Text summarization is one of the well-known issues in natural language processing (NLP) task in recent years. A combination of the term frequency-inverse document frequency (TF-IDF) with a dimension reduction technique named as singular value decomposition (SVD) has shown promising results for extractive text summarization based on diferent Indian languages. Our main goal is to produce an extractive summary of a text document that is succinct, fluid, and stable. In this regard, we have used the Indian Language Summarization (ILSUM)-2024 datasets, which is the third additional task shared by the Forum for Information Retrieval Evaluation (FIRE-2024). Our team, Sangita_NIT_Patna, achieved third place for Bengali and Gujarati languages in task 1. For Hindi and Telugu languages, we secured fourth place, and for Tamil, we ranked fifth. We have used article descriptions as our input data and generated a simple summary of that article description as an output.
task, we developed a single document extractive text summarization framework using TF-IDF and SVD techniques for various Indian languages.
The remaining of the paper is formatted as follows. Section 2 provides a synopsis of the related works. Section 3 presents our proposed framework for ILSUM-2022. Section 4 presents the proposed systems discovery and analysis of the results. Finally, in Section 5, we conclude the paper.
Text summarization is a vibrant research area in Natural Language Processing (NLP), with a focus
on automatically condensing text into concise summaries. This article provides an overview of the
numerous studies conducted in the area of text summarization, such as Kumar et al. [
In this section, we discussed the methodology and datasets. We proposed an extractive text summarization framework for varioues Indian languagess. We will explain each step in detail in the following section, and the overall architecture is shown in Figure 1. So, the proposed model generates multi-sentence summaries.
Datasets m a r g n h it w F D IF T )
D leu (SV a n raV iito l s ug op n m iS co e D
Add these features for sentence ranking
Top n% selected sentences for summay
To evaluate the our model, we utilized the ILSUM-2024 datasets provided by FIRE-2024 [
We preprocessed the dataset by removing the missing and duplicates values from the “Article" or “Summary" columns. We then tokenized the article into sentences and removed punctuation and empty strings to prepare the text for further processing.
In this section, we used TF-IDF technique to respresent the sentence in the vector form of the article. Here, the TF-IDF vectorizer is configured to extract features from the text data using a range of ngrams, including single words (unigrams), two-word combinations (bigrams), three-word combinations (trigrams), and four-word combinations (4-grams). By selecting the top 2000 most frequent terms across all documents, the TF-IDF matrix is truncated, reducing its dimensionality and focusing on the most important terms. This helps conserve memory and boost computational eficiency. It combines two measures: 1. Term Frequency (TF): Term Frequency (TF) is a numerical measure that represents how frequently a term appears in a given document. Here, represents a word (or -gram, which could be a single word or a sequence of words), and represents a specific document. This score increases with the frequency of a word in a document but doesn’t consider whether the word is common across other documents in the corpus. The TF score indicates the relative importance of a term within a document by measuring its occurrence. For a specific -gram g in a document , the TF is calculated as: where
TF(, ) = ∑︀′∈ ′,
, • , is the frequency of -gram in the document . • is the set of all -grams in the document .
• ∑︀′∈ ′, is the total count of all -grams in the document. 2. Inverse Document Frequency (IDF): Measures how common or rare a word is across all documents in the corpus. A word that appears in many documents will have a lower IDF score. If a term appears in almost every document, its IDF score will be close to zero, meaning it has less unique significance. The inverse document frequency of -gram is:
IDF() = log ︂(
)︂ 1 + containing the -gram . scores: where, is the total number of documents in the corpus, is the number of documents 3. TF-IDF Calculation: The TF-IDF score for a term in a document is the product of its TF and IDF - (, ) = TF(, ) × IDF()
Words with high TF-IDF scores are considered important or unique to that document compared to other documents in the corpus. (1) (2) (3)
SVD is a mathematical technique used in linear algebra for decomposing a matrix into three other matrices. SVD is used in Latent Semantic Analysis (LSA) to uncover relationships between terms and documents by reducing dimensionality in text data. SVD can reduce the dimensionality of feature vectors for sentences in the article, produced by the TF-IDF technique, while preserving the essential features and relationships in the article.
= ∑︁ (4) • A is a matrix of dimension M× N. • U: × matrix of the orthonormal eigenvectors of . • ∑︀: diagonal matrix with r elements equal to the root of the positive eigenvalues of or . • : transpose of a × matrix containing the orthonormal eigenvectors of .
The sums the rows of , giving a score for each sentence based on its importance.
In this step, we prioritize the sentences in the article, ranking them in descending order of importance. This produces a list of ranked sentences, with the most important sentences at the top and the less significant ones at the bottom.
To generate the summary, we first calculate the sentence count by taking =15% of the total sentences. We then combine the top-ranked sentences up to this count to create the summary.
In this section, we present the evaluation metrics used to assess the performance of our proposed approach, followed by a detailed discussion of the results obtained.
In this study, we utilized the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) [17] and BERTScore (B) metrics to assess the performance of our model. ROUGE (R) measures the quality of the generated summaries by counting the number of overlapping lexical units between the generated and reference summaries. Unlike R, which relies on exact word or phrase matches, B leverages pretrained contextual embeddings from BERT (Bidirectional Encoder Representations from Transformers) to calculate the semantic similarity between the generated and reference summaries. ROUGE and BERTScore includes precision (Pre), recall (Rec), and F1 score as part of its evaluation metrics. Here, we employed R-N (with N=1,2,4) and R-L (Longest Common Subsequence) to compute the R-1, R-2, and R-L scores based on Pre, Rec and F1 for both the training and validation datasets. For the testing dataset, R-N (with N=1, 2, and 4) and R-L were evaluated based on the F1-score, while B was evaluated using Pre, Rec, and F1-score.
In this section, we discussed about the results obtained on Training, validation and test datasets provided
by ILSUM-2024 [18] [
English results in Table 4 and Table 6, respectively, for Task 1. In the Hindi and Telugu languages category, we secured the 4ℎ position and present the corresponding results in Table 5 and Table 8, respectively, for Task 1. Similarly, for the Tamil language, we achieved the 5ℎ position and show the results in Table 7. In this study, TF-IDF was employed to represent text data by assigning weights to terms based on their importance within the corpus. This method proved efective in reducing the influence of high-frequency, low-relevance terms, resulting in a more meaningful feature space. Singular value decomposition further enhanced the feature set by capturing latent semantic relationships and reducing dimensionality, thereby improving computational eficiency and model generalization.
In this work, we applied TF-IDF and SVD techniques for extractive text summarization in various Indian languages. These findings ofer valuable insights for higher quality summaries, reduction of redundancy, more coherent summaries, when used together, TF-IDF and SVD create summaries that better capture both the key terms (from TF-IDF) and latent concepts (from SVD), producing summaries that are both relevant and coherent. While these traditional methods lack the contextual embeddings provided by deep learning techniques, their simplicity and interpretability make them valuable tools, particularly for resource-constrained applications. Singular value decomposition helps reduce redundancy by identifying overlapping information within sentences. It also captures semantic relationships between words and sentences. However, TF-IDF doesn’t consider sentence structure or semantics, so the resulting summary may miss out on coherence, as it focuses purely on the frequency and rarity of terms. Singular value decomposition can be computationally expensive, especially for large documents. Also, it may struggle with small texts where latent structures are harder to detect.
Several areas can be explored to further enhance the extractive text summarization process for Indian languages. Future work could explore hybrid approaches combining TF-IDF and SVD with contextual word embeddings for improved performance. One promising direction is the integration of more advanced techniques, such as transformer-based architectures (e.g., BERT or GPT), which can capture deeper semantic understanding and sentence structure beyond what TF-IDF and SVD ofer.
This first author would want to acknowledge the Ministry of Education (MOE), Government of India for ifnancial support during the research work through the Rajiv Gandhi fellowship Ph.D scheme (UGC) for computer science & engineering.
The authors confirm that no generative AI tools were used in the writing, editing, or analysis processes of this manuscript. All content was created and reviewed by the authors. [14] V. Ilanchezhiyan, R. Darshan, E. M. Dhitshithaa, B. Bharathi, Text summarization for indian languages: Finetuned transformer model application., in: FIRE (Working Notes), 2023, pp. 766–774. [15] S. Gupta, S. Pal, Named entity-aware abstractive text summarization for hindi language., in: FIRE (Working Notes), 2023, pp. 755–765. [16] S. Satapara, P. Mehta, S. Modha, D. Ganguly, Indian language summarization at FIRE 2023, in: D. Ganguly, S. Majumdar, B. Mitra, P. Gupta, S. Gangopadhyay, P. Majumder (Eds.), Proceedings of the 15th Annual Meeting of the Forum for Information Retrieval Evaluation, FIRE 2023, Panjim, India, December 15-18, 2023, ACM, 2023, pp. 27–29. URL: https://doi.org/10.1145/3632754.3634662. doi:10.1145/3632754.3634662. [17] F. Liu, Y. Liu, Exploring correlation between rouge and human evaluation on meeting summaries,
IEEE Transactions on Audio, Speech, and Language Processing 18 (2009) 187–196. [18] S. Satapara, P. Mehta, S. Modha, A. Hegde, S. HL, D. Ganguly, Key insights from the third ilsum track at fire 2024, in: Proceedings of the 16th Annual Meeting of the Forum for Information Retrieval Evaluation, FIRE 2024, Gandhiinagar, India. December 12-15, 2024, ACM, 2024.