Summarizing Business News: Evaluating BART, T5, and PEGASUS for Effective Information Extraction

ABSTRACT


INTRODUCTION
The demand for precise and effective information extraction and condensation techniques has never been greater than it is today, in an era of information overload where massive amounts of text are produced every day throughout the digital landscape.This problem is addressed by text summarizing, a crucial area of natural language processing [1] (NLP), which creates succinct, coherent summaries from extensive texts automatically.In order to facilitate the extraction of important information and the creation of summaries that are human-like, this research project aims to take advantage of the capabilities of cutting-edge pre-trained models, such as BART [2] (Bidirectional and Auto-Regressive Transformers), T5 (Textto-Text Transfer Transformer) [3], and PEGASUS (Pretraining with Extracted Gap-sentences for Abstractive Summarization) [4].
Information overloading is a prominent and weakly addressed research issue while generating text summaries.Generally, while summarizing the text content, the deep learning models have to refer to enormous corpora of text documents and various text sources.The summarizing model can miss out the important information and may generate unrelated text in the summary.Even the summary could be too short, so essential information in the text summary could be missed.This problem increases when a huge volume of text needs to be processed and summarized.Such issues are critical in business models and can negatively affect the growth of successful business.The described problem is termed as information overloading, and deep learning models are prone to this problem.The need for precise information extraction and condensation techniques has increased to unprecedented levels in the current era of information overload, where an overwhelming volume of text is generated daily across the digital landscape.This problem is addressed by text summarization, a crucial component of natural language processing (NLP), which automatically generates clear and coherent summaries from lengthy texts [1].Pre-trained language models are a notable advancement in natural language processing (NLP) that could lead to significant improvements in text summarization automation.With the help of cutting-edge pre-trained models, such as BART (Bidirectional and Auto-Regressive Transformers) [2], T5 (Text-to-Text Transfer Transformer) [3], and PEGASUS (Pretraining with Extracted Gap-sentences for Abstractive Summarization) [4], this research project hopes to produce summaries that are human-like and make it easier to extract important information.
This research investigates the techniques, approaches, and resources used to apply these models for automated text summarization in order to tackle the problem of information overload.Examining the fundamental framework and mathematical foundations of each model provides insight into how they understand and generate textual content.Additionally, the study shows how these models can be used in practice by producing summaries for a range of input articles.Evaluation standards like METEOR [5] and ROUGE [6] are used to rate the quality of the summaries that are produced.The main objective of this research is to automate the summarization process by quickly extracting important insights from the growing body of textual data by utilising the power of pre-trained models.
The use of state-of-the-art pre-trained language models, such as PEGASUS (Pre-training with Extracted Gapsentences for Abstractive Summarization), T5 (Text-to-Text Transfer Transformer), and BART (Bidirectional and Auto-Regressive Transformers), is the main focus of this study.T5 employs a text-to-text framework, BART is proficient in bidirectional and auto-regressive transformations, and PEGASUS creatively employs pre-training with extracted gap-sentences for abstractive summarization.
Each model-the flexible text-to-text framework of T5, the bidirectional approach of BART, and the specific pre-training with extracted gaps of PEGASUS-brings a special strength to the table.Their comparison sheds light on subtle variations in the way these models interpret and handle text, which advances our knowledge of how these models work to automate text summarization.This paper's later sections go into greater detail about each model, describing how it was used and the results it yielded.This exhaustive investigation establishes the foundation for an in-depth comprehension of how these models work in automating the process of text summarization.

RELATED WORK
An essential task in natural language processing known as automated text summarization [7] has attracted a lot of interest lately.There have been several methods investigated, which can be generically divided into extractive and abstractive summarization.While abstractive summarization creates new phrases that capture the essential information, extractive summarization [1,8] entails choosing and extracting specific sentences from the source text.
The field of text summarization has come a long way from simple rule-based techniques to the sophisticated state-of-theart approaches used today.In its early iterations, rule-based summarization extracted important sentences or phrases from a given text by applying predefined grammatical and linguistic rules.By laying down the foundational ideas of information extraction and condensation, this early period prepared the way for later advances.The development of text summarization techniques reveals an amazing trajectory that reflects the constant quest for better ways to condense information.From the earliest rule-based techniques to the complex, cutting-edge techniques used in modern natural language processing (NLP), this journey can be followed.
Text summarization was first based on crude rule-based techniques, which involved sorting through the text and extracting important sentences or phrases using predetermined grammatical and linguistic rules.By developing the fundamental ideas of information extraction and condensation, this era set the stage for later developments.Rule-based summarization followed preset linguistic structures in an attempt to extract the main ideas of a text.
With the advent of statistical methods, the field experienced a paradigm shift away from strict rule-based approaches.At this stage, algorithms started evaluating a sentence's importance by looking for statistical patterns in the text.During this time, extractive summarization-which involves identifying and presenting existing sentences to compose a concise summary-became increasingly popular.By using statistical analysis to guide sentence selection, this method sought to preserve the most important and instructive passages from the original text.
The introduction of machine learning, and especially the rise of deep learning methods, brought about a paradigm change.The era of abstractive summarization began with this revolution, in which models were able to produce summaries that resembled those of humans by rephrasing and paraphrasing the original text.Abstractive methods, as opposed to extractive summarization, concentrated on comprehending the context and meaning of the content, enabling a more nuanced representation in the summary.This was a big step forward, as models were now able to understand the underlying semantics and produce more concise but still coherent versions of the original content, going beyond simple sentence selection.
The development of text summarization techniques over time can be summarized as a trip from rule-based simplicity to statistical analysis and finally to the revolutionary power of deep learning.The transition from extractive to abstractive summarization is indicative of an ongoing effort to capture the subtle meaning of textual content, which has led to the development of advanced methodologies used in modern NLP.Table 1 shows the summarized literature reviewof related works.
Abstractive summarization: Abstractive summarization is the process of generating more human-like, concise summaries using NLP [9].Text summarization tasks have been very famous due to the fastmoving life in the 21st century.Let's look at the most used approaches: (1) Transformer-Based Techniques: Transformer-based techniques are the techniques we have focused on in this paper.These techniques have gained popularity because of their ability to produce human-like, accurate, and concise summaries.
(2) Pointer Generator Networks: These networks are known to combine abstractive and extractive techniques to address the difficulties in content selection and fluency.These, however, have a few drawbacks, which include limited semantics understanding cannot handle rare words or new words well.
(3) Graph-Based Approaches: Graph based Neural Networks (GNNs) are used when the document structure and relationships between the sentences are very important.
(4) Reinforcement Learning approaches: These models are re-trained to generate more human-like summaries until we achieve a desired output.
Extractive summarization: Using key phrases or sentences that are extracted straight from the original content, extractive summarizing techniques create summaries.Below is a quick summary of the extractive summarization techniques that were discussed: (1) Term Frequency Inverse Document Frequency: The statistical measure known as TF-IDF is used to assess a word's significance in a document in relation to a corpus, or collection of documents.Sentences with higher TF-IDF scores-a measure of their importance in the document-are chosen for the summary in extractive summarization.Easy to understand and effectively computed.It draws attention to terms that are unique to a manuscript.However, it has limited comprehension of semantics; might miss word linkages.
(2) Cluster-Based Method: Sentences with comparable content are grouped into clusters using cluster-based algorithms.Next, sentences that best represent each cluster are chosen to be included in the summary.K-means clustering, and hierarchical clustering are two popular methods.Gather similar sentences to capture thematic consistency.May have trouble with a variety of topics within a document, sensitive to parameter settings and clustering method selection.
(3) Graph Theoretic Approach: Sentences are modelled as nodes in a graph-based approach, while the interactions between sentences are modelled as edges.To find crucial sentences for extraction, centrality metrics like degree centrality and PageRank are employed.Uses graph topology to capture the links and significance of texts.May have trouble with long-range dependencies; sensitive to how relationships are represented.
(4) Latent Semantic Analysis: Singular value decomposition (SDVD) is a technique used in LSA analysis to examine the links between terms and concepts in a document.Sentences that make the most contributions to the underlying latent semantic structure are identified and chosen in extractive summarization.Able to manage polysemy and synonymy; captures semantic links.May have trouble comprehending meanings particular to a certain situation; requires a sizable corpus for training.The article suggests an NLP text summary technique that sees style transfer as a semisupervised learning job and treats text summarization as such.The technique, which is based on a T5 model, is tested on the CSL and LCSTS datasets in China and shows promise for producing summaries of extensive texts.Although current supervised models still outperform the suggested method, it is more practical and efficient for real-world applications since it makes use of unlabeled data and only needs a small number of labelled samples.Reducing the dependency on labelled data and improving the model architecture can be the main goals of future study.[4] The article introduces PEGASUS, a novel abstractive text summarization method using Gap Sentences Generation (GSG) for self-supervised pretraining.GSG involves removing key sentences from input documents and generating summaries from the remaining content.PEGASUS achieved cutting-edge performance across 12 samples, it outperformed prior state-of-the-art results in low-resource summarization.Human evaluations confirmed its summaries matched human-level quality across various datasets.PEGASUS was fine-tuned and compared to pretraining models like BART, T5, and UniLM for optimal performance.[5] The evaluation tool ROUGE (Recall-Oriented Understudy for Gisting Evaluation) presents four distinct methods to automatically assess a summary's quality by contrasting it with other perfect summaries written by humans.These metrics track the number of overlapping units, such as n-grams, word sequences, and word pairs, between the computer-generated summary and the ideal summaries.The metrics have been utilised in extensive summary evaluations and have demonstrated a strong correlation with subjective assessments.Automated evaluation methods like ROUGE, including ROUGE-L, ROUGE-W, and ROUGE-S, are popular due to their efficiency and excel in single document, short, and multi-document summarization across various datasets.Excluding stopwords often enhances performance in ROUGE, which closely aligns with human judgments, especially in single-document and concise summaries, though multi-document summarization correlations remain a challenge.
In order to assess a summary's quality by contrasting it with ideal summaries written by humans, the ROUGE assessment package presents four metrics (ROUGE-N, ROUGE-L, ROUGE-W, and ROUGE-S).These metrics track the number of overlapping units, such as n-grams, word sequences, and word pairs, between the computer-generated summary and the ideal summaries.The ROUGE measures have been examined on several datasets, and the results demonstrate good performance in single-document summarization tasks and very short summary tasks, although human judgement correlations in multi-document summarization tasks remain challenging.[7] A study compared Bart and T5 models for news article summarization using keywords.Bart slightly outperformed T5 among approximately 1000 articles, especially for mid-sized news pieces.The study also highlighted challenges in news summarization, including manual selection and the presence of ads, which were mitigated using the newspaper3k Python tool.Automated summarization offers significant advantages in delivering article highlights.
Researchers experimented with Bart and T5 models using the BBC News Summary dataset through the Huggingface transformers API.Bart consistently outperformed T5 in summarization, with higher Rouge scores and an F1 score averaging 33% compared to T5's 26%.Bart's superior recall and precision make it more efficient for medium-sized news article summarization, saving time and resources.
When it came to creating summaries of news stories, the Bart model fared better than the T5 model, however the difference was not large.Challenges in news summary include the need for manual selection of key elements and the existence of irrelevant content and advertisements.The Bart model was more effective at summarising news because it regularly outperformed the T5 model in terms of summary generation and had higher Rouge scores.

PRE-TRAINED LANGUAGE MODELS
This section introduces three top models-BART [2], T5 [3], and PEGASUS [4] in the quickly developing field of natural language processing and automated text summarization [10].These models are state-of-the-art methods for abstractive summarization, each having an own design and goals.BART excels at producing coherent summaries thanks to its dual nature and encoder-decoder structure.T5's unified text-to-text framework, which offers amazing flexibility, revolutionises NLP [11] tasks.With its unique gap-sentence generating purpose, PEGASUS distinguishes out and excels at producing succinct and insightful summaries.This section offers a succinct summary of these models, emphasising the substantial contributions they have made to automated summarization.

BART (Bidirectional and Auto-Regressive Transformers)
BART is a transformer-based architecture [2] that has shown success in a variety of natural language processing tasks, including text summarization.It is represented by the model name "facebook/bart-large-cnn." BART demonstrates a dual nature because it incorporates both auto-regressive and bidirectional traits.The approach makes use of an encoderdecoder architecture, which is essential for tasks involving sequences of events, including summarization [12,13].The denoising autoencoder objective of BART, which may be mathematically represented as:  While the second part of the objective promotes the creation of coherent summaries, the first term pushes the model to reconstruct the input text from corrupted copies.BART is equipped to perform well on tasks requiring abstractive summarization [14] due to the denoising target and the bidirectional encoding, as shown in Figure 1.
In order to efficiently capture contextual information, BART employs a bidirectional approach to understand and represent input text from both directions.Notably, BART presents an auto-regressive training strategy that enables the creation of coherent and contextually rich abstractive summaries by teaching the model to predict the original sequence from corrupted versions.BART is capable of comprehending and combining textual content for tasks such as text summarization because of its bidirectional architecture and auto-regressive feature.

Text-to-text transfer transformer (T5)
The T5 [15] framework includes the "t5-small" model, a flexible pre-trained model for comprehending and producing natural language.By using a unified text-to-text framework [15], where both input and output are represented as text, T5 completely transforms how NLP tasks are structured.Specifically using an encoder-decoder configuration for sequence-to-sequence activities, the transformer architecture forms the foundation of T5's fundamental architecture.Despite the fact that T5's architecture is not defined by a particular mathematical equation, the text-to-text strategy guarantees consistency in performing a range of NLP [7] tasks, including summarization [16,17].T5 streamlines the model's design and makes it more adaptable to different tasks by articulating them all in the same text-to-text way.This requires only minor taskspecific alterations.The detailed architecture is shown in Figure 2. PEGASUS (Pre-training with Extracted Gap-sentences for Abstractive Summarization) is a pretrained language model intended for abstractive summarization.According to its architecture, a novel gap-sentence generation technique is used to pre-train the model on a sizable corpus of documents.By inserting sentences that have been arbitrarily removed from documents, PEGASUS helps the model comprehend context and produce coherent summaries.PEGASUS is effective for automated text summarization tasks because of this method's ability to capture long-range dependencies and generate abstractive summaries that preserve the important information from the input text.
By successfully modelling the creation of short and cogent summaries, PEGASUS excels at abstractive summarization [12].The method used in this model closely matches the fundamental characteristics of summarizing jobs.
These models, each with a distinct architecture and learning goals, make a substantial contribution to our research's automated summarization process, demonstrating the breadth and diversity of natural language processing methodologies.

METHODOLOGY
The suggested methodology aims to summarize business articles using the text summarization models BART, T5 and Pegasus.This study further aims to evaluate these models based on ROUGE [5] score and METEOR [6] to determine assess the limitations and strengths of each model.This procedure entails the following carefully curated steps to obtain a useful analysis as shown in Figure 3.

Dataset acquisition and preprocessing
This step serves as a crucial step in obtaining an accurate model and analysis.The dataset for this study is acquired from a reputable source i.e., BBC News Articles.This study is focused on the Business genre of the articles.The dataset consisted of news articles and summaries.This study aims to generate new summaries using the BART, T5 and Pegasus model and compare the generated summaries to the already existing summaries.Further, to perform this task efficiently we have used the methodical cleaning process.

Model initialization
This is the core of the methodology as it involves ensuring that the required packages are already installed and involves using these packages to import the models this study aims to compare.The required packages are transformers, rouge and sentencepiece.

BART model
This involves the employment of BART model and then initializing the tokenization.We initialized the model and tokenizer from the hugging face library using a pretrained BART model namely facebook/bart-large-cnn.Further we used this model to generate news article summaries.

T5 model
This involves importing a pretrained model from the hugging face library.The T5 model used in this study is T5small.This also loads tokenizers from the library to preprocess the input text into suitable tokens.Further this model is used to generate the appropriate news article summaries.

Pegasus model
Using the hugging face library, we import a pretrained Pegasus model namely google/pegasus-cnn_dailymail.This step involves the initialization of the model and tokens to be used.

Text summarization task
The input dataset is passed through the three models namely BART, T5 and Pegasus Model.The news articles and summaries are provided as an input.These models generate summaries for all the news articles inputted to the models.Each of these generated summaries are stored using indexing thereby making them easy to access and identify.
These model generated summaries are evaluated against the already existing summaries to evaluate the models performance.

Evaluation metrics
The following evaluation metrics [18] were used to assess and perform a comparison between the models.

Rouge-1
In the field of natural language processing, ROUGE-1, a variation of the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) metric [5], is an extensively used evaluation measure, particularly in the evaluation of automatic text summarization systems.Its main objective is to measure the degree of similarity and overlap between the words (or unigrams) in the automatically generated summary and those in the reference summary that was manually crafted.ROUGE-1 is specifically concerned with assessing the sufficiency of individual words in the generated summary in relation to the reference summary.
where, Precision: The ratio of the number of matched unigrams to the total number of unigrams in the generated text.
Recall: The ratio of the number of matched unigrams to the total number of unigrams in the reference text.
Penalty: The penalty term to account for unaligned words.α: A tuneable parameter to control the importance of precision versus recall.

Rouge-P
In the realm of natural language processing, ROUGE-P, a ROUGE [5] metric version, is a widely used evaluation metric, particularly for evaluating the calibre of automatically generated summaries.ROUGE-P measures the overlap of ngrams (contiguous sequences of n items, often words) between the machine-generated summary and the reference (humanwritten) summary with a focus on the precision component of the evaluation.
ROUGE-P analyses an automated summarising system's capacity to generate n-grams that are found in the reference summary.The generated summary more closely resembles the reference summary in terms of shared n-grams the higher the ROUGE-P score.

Exploratory data analysis
Figure 4 shows the creation of summaries for all articles, an exploratory data analysis (EDA) phase is carried out to learn more about the outcomes of the summary [7] and the distribution of the different parameters.The retrieved summaries serve as the starting point for collecting insights, patterns, and trends from the summarised information, along with any embedded metadata [19,20].
Figure 4 represents the word cloud of the articles.A word cloud representation helps us understand the frequently repeated words in our dataset.
Figure 5 histogram gives us the distribution of summary lengths as we can infer from the image that most of the summaries range between 100 to 175 words.
The boxplot in Figure 6 shows us the distribution of article length and summary lengths.As inferred from the image and calculated the average article length is 329 words and average summary length is 140 words.

RESULTS
In this paper, we carried out a thorough analysis of three cutting-edge models of abstractive summarization: BART, T5, and PEGASUS.The study, which sought to gauge these models' capacity for summarization, was based on a number of significant criteria, including METEOR, ROUGE-P, and ROUGE-1.

Meteor scores
Meteor scores often range between 0 and 1, although it isn't universally defined what is the range of a good meteor score, but it considered that a meteor score greater than 0.25 is considered acceptable in this study.By comparing the terminology and phrasing of generated summaries to humanwritten reference summaries, METEOR evaluates the generated summaries' quality.With the greatest METEOR score, T5 emerged as the winner, closely followed by BART.PEGASUS has a marginally lower score, indicating that its vocabulary and phrase alignment could use some improvement as shown in Table 2.

Rouge-P scores
The generated summaries' ability to accurately extract ngrams from the reference summaries is measured by ROUGE-P (Precision).BART demonstrated the highest level of precision, demonstrating its capacity to accurately copy information from the reference summary.T5 and PEGASUS came in second and third, respectively, demonstrating a marginally lower precision, as shown in Table 3.

Rouge-1 scores
Rouge scores in general range from 0 to 1, (rouge-1 and rouge-p included).A rouge score that ranges between 0 to 0.3 is considered as poor, 0.3 to 1 is considered acceptable in this study.Between the generated summaries and the reference summaries, ROUGE-1 measures the unigram overlap.T5 had the greatest unigram overlap in this area BART took second place, while PEGASUS, indicating the ability to align unigrams, as shown in Table 4.

Compare and contrast
BART, T5, and PEGASUS have all shown considerable gains in their performance after fine-tuning for summarizing news business items.Figure 7 shows a sample summary output of the models.With the highest METEOR score, T5 distinguishes out as having a thorough comprehension of language and semantics.BART performs well in ROUGE-1, showing significant word overlap with reference summary, while T5 performs well in ROUGE-P, indicating superior precision.PEGASUS is a good option since it retains good precision despite having somewhat lower ratings in other areas.Depending on the precise requirements of the summarizing assignment, the decision between these models should be based on priorities-comprehensive understanding, precision, or a balance between the two.
We would recommend the usage of T5 model for news summarization because of its high performance in ROUGE-1 and also shows us the best results in METEOR scores as well.Though it lagged in ROUGE-P scores, the scores were good enough.As we know, METEOR scores and ROUGE-1 scores higher than 0.30 are considered good.

CONCLUSION
Abstractive and extractive summarization techniques have several real-life applications, such as News Summarization, Meeting Summarization, Legal Case Summarization, etc.In this study, the abstractive summarization performance of three top models-BART, T5, and PEGASUS-was thoroughly assessed.Essential metrics METEOR, ROUGE-P (Precision), and ROUGE-1 served as the foundation for the evaluation.We evaluated them in order to determine their relative performance and to help practitioners choose the best model for their summarising requirements.
In line with human-written summaries, T5 has the highest METEOR score, which was revealed by our research.This indicates that it has the best vocabulary and phrasing alignment.BART, on the other hand, demonstrated the highest level of precision (ROUGE-P), excelling in the exact replication of reference n-grams.A significant alignment with the reference summaries is also indicated by the fact that BART consistently produced the largest unigram overlap (ROUGE-1).
In conclusion, the particular needs of the work at hand determine the best summarization model to use.BART is a strong contender if the goal is to faithfully recreate reference content.T5 is a strong option, though, if lexical and phrase fidelity are the primary concerns.PEGASUS offers a balanced performance and may represent a workable compromise, although somewhat falling short in some criteria.

FUTURE WORK
In the future, investigating ensemble strategies that take advantage of the capabilities of several models may lead to improvements in abstractive summarization.Additionally, examining these models across a wider range of datasets and domains may offer a more thorough knowledge of their strengths and weaknesses.
Future studies could examine real-world situations and explore how these models might be used to improve the practical implementation of these models.Valuable insights could be gained by identifying the research that is required to enable the integration of these models into practical applications.By setting the foundation for wise decisionmaking in abstractive summarizing, this work supports ongoing efforts to improve the caliber and effectiveness of automated summarization systems.Furthermore, evaluating the moral ramifications and any biases related to the deployment of abstractive summarization models in practical settings is an important area for further research.To guarantee responsible deployment and avoid unexpected repercussions, research addressing fairness, accountability, and transparency in automated summarization systems is crucial.
represents the number of training examples.  is the input text.  is the corresponding target text (summary).θ denotes the model parameters.λ is a regularization hyperparameter.

Figure 2 .
Figure 2. T5 model architecture 3.3 PEGASUS (Pre-training for abstractive summarization using extracted gap-sentences) PEGASUS [4] represented by the model 'google/pegasus-cnn_dailymail,' is specifically created for tasks requiring abstractive text summary.During pre-training, it uses a gapsentence generation aim, which enables it to efficiently capture the document-level structure.The unique PEGASUS objective, which involves randomly hiding off sentences in a document and training the model to produce those sentences, is where the essential innovation lies.This goal can be expressed numerically as:

Figure 3 .
Figure 3. Steps to obtain a useful analysis

Table 1 .
Displays the literature review of related works

Table 2 .
Comparison of average METEOR scores

Table 3 .
Comparison of average ROUGE-P scores

Table 4 .
Comparison of average ROUGE-I scores