in Bengali [ 22 ] . Our system utilizes the concepts Lopyrev et al. [ 23 ] discussed this in their work. We adopt the seq2seq model and LSTM encoder-decoder architecture based on the frameworks introduced by Sutskever et. al. [ 24 ] and Bahdanau et. al. [ 25 ] , respectively. Additionally, the attention techniques for the components comprising the encoder and decoder based on the concepts from Luong et al. [ 26 ] and Rush et al. [ 27 ] have also explored. We used a special computer program to understand how sentences are built, like Vinyals and his team did.We used a special computer program to understand how sentences are built, like Vinyals and his team did [ 28 ] . III Methodology This section delineates the methodology proposed for a Bengali news summarizer employing a neural network-based approach. III-A DataSet The presence of a standardized dataset is crucial for enabling effective text summarization. Figure 1: Dataset Overview Inspired by Hermann et al. [ 29 ] and drawing insights from established English datasets such as the CNN-Daily Mail dataset, we conceived the idea of creating a standardized dataset [ 30 ] . Given the scarcity of publicly available datasets for Bengali summarization, our focus turned to the online news portal bangla.bdnews24.com, which offered a repository of news articles and corresponding summaries [ 31 ] . Employing a custom crawler, we collected 19,352 articles and summaries spanning various domains including sports, politics, and economics. Online news sources often feature extraneous content such as advertisements, non-Bengali words, and hyperlinks. To mitigate this, we developed a data cleaning program to systematically filter out irrelevant information from our dataset. TABLE 1: Statistics of the Dataset Total Articles 19,096 Number of Summaries 19,096 Maximum words in an article 76 Maximum words in a summary 12 Minimum words in an article 5 Minimum words in a summary 3 III-B Pre-processing and cleaning Preprocessing steps are applied to cleanse our dataset, underscoring the pivotal role of cleaning in enhancing outcomes before model training. These steps encompass the removal of extraneous spaces, unnecessary characters, and similar adjustments. Subsequently, start and end tokens are incorporated to delineate the beginning and conclusion of sentences. III-C Model Architecture Inspired by the impressive results achieved by the LSTM encoder-decoder with attention mechanism as reported in the study by Lopyrev et al. [20], we have developed a similar neural attention model architecture for our research. Our model comprises both LSTM Encoder and LSTM Decoder components, each integrated with attention mechanisms. TensorFlow’s embedding layer, embedding_attention_seq2seq was employed to numerically represent words for input into the encoders. Following the generation of the decoder’s output, a comparison between the actual and predicted summaries was conducted using the softmax loss function,

Read the complete text