abstract: For the past three decades, the design of an effective strategy for generating poetry that matches that of a human’s creative capabilities and complexities has been an elusive goal in artificial intelligence (AI) and natural language generation (NLG) research, and among linguistic creativity researchers in particular. This thesis presents a novel approach to fixed verse poetry generation using neural word embeddings. During the course of generation, a two layered poetry classifier is developed. The first layer uses a lexicon based method to classify poems into types based on form and structure, and the second layer uses a supervised classification method to classify poems into subtypes based on content with an accuracy of 92%. The system then uses a two-layer neural network to generate poetry based on word similarities and word movements in a 50-dimensional vector space. The verses generated by the system are evaluated using rhyme, rhythm, syllable counts and stress patterns. These computational features of language are considered for generating haikus, limericks and iambic pentameter verses. The generated poems are evaluated using a Turing test on both experts and non-experts. The user study finds that only 38% computer generated poems were correctly identified by nonexperts while 65% of the computer generated poems were correctly identified by experts. Although the system does not pass the Turing test, the results from the Turing test suggest an improvement of over 17% when compared to previous methods which use Turing tests to evaluate poetry generators. / Dissertation/Thesis / Masters Thesis Computer Science 2016

O que as outras pessoas pensam sempre foi uma parte importante do processo de tomada de decisão. Por exemplo, as pessoas costumam consultar seus amigos para obter um parecer sobre um livro ou um filme ou um restaurante. Hoje em dia, os usuários publicam suas opiniões em sites de revisão colaborativa, como IMDB para filmes, Yelp para restaurantes e TripAdiviser para hotéis. Ao longo do tempo, esses sites têm construído um enorme banco de dados que conecta usuários, artigos e opiniões expressas por uma classificação numérica e um comentário de texto livre que explicam por que eles gostam ou não gostam de um item. Mas essa vasta quantidade de dados pode prejudicar o usuário a obter uma opinião. Muitos trabalhos relacionados fornecem uma interpretações de revisões para os usuários. Eles oferecem vantagens diferentes para vários tipos de resumos. No entanto, todos eles têm a mesma limitação: eles não fornecem resumos personalizados nem contrastantes comentários escritos por diferentes segmentos de colaboradores. Compreeder e contrastar comentários escritos por diferentes segmentos de revisores ainda é um problema de pesquisa em aberto. Assim, nosso trabalho propõe uma nova arquitetura, chamado Tell Me Why. TMW é um projeto desenvolvido no Laboratório de Informática Grenoble em cooperação com a Universidade Federal do Rio Grande do Sul para fornecer aos usuários uma melhor compreensão dos comentários. Propomos uma combinação de análise de texto a partir de comentários com a mineração de dados estruturado resultante do cruzamento de dimensões do avaliador e item. Além disso, este trabalho realiza uma investigação sobre métodos de sumarização utilizados na revisão de produtos. A saída de nossa arquitetura consiste em declarações personalizadas de texto usando Geração de Linguagem Natural composto por atributos de itens e comentários resumidos que explicam a opinião das pessoas sobre um determinado assunto. Os resultados obtidos a partir de uma avaliação comparativa com a Revisão Mais Útil da Amazon revelam que é uma abordagem promissora e útil na opinião do usuário. / What other people think has been always an important part of the process of decision-making. For instance, people usually consult their friends to get an opinion about a book, or a movie or a restaurant. Nowadays, users publish their opinions on collaborative reviewing sites such as IMDB for movies, Yelp for restaurants and TripAdvisor for hotels. Over the time, these sites have built a massive database that connects users, items and opinions expressed by a numeric rating and a free text review that explain why they like or dislike a specific item. But this vast amount of data can hamper the user to get an opinion. Several related work provide a review interpretations to the users. They offer different advantages for various types of summaries. However, they all have the same limitation: they do not provide personalized summaries nor contrasting reviews written by different segments of reviewers. Understanding and contrast reviews written by different segments of reviewers is still an open research problem. Our work proposes a new architecture, called Tell Me Why, which is a project developed at Grenoble Informatics Laboratory in cooperation with Federal University of Rio Grande do Sul to provide users a better understanding of reviews. We propose a combination of text analysis from reviews with mining structured data resulting from crossing reviewer and item dimensions. Additionally, this work performs an investigation of summarization methods utilized in review domain. The output of our architecture consists of personalized statement using Natural Language Generation that explain people’s opinion about a particular item. The evaluation reveal that it is a promising approach and useful in user’s opinion.

O que as outras pessoas pensam sempre foi uma parte importante do processo de tomada de decisão. Por exemplo, as pessoas costumam consultar seus amigos para obter um parecer sobre um livro ou um filme ou um restaurante. Hoje em dia, os usuários publicam suas opiniões em sites de revisão colaborativa, como IMDB para filmes, Yelp para restaurantes e TripAdiviser para hotéis. Ao longo do tempo, esses sites têm construído um enorme banco de dados que conecta usuários, artigos e opiniões expressas por uma classificação numérica e um comentário de texto livre que explicam por que eles gostam ou não gostam de um item. Mas essa vasta quantidade de dados pode prejudicar o usuário a obter uma opinião. Muitos trabalhos relacionados fornecem uma interpretações de revisões para os usuários. Eles oferecem vantagens diferentes para vários tipos de resumos. No entanto, todos eles têm a mesma limitação: eles não fornecem resumos personalizados nem contrastantes comentários escritos por diferentes segmentos de colaboradores. Compreeder e contrastar comentários escritos por diferentes segmentos de revisores ainda é um problema de pesquisa em aberto. Assim, nosso trabalho propõe uma nova arquitetura, chamado Tell Me Why. TMW é um projeto desenvolvido no Laboratório de Informática Grenoble em cooperação com a Universidade Federal do Rio Grande do Sul para fornecer aos usuários uma melhor compreensão dos comentários. Propomos uma combinação de análise de texto a partir de comentários com a mineração de dados estruturado resultante do cruzamento de dimensões do avaliador e item. Além disso, este trabalho realiza uma investigação sobre métodos de sumarização utilizados na revisão de produtos. A saída de nossa arquitetura consiste em declarações personalizadas de texto usando Geração de Linguagem Natural composto por atributos de itens e comentários resumidos que explicam a opinião das pessoas sobre um determinado assunto. Os resultados obtidos a partir de uma avaliação comparativa com a Revisão Mais Útil da Amazon revelam que é uma abordagem promissora e útil na opinião do usuário. / What other people think has been always an important part of the process of decision-making. For instance, people usually consult their friends to get an opinion about a book, or a movie or a restaurant. Nowadays, users publish their opinions on collaborative reviewing sites such as IMDB for movies, Yelp for restaurants and TripAdvisor for hotels. Over the time, these sites have built a massive database that connects users, items and opinions expressed by a numeric rating and a free text review that explain why they like or dislike a specific item. But this vast amount of data can hamper the user to get an opinion. Several related work provide a review interpretations to the users. They offer different advantages for various types of summaries. However, they all have the same limitation: they do not provide personalized summaries nor contrasting reviews written by different segments of reviewers. Understanding and contrast reviews written by different segments of reviewers is still an open research problem. Our work proposes a new architecture, called Tell Me Why, which is a project developed at Grenoble Informatics Laboratory in cooperation with Federal University of Rio Grande do Sul to provide users a better understanding of reviews. We propose a combination of text analysis from reviews with mining structured data resulting from crossing reviewer and item dimensions. Additionally, this work performs an investigation of summarization methods utilized in review domain. The output of our architecture consists of personalized statement using Natural Language Generation that explain people’s opinion about a particular item. The evaluation reveal that it is a promising approach and useful in user’s opinion.

L'humanité a longtemps été passionnée par la création de machines intellectuelles qui peuvent librement intéragir avec nous dans notre langue. Tous les systèmes modernes qui communiquent directement avec l'utilisateur partagent une caractéristique commune: ils ont un système de dialogue à la base. Aujourd'hui pratiquement tous les composants d'un système de dialogue ont adopté des méthodes statistiques et les utilisent largement comme leurs modèles de base. Jusqu'à récemment la génération de langage naturel (GLN) utilisait pour la plupart des patrons/modèles codés manuellement, qui représentaient des phrases types mappées à des réalisations sémantiques particulières. C'était le cas jusqu'à ce que les approches statistiques aient envahi la communauté de recherche en systèmes de dialogue. Dans cette thèse, nous suivons cette ligne de recherche et présentons une nouvelle approche à la génération de la langue naturelle. Au cours de notre travail, nous nous concentrons sur deux aspects importants du développement des systèmes de génération: construire un générateur performant et diversifier sa production. Deux idées principales que nous défendons ici sont les suivantes: d'abord, la tâche de GLN peut être vue comme la traduction entre une langue naturelle et une représentation formelle de sens, et en second lieu, l'extension du corpus qui impliquait traditionnellement des paraphrases définies manuellement et des règles spécialisées peut être effectuée automatiquement en utilisant des méthodes automatiques d'extraction des synonymes et des paraphrases bien connues et largement utilisées. En ce qui concerne notre première idée, nous étudions la possibilité d'utiliser le cadre de la traduction automatique basé sur des modèles ngrams; nous explorons également le potentiel de l'apprentissage discriminant (notamment les champs aléatoires markoviens) appliqué à la GLN; nous construisons un système de génération qui permet l'inclusion et la combinaison des différents modèles et qui utilise un cadre de décodage efficace (automate à état fini). En ce qui concerne le second objectif, qui est l'extension du corpus, nous proposons d'élargir la taille du vocabulaire et le nombre de l'ensemble des structures syntaxiques disponibles via l'intégration des synonymes et des paraphrases. À notre connaissance, il n'y a pas eu de tentatives d'augmenter la taille du vocabulaire d'un système de GLN en incorporant les synonymes. À ce jour, la plupart d'études sur l'extension du corpus visent les paraphrases et recourent au crowdsourcing pour les obtenir, ce qui nécessite une validation supplémentaire effectuée par les développeurs du système. Nous montrons que l'extension du corpus au moyen d'extraction automatique de paraphrases et la validation automatique sont tout aussi efficaces, étant en même temps moins coûteux en termes de temps de développement et de ressources. Au cours d'expériences intermédiaires nos modèles ont montré une meilleure performance que celle obtenue par le modèle de référence basé sur les syntagmes et se sont révélés d'être plus robustes, pour le traitement des combinaisons inconnues de concepts, que le générateur à base des règles. L'évaluation humaine finale a prouvé que les modèles représent une alternative solide au générateur à base des règles / The humanity has long been passionate about creating intellectual machines that can freely communicate with us in our language. Most modern systems communicating directly with the user share one common feature: they have a dialog system (DS) at their base. As of today almost all DS components embraced statistical methods and widely use them as their core models. Until recently Natural Language Generation (NLG) component of a dialog system used primarily hand-coded generation templates, which represented model phrases in a natural language mapped to a particular semantic content. Today data-driven models are making their way into the NLG domain. In this thesis, we follow along this new line of research and present several novel data-driven approaches to natural language generation. In our work we focus on two important aspects of NLG systems development: building an efficient generator and diversifying its output. Two key ideas that we defend here are the following: first, the task of NLG can be regarded as the translation between a natural language and a formal meaning representation, and therefore, can be performed using statistical machine translation techniques, and second, corpus extension and diversification which traditionally involved manual paraphrasing and rule crafting can be performed automatically using well-known and widely used synonym and paraphrase extraction methods. Concerning our first idea, we investigate the possibility of using NGRAM translation framework and explore the potential of discriminative learning, notably Conditional Random Fields (CRF) models, as applied to NLG; we build a generation pipeline which allows for inclusion and combination of different generation models (NGRAM and CRF) and which uses an efficient decoding framework (finite-state transducers' best path search). Regarding the second objective, namely corpus extension, we propose to enlarge the system's vocabulary and the set of available syntactic structures via integrating automatically obtained synonyms and paraphrases into the training corpus. To our knowledge, there have been no attempts to increase the size of the system vocabulary by incorporating synonyms. To date most studies on corpus extension focused on paraphrasing and resorted to crowd-sourcing in order to obtain paraphrases, which then required additional manual validation often performed by system developers. We prove that automatic corpus extension by means of paraphrase extraction and validation is just as effective as crowd-sourcing, being at the same time less costly in terms of development time and resources. During intermediate experiments our generation models showed a significantly better performance than the phrase-based baseline model and appeared to be more robust in handling unknown combinations of concepts than the current in-house rule-based generator. The final human evaluation confirmed that our data-driven NLG models is a viable alternative to rule-based generators.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第22799号 / 情博第729号 / 新制||情||125(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 河原 達也, 教授 黒橋 禎夫, 教授 森 信介 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Language is critical to establishing long-term cooperative relationships among intelligent agents (including people), particularly when the agents' preferences are in conflict. In such scenarios, an agent uses speech to coordinate and negotiate behavior with its partner(s). While recent work has shown that neural language modeling can produce effective speech agents, such algorithms typically only accept previous text as input. However, in relationships among intelligent agents, not all relevant context is expressed in conversation. Thus, in this paper, we propose and analyze an algorithm, called Llumi, that incorporates other forms of context to learn to speak in long-term relationships modeled as repeated games with cheap talk. Llumi combines models of intentionality with neural language modeling techniques to learn speech from data that is relevant to the agent's current context. A user study illustrates that, while imperfect, Llumi does learn context-aware speech repeated games with cheap talk when partnered with people, including games in which it was not trained. We believe these results are useful in determining how autonomous agents can learn to use speech to facilitate successful human-agent teaming.

Nos travaux s'inscrivent dans le domaine des résumés linguistiques flous (RLF) qui permettent la génération de phrases en langage naturel, descriptives de données numériques, et offrent ainsi une vision synthétique et compréhensible de grandes masses d'information. Nous nous intéressons d'abord à l'interprétabilité des RLF, capitale pour fournir une vision simplement appréhendable de l'information à un utilisateur humain et complexe du fait de sa formulation linguistique. En plus des travaux existant à ce sujet sur les composants élémentaires des RLF, nous proposons une approche globale de l'interprétabilité des résumés vus comme un ensemble de phrases et nous intéressons plus spécifiquement à la question de leur cohérence. Afin de la garantir dans le cadre de la logique floue standard, nous introduisons une formalisation originale de l'opposition entre phrases de complexité croissante. Ce formalisme nous permet de démontrer que les propriétés de cohérence sont vérifiables par le choix d'un modèle de négation spécifique. D'autre part, nous proposons sur cette base un cube en 4 dimensions mettant en relation toutes les oppositions possibles entre les phrases d'un RLF et montrons que ce cube généralise plusieurs structures d'opposition logiques existantes. Nous considérons ensuite le cas de données sous forme de séries numériques et nous intéressons à des résumés linguistiques portant sur leur périodicité : les phrases que nous proposons indiquent à quel point une série est périodique et proposent une formulation linguistique appropriée de sa période. La méthode d’extraction proposée, nommée DPE pour Detection of Periodic Events, permet de segmenter les données de manière adaptative et sans paramètre utilisateur, en utilisant des outils issus de la morphologie mathématique. Ces segments sont ensuite utilisés pour calculer la période de la série temporelle ainsi que sa périodicité, calculée comme un degré de qualité sur le résultat renvoyé mesurant à quel point la série est périodique. Enfin, DPE génère des phrases comme « Environ toutes les 2 heures, l'afflux de client est important ». Des expériences sur des données artificielles et réelles confirment la pertinence de l'approche. D’un point de vue algorithmique, nous proposons une implémentation incrémentale et efficace de DPE, basée sur l’établissement de formules permettant le calcul de mises à jour des variables. Cette implémentation permet le passage à l'échelle de la méthode ainsi que l'analyse en temps réel de flux de données. Nous proposons également une extension de DPE basée sur le concept de périodicité locale permettant d'identifier les sous-séquences périodiques d'une série temporelle par l’utilisation d’un test statistique original. La méthode, validée sur des données artificielles et réelles, génère des phrases en langage naturel permettant d’extraire des informations du type « Toutes les deux semaines sur le premier semestre de l'année, les ventes sont élevées ». / Our research is in the field of fuzzy linguistic summaries (FLS) that allow to generate natural language sentences to describe very large amounts of numerical data, providing concise and intelligible views of these data. We first focus on the interpretability of FLS, crucial to provide end-users with an easily understandable text, but hard to achieve due to its linguistic form. Beyond existing works on that topic, based on the basic components of FLS, we propose a general approach for the interpretability of summaries, considering them globally as groups of sentences. We focus more specifically on their consistency. In order to guarantee it in the framework of standard fuzzy logic, we introduce a new model of oppositions between increasingly complex sentences. The model allows us to show that these consistency properties can be satisfied by selecting a specific negation approach. Moreover, based on this model, we design a 4-dimensional cube displaying all the possible oppositions between sentences in a FLS and show that it generalises several existing logical opposition structures. We then consider the case of data in the form of numerical series and focus on linguistic summaries about their periodicity: the sentences we propose indicate the extent to which the series are periodic and offer an appropriate linguistic expression of their periods. The proposed extraction method, called DPE, standing for Detection of Periodic Events, splits the data in an adaptive manner and without any prior information, using tools from mathematical morphology. The segments are then exploited to compute the period and the periodicity, measuring the quality of the estimation and the extent to which the series is periodic. Lastly, DPE returns descriptive sentences of the form ``Approximately every 2 hours, the customer arrival is important''. Experiments with artificial and real data show the relevance of the proposed DPE method. From an algorithmic point of view, we propose an incremental and efficient implementation of DPE, based on established update formulas. This implementation makes DPE scalable and allows it to process real-time streams of data. We also present an extension of DPE based on the local periodicity concept, allowing the identification of local periodic subsequences in a numerical series, using an original statistical test. The method validated on artificial and real data returns natural language sentences that extract information of the form ``Every two weeks during the first semester of the year, sales are high''.

Natural language generation is a method used to produce understandable texts in human languages from data [1]. Grammatical Framework is a grammar formalism and a functional programming language using a nonstatistical approach to build natural language applications. It separates the semantics and the syntax - achieving multilingualism by mapping the same semantic model to several syntaxes [2]. Grammatical Framework also has a large library called the Resource Grammar Library which serves the programmer pre-made functions in over 30 languages ready to be used to build words and sentences [3]. This report investigates if Grammatical Framework can be successfully used to perform natural language generation in order to create Wikipedia articles from data taken from Wikidata. A grammar and a program has been built to generate articles in Swedish for urban areas in Sweden. The grammar has been built around the structure of the first three sentences in the Swedish article about the urban area Linköping. Furthermore, the grammar and program is extended in order to support generation of the same articles in English and French. The results show that Grammatical Framework can be somewhat successfully used to generate small Wikipedia articles in different languages using data from Wikidata as input. While all texts were coherent, the Swedish texts were the ones having the least amount of grammatical mistakes. The biggest drawback is the rule of no pattern matching on run-time arguments, which severely limits the programmer since many functions in the resource grammar library use pattern matching internally. Even though Grammatical Framework does not solve the whole problem, it serves as a powerful enough tool to be suitable for natural language generation, with the main advantage being that it relieves the programmer from needing to pay attention to tasks related to grammar such as inflection and gender agreement. / Textgeneration är en metod som används för att generera naturlig text från data. Grammatical Framework är en grammatikformalism och ett funktionellt programmeringsspråk som använder ett ickestatistiskt tillvägagångssätt för att skapa språkteknologiska applikationer. Grammatical Framework separerar semantik och syntax, och uppnår flerspråkighet genom att länka samma semantiska model till flera syntaxer. Grammatical Framework har också ett stort bibliotek, en resursgrammatik, kallad Resource Grammar Library, som tillhandahåller applikations-programmeraren färdiga funktioner i över 30 språk redo att användas för att skapa ord och meningar. Syftet med den här rapporten är att undersöka om Grammatical Framework på ett framgångsrikt sätt kan användas för att generera Wikipedia-artiklar genom att använda data taget från Wikidata. En grammatik och ett program har skapats för att generera artiklar på svenska för svenska tätorter. Grammatiken använder de tre första meningarna i den svenska artikeln om tätorten Linköping som textstruktur. Vidare utökas grammatiken och programmet till att kunna generera samma artiklar på engelska och franska. Resultaten visar att Grammatical Framework är någorlunda framgångsrik när det kommer till att generera små Wikipedia-artiklar på olika språk. Fastän alla texter var läsbara, så hade de svenska texterna minst antal grammatiska fel. Den största nackdelen är den regel i Grammatical Framework som inte tillåter mönstermatchning med run-time argument, vilket begränsar programmeraren då många funktioner i resursgrammatiken använder möstermatching internt på sina argument. Även om Grammatical Framework inte löser hela problemet så är det ett tillräckligt kraftfullt verktyg för att vara lämpat till att användas vid textgenerering, där den största fördelen är att den avlastar programmeraren från att behöva tänka på böjning och andra grammatiska aspekter.

In this thesis, we propose an end-to-end deep learning model for a question generation task. Given a Wikipedia article written in English and a segment of text appearing in the article, the model can generate a simple question whose answer is the given text segment. The model is based on an encoder-decoder architecture. Our experiments show that a model with a fine-tuned BERT encoder and a self-attention decoder give the best performance. We also propose an evaluation metric for the question generation task, which evaluates both syntactic correctness and relevance of the generated questions. According to our analysis on sampled data, the new metric is found to give better evaluation compared to other popular metrics for sequence to sequence tasks. / I den här avhandlingen presenteras en djup neural nätverksmodell för en frågeställningsuppgift. Givet en Wikipediaartikel skriven på engelska och ett textsegment i artikeln kan modellen generera en enkel fråga vars svar är det givna textsegmentet. Modellen är baserad på en kodar-avkodararkitektur (encoderdecoder architecture). Våra experiment visar att en modell med en finjusterad BERT-kodare och en självuppmärksamhetsavkodare (self-attention decoder) ger bästa prestanda. Vi föreslår också en utvärderingsmetrik för frågeställningsuppgiften, som utvärderar både syntaktisk korrekthet och relevans för de genererade frågorna. Enligt vår analys av samplade data visar det sig att den nya metriken ger bättre utvärdering jämfört med andra populära metriker för utvärdering.

Ce mémoire de maîtrise présente le développement d’un système de génération de la langue naturelle pour automatiser les lettres de contact envoyées par les chasseurs de tête. Les travaux de Ehud Reiter ont inspiré la portion de génération de texte. La génération du contenu est basée sur des règles d’associations obtenues par l’analyse statistique d’une base de données de profils LinkedIn. Le système écrit des lettres en anglais mais peut être facilement étendu à la langue française. Ce projet s’inscrit dans le cadre du Butterfly Predictive Project, une collaboration entre l’Université de Montréal et LittleBIGJob. / This master’s thesis presents the development of a Natural Language Generation system designed to automate the writing of first-contact letters by professional headhunters. A top-down approach modelled on Ehud Reiter’s work handles the Natural Language portion of the system. Content generation is based on association rules obtained by statistical analysis of a large database of LinkedIn profiles. The system writes English letters but can easily be extended to French. This project is part of the Butterfly Predictive Project, a collaboration between Université de Montréal and LittleBIGJob.