Génération de résumés par abstraction

Cette thèse présente le résultat de plusieurs années de recherche dans le domaine de la génération automatique de résumés. Trois contributions majeures, présentées sous la forme d'articles publiés ou soumis pour publication, en forment le coeur. Elles retracent un cheminement qui part des méthodes par extraction en résumé jusqu'aux méthodes par abstraction. L'expérience HexTac, sujet du premier article, a d'abord été menée pour évaluer le niveau de performance des êtres humains dans la rédaction de résumés par extraction de phrases. Les résultats montrent un écart important entre la performance humaine sous la contrainte d'extraire des phrases du texte source par rapport à la rédaction de résumés sans contrainte. Cette limite à la rédaction de résumés par extraction de phrases, observée empiriquement, démontre l'intérêt de développer d'autres approches automatiques pour le résumé. Nous avons ensuite développé un premier système selon l'approche Fully Abstractive Summarization, qui se situe dans la catégorie des approches semi-extractives, comme la compression de phrases et la fusion de phrases. Le développement et l'évaluation du système, décrits dans le second article, ont permis de constater le grand défi de générer un résumé facile à lire sans faire de l'extraction de phrases. Dans cette approche, le niveau de compréhension du contenu du texte source demeure insuffisant pour guider le processus de sélection du contenu pour le résumé, comme dans les approches par extraction de phrases. Enfin, l'approche par abstraction basée sur des connaissances nommée K-BABS est proposée dans un troisième article. Un repérage des éléments d'information pertinents est effectué, menant directement à la génération de phrases pour le résumé. Cette approche a été implémentée dans le système ABSUM, qui produit des résumés très courts mais riches en contenu. Ils ont été évalués selon les standards d'aujourd'hui et cette évaluation montre que des résumés hybrides formés à la fois de la sortie d'ABSUM et de phrases extraites ont un contenu informatif significativement plus élevé qu'un système provenant de l'état de l'art en extraction de phrases.

Model Selection in Summary Evaluation

A difficulty in the design of automated text summarization algorithms is in the objective evaluation. Viewing summarization as a tradeoff between length and information content, we introduce a technique based on a hierarchy of classifiers to rank, through model selection, different summarization methods. This summary evaluation technique allows for broader comparison of summarization methods than the traditional techniques of summary evaluation. We present an empirical study of two simple, albeit widely used, summarization methods that shows the different usages of this automated task-based evaluation system and confirms the results obtained with human-based evaluation methods over smaller corpora.

COMPENDIUM: una herramienta de generación de resúmenes modular

En este artículo presentamos COMPENDIUM, una herramienta de generación de resúmenes de textos modular. Esta herramienta se compone de un módulo central con cinco etapas bien diferenciadas: i) análisis lingüístico; ii) detección de redundancia; iii) identificación del tópico; iv) detección de relevancia; y v) generación del resumen, y una serie de módulos adicionales que permiten incrementar las funcionalidades de la herramienta permitiendo la generación de distintos tipos de resúmenes, como por ejemplo orientados a un tema concreto. Realizamos una evaluación exhaustiva en dos dominios distintos (noticias de prensa y documentos sobre lugares turísticos) y analizamos diferentes tipos de resúmenes generados con COMPENDIUM (mono-documento, multi-documento, genéricos y orientados a un tema). Además, comparamos nuestro sistema con otros sistemas de generación de resúmenes actuales. Los resultados que se obtienen demuestran que la herramienta COMPENDIUM es capaz de generar resúmenes competitivos para los distintos tipos de resúmenes propuestos.

Topic Segmentation: How Much Can We Do by Counting Words and Sequences of Words

In this paper, we present an innovative topic segmentation system based on a new informative similarity measure that takes into account word co-occurrence in order to avoid the accessibility to existing linguistic resources such as electronic dictionaries or lexico-semantic databases such as thesauri or ontology. Topic segmentation is the task of breaking documents into topically coherent multi-paragraph subparts. Topic segmentation has extensively been used in information retrieval and text summarization. In particular, our architecture proposes a language-independent topic segmentation system that solves three main problems evidenced by previous research: systems based uniquely on lexical repetition that show reliability problems, systems based on lexical cohesion using existing linguistic resources that are usually available only for dominating languages and as a consequence do not apply to less favored languages and finally systems that need previously existing harvesting training data. For that purpose, we only use statistics on words and sequences of words based on a set of texts. This solution provides a flexible solution that may narrow the gap between dominating languages and less favored languages thus allowing equivalent access to information.

prism: rinfrangere distribuzioni di probabilità in riassunti astrattivi mediante strategie di decoding

Negli ultimi quattro anni la summarization astrattiva è stata protagonista di una evoluzione senza precedenti dettata da nuovi language model neurali, architetture transformer-based, elevati spazi dimensionali, ampi dataset e innovativi task di pre-training. In questo contesto, le strategie di decoding convertono le distribuzioni di probabilità predette da un modello in un testo artificiale, il quale viene composto in modo auto regressivo. Nonostante il loro cruciale impatto sulla qualità dei riassunti inferiti, il ruolo delle strategie di decoding è frequentemente trascurato e sottovalutato. Di fronte all'elevato numero di tecniche e iperparametri, i ricercatori necessitano di operare scelte consapevoli per ottenere risultati più affini agli obiettivi di generazione. Questa tesi propone il primo studio altamente comprensivo sull'efficacia ed efficienza delle strategie di decoding in task di short, long e multi-document abstractive summarization. Diversamente dalle pubblicazioni disponibili in letteratura, la valutazione quantitativa comprende 5 metriche automatiche, analisi temporali e carbon footprint. I risultati ottenuti dimostrano come non vi sia una strategia di decoding dominante, ma come ciascuna possieda delle caratteristiche adatte a task e dataset specifici. I contributi proposti hanno l'obiettivo di neutralizzare il gap di conoscenza attuale e stimolare lo sviluppo di nuove tecniche di decoding.

Text analytics of social media: Sentiment analysis, event detection and summarization

In the last decade, large numbers of social media services have emerged and been widely used in people's daily life as important information sharing and acquisition tools. With a substantial amount of user-contributed text data on social media, it becomes a necessity to develop methods and tools for text analysis for this emerging data, in order to better utilize it to deliver meaningful information to users. ^ Previous work on text analytics in last several decades is mainly focused on traditional types of text like emails, news and academic literatures, and several critical issues to text data on social media have not been well explored: 1) how to detect sentiment from text on social media; 2) how to make use of social media's real-time nature; 3) how to address information overload for flexible information needs. ^ In this dissertation, we focus on these three problems. First, to detect sentiment of text on social media, we propose a non-negative matrix tri-factorization (tri-NMF) based dual active supervision method to minimize human labeling efforts for the new type of data. Second, to make use of social media's real-time nature, we propose approaches to detect events from text streams on social media. Third, to address information overload for flexible information needs, we propose two summarization framework, dominating set based summarization framework and learning-to-rank based summarization framework. The dominating set based summarization framework can be applied for different types of summarization problems, while the learning-to-rank based summarization framework helps utilize the existing training data to guild the new summarization tasks. In addition, we integrate these techneques in an application study of event summarization for sports games as an example of how to better utilize social media data. ^

Text Analytics of Social Media: Sentiment Analysis, Event Detection and Summarization

In the last decade, large numbers of social media services have emerged and been widely used in people's daily life as important information sharing and acquisition tools. With a substantial amount of user-contributed text data on social media, it becomes a necessity to develop methods and tools for text analysis for this emerging data, in order to better utilize it to deliver meaningful information to users. Previous work on text analytics in last several decades is mainly focused on traditional types of text like emails, news and academic literatures, and several critical issues to text data on social media have not been well explored: 1) how to detect sentiment from text on social media; 2) how to make use of social media's real-time nature; 3) how to address information overload for flexible information needs. In this dissertation, we focus on these three problems. First, to detect sentiment of text on social media, we propose a non-negative matrix tri-factorization (tri-NMF) based dual active supervision method to minimize human labeling efforts for the new type of data. Second, to make use of social media's real-time nature, we propose approaches to detect events from text streams on social media. Third, to address information overload for flexible information needs, we propose two summarization framework, dominating set based summarization framework and learning-to-rank based summarization framework. The dominating set based summarization framework can be applied for different types of summarization problems, while the learning-to-rank based summarization framework helps utilize the existing training data to guild the new summarization tasks. In addition, we integrate these techneques in an application study of event summarization for sports games as an example of how to better utilize social media data.

Summarization in Math Class

This action research study of twenty students in my sixth grade mathematics classroom examines the implementation of summarization strategies. Students were taught how to summarize concepts and how to explain their thinking in different ways to the teacher and their peers. Through analysis of students’ summaries of concepts from lessons that I taught, tests scores, and student journals and interviews, I discovered that summarizing mathematical concepts offers students an engaging opportunity to better understand those concepts and render that understanding more visible to the teacher. This analysis suggests that non-traditional summarization, such as verbal and written strategies, and strategies involving movement and discussions, can be useful in mathematics classrooms to improve student understanding, engagement in learning tasks, and as a form of formative assessment.

Extractive summarization using complex networks and syntactic dependency

The realization that statistical physics methods can be applied to analyze written texts represented as complex networks has led to several developments in natural language processing, including automatic summarization and evaluation of machine translation. Most importantly, so far only a few metrics of complex networks have been used and therefore there is ample opportunity to enhance the statistics-based methods as new measures of network topology and dynamics are created. In this paper, we employ for the first time the metrics betweenness, vulnerability and diversity to analyze written texts in Brazilian Portuguese. Using strategies based on diversity metrics, a better performance in automatic summarization is achieved in comparison to previous work employing complex networks. With an optimized method the Rouge score (an automatic evaluation method used in summarization) was 0.5089, which is the best value ever achieved for an extractive summarizer with statistical methods based on complex networks for Brazilian Portuguese. Furthermore, the diversity metric can detect keywords with high precision, which is why we believe it is suitable to produce good summaries. It is also shown that incorporating linguistic knowledge through a syntactic parser does enhance the performance of the automatic summarizers, as expected, but the increase in the Rouge score is only minor. These results reinforce the suitability of complex network methods for improving automatic summarizers in particular, and treating text in general. (C) 2011 Elsevier B.V. All rights reserved.

Conditions of labor in American industries: a summarization of the results of recent investigations

by W. Jett Lauck and Edgar Sydenstricker.

Data catalog for the Marine Ecosystems Analysis Puget Sound Project : distribution and summarization of digital data /

The MESA Puget Sound Project is sponsored by the National Oceanic and Atmospheric Administration and the Environmental Protection Agency.

Mining the online social network data: Influence, summarization, and organization

Online Social Network (OSN) services provided by Internet companies bring people together to chat, share the information, and enjoy the information. Meanwhile, huge amounts of data are generated by those services (they can be regarded as the social media ) every day, every hour, even every minute, and every second. Currently, researchers are interested in analyzing the OSN data, extracting interesting patterns from it, and applying those patterns to real-world applications. However, due to the large-scale property of the OSN data, it is difficult to effectively analyze it. This dissertation focuses on applying data mining and information retrieval techniques to mine two key components in the social media data — users and user-generated contents. Specifically, it aims at addressing three problems related to the social media users and contents: (1) how does one organize the users and the contents? (2) how does one summarize the textual contents so that users do not have to go over every post to capture the general idea? (3) how does one identify the influential users in the social media to benefit other applications, e.g., Marketing Campaign? The contribution of this dissertation is briefly summarized as follows. (1) It provides a comprehensive and versatile data mining framework to analyze the users and user-generated contents from the social media. (2) It designs a hierarchical co-clustering algorithm to organize the users and contents. (3) It proposes multi-document summarization methods to extract core information from the social network contents. (4) It introduces three important dimensions of social influence, and a dynamic influence model for identifying influential users.

Mining the Online Social Network Data: Influence, Summarization, and Organization

Online Social Network (OSN) services provided by Internet companies bring people together to chat, share the information, and enjoy the information. Meanwhile, huge amounts of data are generated by those services (they can be regarded as the social media ) every day, every hour, even every minute, and every second. Currently, researchers are interested in analyzing the OSN data, extracting interesting patterns from it, and applying those patterns to real-world applications. However, due to the large-scale property of the OSN data, it is difficult to effectively analyze it. This dissertation focuses on applying data mining and information retrieval techniques to mine two key components in the social media data — users and user-generated contents. Specifically, it aims at addressing three problems related to the social media users and contents: (1) how does one organize the users and the contents? (2) how does one summarize the textual contents so that users do not have to go over every post to capture the general idea? (3) how does one identify the influential users in the social media to benefit other applications, e.g., Marketing Campaign? The contribution of this dissertation is briefly summarized as follows. (1) It provides a comprehensive and versatile data mining framework to analyze the users and user-generated contents from the social media. (2) It designs a hierarchical co-clustering algorithm to organize the users and contents. (3) It proposes multi-document summarization methods to extract core information from the social network contents. (4) It introduces three important dimensions of social influence, and a dynamic influence model for identifying influential users.

Verbalizzazione di eventi biomedici espressi nella letteratura scientifica: generazione controllata di linguaggio naturale da grafi di conoscenza mediante transformer text-to-text

Il periodo in cui viviamo rappresenta la cuspide di una forte e rapida evoluzione nella comprensione del linguaggio naturale, raggiuntasi prevalentemente grazie allo sviluppo di modelli neurali. Nell'ambito dell'information extraction, tali progressi hanno recentemente consentito di riconoscere efficacemente relazioni semantiche complesse tra entità menzionate nel testo, quali proteine, sintomi e farmaci. Tale task -- reso possibile dalla modellazione ad eventi -- è fondamentale in biomedicina, dove la crescita esponenziale del numero di pubblicazioni scientifiche accresce ulteriormente il bisogno di sistemi per l'estrazione automatica delle interazioni racchiuse nei documenti testuali. La combinazione di AI simbolica e sub-simbolica può consentire l'introduzione di conoscenza strutturata nota all'interno di language model, rendendo quest'ultimi più robusti, fattuali e interpretabili. In tale contesto, la verbalizzazione di grafi è uno dei task su cui si riversano maggiori aspettative. Nonostante l'importanza di tali contributi (dallo sviluppo di chatbot alla formulazione di nuove ipotesi di ricerca), ad oggi, risultano assenti contributi capaci di verbalizzare gli eventi biomedici espressi in letteratura, apprendendo il legame tra le interazioni espresse in forma a grafo e la loro controparte testuale. La tesi propone il primo dataset altamente comprensivo su coppie evento-testo, includendo diverse sotto-aree biomediche, quali malattie infettive, ricerca oncologica e biologia molecolare. Il dataset introdotto viene usato come base per l'addestramento di modelli generativi allo stato dell'arte sul task di verbalizzazione, adottando un approccio text-to-text e illustrando una tecnica formale per la codifica di grafi evento mediante testo aumentato. Infine, si dimostra la validità degli eventi per il miglioramento delle capacità di comprensione dei modelli neurali su altri task NLP, focalizzandosi su single-document summarization e multi-task learning.

EOPAS, the EthnoER online representation of interlinear text

One of the goals of the ARC funded Eresearch project called Sharing access and analytical tools for ethnographic digital media using high speed networks, or simply EthnoER is to take outputs of normal linguistic analytical processes and present them online in a system we have called the EthnoER online presentation and annotation system, or EOPAS.