English. A summary that aims at preserving the emotions of the original text can be interesting in certain application scenarios, such as in the generation of metareviews, both in academic and commercial domains. TextRank is a well-studied algorithm for automatic extractive summarization. This work introduces SentiTextRank, an emotional variant of TextRank, to enhance the extractive technique for both single-document and multi-document summarization. SentiTextRank incorporates emotions into the summarization process by classifying sentences into the eight emotional categories used in SenticNet. The preliminary evaluation of SentiTextRank yields encouraging results. In particular, our method generates informative summaries composed of sentences that preserve the emotional content of the original document. Italian. Un riassunto che miri a preservare le emozioni del testo originale può essere interessante in alcuni scenari applicativi, come ad esempio nella generazione di meta-recensioni sian nel dominio accademico che in quello commerciale. TextRank è un algoritmo per il riassunto automatico estrattivo molto studiato. Questo lavoro introduce SentiTextRank, una variante emozionale di TextRank, per potenziare la tecnica estrattiva sia per il riassunto di singoli documenti che per il riassunto di documenti multipli: SentiTextRank integra le emozioni nel processo di sintesi, classificando le frasi nelle otto categorie emotive utilizzate in SenticNet. La valutazione preliminare di SentiTextRank produce dei risultati incoraggianti. In particolare, il nostro metodo produce dei riassunti informativi formati da frasi che rispettano il contenuto emozionale del documento originale.
important to consider emotions to create summaries that truly reflect the essence of the source texts.
In this paper, by incorporating sentiment scoring be- SentiTextRank: Input=Source, Output= tween sentences, we generate summaries that capture 1. Set compression ratio parameter between the the emotional tone and impact of the original text. We source(s) and the final summary . believe that exploring this aspect further would lead to 2. Classify sentences of the source(s) into diferent more comprehensive and efective text summarization SenticNet emotion categories with ∈ methods. {joy, admiration, surprise, fear, disgust, anger, sad
This paper has two main goals. First, we define a new ness, interest}. algorithm called SentiTextRank, which is an emotional 3. Generate a summary for each emotion variant of TextRank [5]. Second, we provide an initial category by using TextRank. evaluation of SentiTextRank by considering two auto- 4. Build by picking a number of sentences matic metrics based on content distance. proportional to from each maintaining
Note that modern LLMs showed some abilities in sum- the original sentence order of the source documarizing texts by using a specific style, with some limi- ment. tations in producing a summary that is truly extractive.
Moreover, LLMs showed also a big impact from the point of view of the required computational resources. We be- 3. Experimental Result and lieve that the work presented in this paper, that requires Discussion just few hours to conduct all the experiments, can be seen as a cheap (in many senses) alternative to the use of modern expensive (in many senses) LLMs1.
The paper is structured as follows. In Section 2, we define the new SentiTextRank algorithm, in Section 3 we report the result of a first experimental evaluation of the SentiTextRank algorithm and the Section 4 ends the paper pointing out to work in progress.
of single-document summarization using two datasets, the CNN/Daily Mail dataset (CNN) and the DUC2001 single document dataset (D01), as well as the results of multi-document summarization using two datasets, the DUC2001 multi-document dataset (MD01) and the DUC2004 multi-document dataset (MD04).
The DUC 2001 single document and DUC 2001 Multidocument datasets were collected from the website2 and
consist of news datasets. For our experiments with single TextRank accounting for documents, we utilized a sample data set of 54 documents. emotions The DUC 2004 multi-document summarization dataset3 includes 50 items with multiple files and four reference TextRank is a popular algorithm for extractive summa- files per item, from which we utilized the first reference rization which constructs a graph of sentences or words for each item. Additionally, we used the CNN/Daily mail from a text and assigns scores to each node based on dataset4, where we considered the “highlights” column their importance in the graph structure. Finally, it ranks as the reference summary. Our experiments were conthe nodes and selects the top-ranked sentences or words ducted on the first 100 rows of text in the CNN/Daily as the summary [5]. The TextRank algorithm is based mail dataset. Since the datasets provide just abstractive on the PageRank algorithm, where the sentences of the gold summaries, in order to provide a fair comparison we documents play the role of web pages, and a similarity have converted the abstractive summaries into extractive score plays the role of hyperlink connectivity. Our ap- summaries. This procedure has been proposed in [7, 8]. proach enhances traditional TextRank by incorporating An extractive reference summary should yield the highemotions. In particular, we categorize sentences of the est Rouge score when compared to the gold abstractive original source(s) on the basis of emotions using Sentic- summary. As finding the globally optimal subset of senNet [6]. On the basis of this classification, we obtain a tences that maximize the Rouge score is computationally number of distinct emotion sets of sentences. The main intractable, we adopt a greedy approach: we iteratively idea is to build one single final summary by merging in add one sentence at a time to the summary, ensuring that a selective way the results of TextRank on each one of the Rouge score of the current set of selected sentences these emotions sets. is maximized in relation to the entire gold summary. We
So, the proposed SentiTextRank algorithm generates repeat this process until there are no more candidate senextractive summaries with emphasis on emotion tences that could enhance the Rouge score when added categories through the following steps:
tive Summary
Ever noticed how plane seats appear to be getting smaller and smaller? With increasing numbers of people taking to the skies, some experts are questioning if having such packed out planes is putting passengers at risk.
Experts question if packed out planes are putting passengers at risk. U.S consumer advisory group says minimum space must be stipulated.
Ever noticed how plane seats appear to be getting smaller and smaller?. This week, a U.S consumer advisory group set up by the Department of Transportation said at a public hearing that while the government is happy to set standards for animals flying on planes, it doesn’t stipulate a minimum amount of space for humans.
Ever noticed how plane seats appear to be getting smaller and smaller? With increasing numbers of people taking to the skies, some experts are questioning if having such packed out planes is putting passengers at risk.
They say that the shrinking space on aeroplanes is not only uncomfortable - it’s putting our health and safety in danger. This week, a U.S consumer advisory group set up by the Department of Transportation said at a public hearing that while the government is happy to set standards for animals flying on planes, it doesn’t stipulate a minimum amount of space for humans.
They say that the shrinking space on aeroplanes is not only uncomfortable - it’s putting our health and safety in danger. ’It is time that the DOT and FAA take a stand for humane treatment of passengers. to the current summary set. The subset of sentences that summarization methods, namely the Baseline (Lead), Texwe have at this point is then considered the extractive ref- tRank (TR), and our proposed method, SentiTextRank erence summary for the evaluation. In Table 1 we report (STR) evaluated on single-document datasets DUC-2001 an example of summaries generated with the diferent (D01) and CNN, and the multi-document datasets DUCmethods. 2001 (MD01) and DUC-2004 (MD04). As a baseline, we
Table 1 reports an excerpt from the CNN dataset (Orig- selected the leading sentences from the original docuinal Text) and the corresponding reference summary ments based on the compression ratio. We evaluated (Gold Abstractive Summary). Moreover, Table 1 con- the summaries using two measures: Rouge-L F1 (RL-F1) tains the corresponding generated reference extractive and BERT F1. ROUGE (Recall Oriented Understudy for summary (Reference Extractive Summary), the prefix Gisting Evaluation) is frequently used to assess how well baseline (Lead), the text generated with the TextRank summarization techniques perform. Rouge-L computes baseline (TR) and, finally, the SentiTextRank generated ROUGE for the longest sequence of n-grams [9]. BERT summary (STR). F1 score is a metric commonly used in text classification tasks. It measures the token-level similarity between the Dataset Algorithm RL-F1 BERT-F1 generated summary and the reference summary, consid
Lead 0.600 0.729 ering both precision and recall [10].
D01 TR 0.382 0.649 The results consistently indicate that the Lead method STR 0.366 0.605 outperforms the other methods across the datasets, showLead 0.711 0.794 casing its superiority in generating high-quality sumCNN TR 0.345 0.642 maries. Specifically, Lead achieves the highest scores in STR 0.372 0.608 Rouge-L F1 and BERT-F1 for D01, CNN, MD01, and MD04.
Lead 0.802 0.851 The TR and STR methods exhibit moderate performance MD01 STTRR 00..016631 00..550605 in specific evaluation metrics.
Lead 0.511 0.683 Note that the better performance of the Lead method MD04 TR 0.123 0.575 can be attributed to the fact that all the experiments were STR 0.227 0.542 conducted using news datasets; indeed this result is consistent with the results reported in the literature, where Table 2 a baseline composed of the leading sentences frequently The results of summarization experiments. Lead = Lead Base- outperforms extractive and abstractive models on news line, TR = TextRank, STR = SentiTextRank. datasets [11]. However, we think that the comparison between original TR and STR shows encouraging results.
Table 2 presents the experimental results of diferent Indeed, the fact that using emotions does not degrade the performance with regard to TextRank shows that we can produce a summary that represents the content as well as the emotions of the source documents. In order to experimentally prove this intuition, we need to formalize an emotional distance between summaries and source documents. We plan to develop this point in the future by both (1) using LLM and (2) considering human evaluation.
Further research is necessary to evaluate the performance of our proposed STR method on another domain dataset to provide a comprehensive understanding of its efectiveness.
which integrates emotions into the extractive summarization process to create more informative and emotionally rich summaries. The experimental results are encouraging with respect the efectiveness of SentiTextRank in capturing factual information.
The ongoing work on SentiTextRank is following different directions.
First, we want to design a new version of the algorithm that will not be based on the classification of a sentence in one single prevalent emotion. The idea that we want to develop is to define one single measure that combines both content and emotion similarities. By using this combined measure, we can apply the original TextRank algorithm on the entire set of sentences from the source(s) and obtain one single ranking structure accounting for both content and emotion.
Second, we want to conduct more extensive experiments also on datasets from diferent domains. In particular, we are considering medical applications since the afective component of medical information can represent a relevant biopsychosocial feature [12].
Third, we are aware that automatic metrics not always measure a real quality of the summarized text with respect to human judgment [13, 14]. So, we plan in future to conduct human-based evaluation too.