La tâche de question-réponse (Question-Answering, QA) est bien ancrée dans la communauté de Traitement Automatique du Langage Naturel (TALN) depuis de nombreuses années. De manière générale, celle-ci consiste à répondre à des questions données à l’aide de documents (textuels ou autres) ou de conversations en faisant au besoin usage de connaissances et en mettant en oeuvre des mécanismes d’inférence. Ainsi, dépendamment du jeu de données et de la tâche lui étant associée, il faut que le système puisse détecter et comprendre les éléments utiles pour répondre correctement à chacune des questions posées. De nombreux progrès ont été réalisés depuis quelques années avec des modèles neuronaux de plus en plus complexes, ces derniers sont cependant coûteux en production, et relativement opaques. Du à leur opacité, il est difficile d’anticiper avec précision le comportement de certains modèles et d’ainsi prévoir quand ces systèmes vont retourner de mauvaises réponses. Contrairement à la très grande majorité des systèmes proposés actuellement, nous allons dans ce mémoire tenter de résoudre cette tâche avec des modèles de taille contrôlable, on s’intéressera principalement aux approches basées sur les traits (features). Le but visé en restreignant la taille des modèles est qu’ils généralisent mieux. On pourra alors mesurer ce que ces modèles capturent afin d’évaluer la granularité de leur "compréhension" de la langue. Aussi, en analysant les lacunes de modèles de taille contrôlable, on pourra mettre en valeur ce que des modèles plus complexes ont capturé. Pour réaliser notre étude, on s’évalue ici sur SQuAD: un jeu de données populaire proposé par l’Université Standford. / The Question-Answering task (QA) is a well established Natural Language Processing (NLP) task. Generally speaking, it consists in answering questions using documents (textual or otherwise) or conversations, making use of knowledge if necessary and implementing inference mechanisms. Thus, depending on the data set and the task associated with it, the system must be able to detect and understand the useful elements to correctly answer each of the questions asked. A lot of progress has been made in recent years with increasingly complex neural models. They are however expensive in production, and relatively opaque. Due to this opacity, it is diÿcult to accurately predict the behavior of some models and thus, to predict when these systems will return wrong answers. Unlike the vast majority of systems currently proposed, in this thesis we will try to solve this task with models with controllable size. We will focus mainly on feature-based approaches. The goal in restricting the size of the models is that they generalize better. So we will measure what these models capture in order to assess the granularity of their "understanding" of the language. Also, by analyzing the gaps of controllable size models, we will be able to highlight what more complex models have captured. To carry out our study, we evaluate ourselves here on SQuAD: a popular data set o˙ered by Standford University.

Le volume et la complexité des données générées par les systèmes d’information croissent de façon singulière dans les entrepôts de données. Le domaine de l’informatique décisionnelle (aussi appelé BI) a pour objectif d’apporter des méthodes et des outils pour assister les utilisateurs dans leur tâche de recherche d’information. En effet, les sources de données ne sont en général pas centralisées, et il est souvent nécessaire d’interagir avec diverses applications. Accéder à l’information est alors une tâche ardue, alors que les employés d’une entreprise cherchent généralement à réduire leur charge de travail. Pour faire face à ce constat, le domaine « Enterprise Search » s’est développé récemment, et prend en compte les différentes sources de données appartenant aussi bien au réseau privé d’entreprise qu’au domaine public (telles que les pages Internet). Pourtant, les utilisateurs de moteurs de recherche actuels souffrent toujours de du volume trop important d’information à disposition. Nous pensons que de tels systèmes pourraient tirer parti des méthodes du traitement naturel des langues associées à celles des systèmes de questions/réponses. En effet, les interfaces en langue naturelle permettent aux utilisateurs de rechercher de l’information en utilisant leurs propres termes, et d’obtenir des réponses concises et non une liste de documents dans laquelle l’éventuelle bonne réponse doit être identifiée. De cette façon, les utilisateurs n’ont pas besoin d’employer une terminologie figée, ni de formuler des requêtes selon une syntaxe très précise, et peuvent de plus accéder plus rapidement à l’information désirée. Un challenge lors de la construction d’un tel système consiste à interagir avec les différentes applications, et donc avec les langages utilisés par ces applications d’une part, et d’être en mesure de s’adapter facilement à de nouveaux domaines d’application d’autre part. Notre rapport détaille un système de questions/réponses configurable pour des cas d’utilisation d’entreprise, et le décrit dans son intégralité. Dans les systèmes traditionnels de l’informatique décisionnelle, les préférences utilisateurs ne sont généralement pas prises en compte, ni d’ailleurs leurs situations ou leur contexte. Les systèmes état-de-l’art du domaine tels que Soda ou Safe ne génèrent pas de résultats calculés à partir de l’analyse de la situation des utilisateurs. Ce rapport introduit une approche plus personnalisée, qui convient mieux aux utilisateurs finaux. Notre expérimentation principale se traduit par une interface de type search qui affiche les résultats dans un dashboard sous la forme de graphes, de tables de faits ou encore de miniatures de pages Internet. En fonction des requêtes initiales des utilisateurs, des recommandations de requêtes sont aussi affichées en sus, et ce dans le but de réduire le temps de réponse global du système. En ce sens, ces recommandations sont comparables à des prédictions. Notre travail se traduit par les contributions suivantes : tout d’abord, une architecture implémentée via des algorithmes parallélisés et qui prend en compte la diversité des sources de données, à savoir des données structurées ou non structurées dans le cadre d’un framework de questions/réponses qui peut être facilement configuré dans des environnements différents. De plus, une approche de traduction basée sur la résolution de contrainte, qui remplace le traditionnel langage-pivot par un modèle conceptuel et qui conduit à des requêtes multidimensionnelles mieux personnalisées. En outre, en ensemble de patrons linguistiques utilisés pour traduire des questions BI en des requêtes pour bases de données, qui peuvent être facilement adaptés dans le cas de configurations différentes. / The amount and complexity of data generated by information systems keep increasing in Warehouses. The domain of Business Intelligence (BI) aims at providing methods and tools to better help users in retrieving those data. Data sources are distributed over distinct locations and are usually accessible through various applications. Looking for new information could be a tedious task, because business users try to reduce their work overload. To tackle this problem, Enterprise Search is a field that has emerged in the last few years, and that takes into consideration the different corporate data sources as well as sources available to the public (e.g. World Wide Web pages). However, corporate retrieval systems nowadays still suffer from information overload. We believe that such systems would benefit from Natural Language (NL) approaches combined with Q&A techniques. Indeed, NL interfaces allow users to search new information in their own terms, and thus obtain precise answers instead of turning to a plethora of documents. In this way, users do not have to employ exact keywords or appropriate syntax, and can have faster access to new information. Major challenges for designing such a system are to interface different applications and their underlying query languages on the one hand, and to support users’ vocabulary and to be easily configured for new application domains on the other hand. This thesis outlines an end-to-end Q&A framework for corporate use-cases that can be configured in different settings. In traditional BI systems, user-preferences are usually not taken into account, nor are their specific contextual situations. State-of-the art systems in this field, Soda and Safe do not compute search results on the basis of users’ situation. This thesis introduces a more personalized approach, which better speaks to end-users’ situations. Our main experimentation, in this case, works as a search interface, which displays search results on a dashboard that usually takes the form of charts, fact tables, and thumbnails of unstructured documents. Depending on users’ initial queries, recommendations for alternatives are also displayed, so as to reduce response time of the overall system. This process is often seen as a kind of prediction model. Our work contributes to the following: first, an architecture, implemented with parallel algorithms, that leverages different data sources, namely structured and unstructured document repositories through an extensible Q&A framework, and this framework can be easily configured for distinct corporate settings; secondly, a constraint-matching-based translation approach, which replaces a pivot language with a conceptual model and leads to more personalized multidimensional queries; thirdly, a set of NL patterns for translating BI questions in structured queries that can be easily configured in specific settings. In addition, we have implemented an iPhone/iPad™ application and an HTML front-end that demonstrate the feasibility of the various approaches developed through a series of evaluation metrics for the core component and scenario of the Q&A framework. To this end, we elaborate on a range of gold-standard queries that can be used as a basis for evaluating retrieval systems in this area, and show that our system behave similarly as the well-known WolframAlpha™ system, depending on the evaluation settings.

abstract: While in recent years deep learning (DL) based approaches have been the popular approach in developing end-to-end question answering (QA) systems, such systems lack several desired properties, such as the ability to do sophisticated reasoning with knowledge, the ability to learn using less resources and interpretability. In this thesis, I explore solutions that aim to address these drawbacks. Towards this goal, I work with a specific family of reading comprehension tasks, normally referred to as the Non-Extractive Reading Comprehension (NRC), where the given passage does not contain enough information and to correctly answer sophisticated reasoning and ``additional knowledge" is required. I have organized the NRC tasks into three categories. Here I present my solutions to the first two categories and some preliminary results on the third category. Category 1 NRC tasks refer to the scenarios where the required ``additional knowledge" is missing but there exists a decent natural language parser. For these tasks, I learn the missing ``additional knowledge" with the help of the parser and a novel inductive logic programming. The learned knowledge is then used to answer new questions. Experiments on three NRC tasks show that this approach along with providing an interpretable solution achieves better or comparable accuracy to that of the state-of-the-art DL based approaches. The category 2 NRC tasks refer to the alternate scenario where the ``additional knowledge" is available but no natural language parser works well for the sentences of the target domain. To deal with these tasks, I present a novel hybrid reasoning approach which combines symbolic and natural language inference (neural reasoning) and ultimately allows symbolic modules to reason over raw text without requiring any translation. Experiments on two NRC tasks shows its effectiveness. The category 3 neither provide the ``missing knowledge" and nor a good parser. This thesis does not provide an interpretable solution for this category but some preliminary results and analysis of a pure DL based approach. Nonetheless, the thesis shows beyond the world of pure DL based approaches, there are tools that can offer interpretable solutions for challenging tasks without using much resource and possibly with better accuracy. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2019

No description available.

Question answering systems can be seen as the next step in information retrieval, allowing users to pose questions in natural language and receive succinct answers. In order for a question answering system as a whole to be successful, research has shown that the correct classification of questions with regards to the expected answer type is imperative. Question classification has two components: a taxonomy of answer types, and a machinery for making the classifications. This thesis focuses on five different machine learning algorithms for the question classification task. The algorithms are k nearest neighbours, naïve bayes, decision tree learning, sparse network of winnows, and support vector machines. These algorithms have been applied to two different corpora, one of which has been used extensively in previous work and has been constructed for a specific agenda. The other corpus is drawn from a set of users' questions posed to a running online system. The results showed that the performance of the algorithms on the different corpora differs both in absolute terms, as well as with regards to the relative ranking of them. On the novel corpus, naïve bayes, decision tree learning, and support vector machines perform on par with each other, while on the biased corpus there is a clear difference between them, with support vector machines being the best and naïve bayes being the worst. The thesis also presents an analysis of questions that are problematic for all learning algorithms. The errors can roughly be divided as due to categories with few members, variations in question formulation, the actual usage of the taxonomy, keyword errors, and spelling errors. A large portion of the errors were also hard to explain. / <p>Report code: LiU-Tek-Lic-2007:29.</p>

The complexity for being at the forefront regarding information retrieval systems are constantly increasing. Recent technology of natural language processing called BERT has reached superhuman performance in high resource languages for reading comprehension tasks. However, several researchers has stated that multilingual model’s are not enough for low-resource languages, since they are lacking a thorough understanding of those languages. Recently, a Swedish pre-trained BERT model has been introduced which is trained on significantly more Swedish data than the multilingual models currently available. This study compares both multilingual and Swedish monolingual inherited BERT model’s for question answering utilizing both a English and a Swedish machine translated SQuADv2 data set during its fine-tuning process. The models are evaluated with SQuADv2 benchmark and within a implemented question answering system built upon the classical retriever-reader methodology. This study introduces a naive and more robust prediction method for the proposed question answering system as well finding a sweet spot for each individual model approach integrated into the system. The question answering system is evaluated and compared against another question answering library at the leading edge within the area, applying a custom crafted Swedish evaluation data set. The results show that the fine-tuned model based on the Swedish pre-trained model and the Swedish SQuADv2 data set were superior in all evaluation metrics except speed. The comparison between the different systems resulted in a higher evaluation score but a slower prediction time for this study’s system.

A natural application of artificial intelligence is to help blind people overcome their daily visual challenges through AI-based assistive technologies. In this regard, one of the most promising tasks is Visual Question Answering (VQA): the model is presented with an image and a question about this image. It must then predict the correct answer. Recently has been introduced the VizWiz dataset, a collection of images and questions originating from blind people. Being the first VQA dataset deriving from a natural setting, VizWiz presents many limitations and peculiarities. More specifically, the characteristics observed are the high uncertainty of the answers, the conversational aspect of questions, the relatively small size of the datasets and ultimately, the imbalance between answerable and unanswerable classes. These characteristics could be observed, individually or jointly, in other VQA datasets, resulting in a burden when solving the VQA task. Particularly suitable to address these aspects of the data are data science pre-processing techniques. Therefore, to provide a solid contribution to the VQA task, we answered the research question “Can data science pre-processing techniques improve the VQA task?” by proposing and studying the effects of four different pre-processing techniques. To address the high uncertainty of answers we employed a pre-processing step in which it is computed the uncertainty of each answer and used this measure to weight the soft scores of our model during training. The adoption of an “uncertainty-aware” training procedure boosted the predictive accuracy of our model of 10% providing a new state-of-the-art when evaluated on the test split of the VizWiz dataset. In order to overcome the limited amount of data, we designed and tested a new pre-processing procedure able to augment the training set and almost double its data points by computing the cosine similarity between answers representation. We addressed also the conversational aspect of questions collected from real world verbal conversations by proposing an alternative question pre-processing pipeline in which conversational terms are removed. This led in a further improvement: from a predictive accuracy of 0.516 with the standard question processing pipeline, we were able to achieve 0.527 predictive accuracy when employing the new pre-processing pipeline. Ultimately, we addressed the imbalance between answerable and unanswerable classes when predicting the answerability of a visual question. We tested two standard pre-processing techniques to adjust the dataset class distribution: oversampling and undersampling. Oversampling provided an albeit small improvement in both average precision and F1 score. / En naturlig tillämpning av artificiell intelligens är att hjälpa blinda med deras dagliga visuella utmaningar genom AI-baserad hjälpmedelsteknik. I detta avseende, är en av de mest lovande uppgifterna Visual Question Answering (VQA): modellen presenteras med en bild och en fråga om denna bild, och måste sedan förutspå det korrekta svaret. Nyligen introducerades VizWiz-datamängd, en samling bilder och frågor till dessa från blinda personer. Då detta är det första VQA-datamängden som härstammar från en naturlig miljö, har det många begränsningar och särdrag. Mer specifikt är de observerade egenskaperna: hög osäkerhet i svaren, informell samtalston i frågorna, relativt liten datamängd och slutligen obalans mellan svarbara och icke svarbara klasser. Dessa egenskaper kan även observeras, enskilda eller tillsammans, i andra VQA-datamängd, vilket utgör särskilda utmaningar vid lösning av VQA-uppgiften. Särskilt lämplig för att hantera dessa aspekter av data är förbehandlingsteknik från området data science. För att bidra till VQA-uppgiften, svarade vi därför på frågan “Kan förbehandlingstekniker från området data science bidra till lösningen av VQA-uppgiften?” genom att föreslå och studera effekten av fyra olika förbehandlingstekniker. För att hantera den höga osäkerheten i svaren använde vi ett förbehandlingssteg där vi beräknade osäkerheten i varje svar och använde detta mått för att vikta modellens utdata-värden under träning. Användandet av en ”osäkerhetsmedveten” träningsprocedur förstärkte den förutsägbara noggrannheten hos vår modell med 10%. Med detta nådde vi ett toppresultat när modellen utvärderades på testdelen av VizWiz-datamängden. För att övervinna problemet med den begränsade mängden data, konstruerade och testade vi en ny förbehandlingsprocedur som nästan dubblerar datapunkterna genom att beräkna cosinuslikheten mellan svarens vektorer. Vi hanterade även problemet med den informella samtalstonen i frågorna, som samlats in från den verkliga världens verbala konversationer, genom att föreslå en alternativ väg att förbehandla frågorna, där samtalstermer är borttagna. Detta ledde till en ytterligare förbättring: från en förutsägbar noggrannhet på 0.516 med det vanliga sättet att bearbeta frågorna kunde vi uppnå 0.527 prediktiv noggrannhet vid användning av det nya sättet att förbehandla frågorna. Slutligen hanterade vi obalansen mellan svarbara och icke svarbara klasser genom att förutse om en visuell fråga har ett möjligt svar. Vi testade två standard-förbehandlingstekniker för att justeradatamängdens klassdistribution: översampling och undersampling. Översamplingen gav en om än liten förbättring i både genomsnittlig precision och F1-poäng.

Question answering (QA) stands squarely along the path from document retrieval to text understanding. As an area of research interest, it serves as a proving ground where strategies for document processing, knowledge representation, question analysis, and answer extraction may be evaluated in real world information extraction contexts. The task is to go beyond the representation of text documents as "bags of words" or data blobs that can be scanned for keyword combinations and word collocations in the manner of internet search engines. Instead, the goal is to recognize and extract the semantic content of the text, and to organize it in a manner that supports reasoning about the concepts represented. The issue presented is how to obtain and query such a structure without either a predefined set of concepts or a predefined set of relationships among concepts. This research investigates a means for acquiring from text documents both the underlying concepts and their interrelationships. Specifically, a syntax-based formalism for representing atomic propositions that are extracted from text documents is presented, together with a method for constructing a network of concept nodes for indexing such logical forms based on the discourse entities they contain. It is shown that meaningful questions can be decomposed into Boolean combinations of question patterns using the same formalism, with free variables representing the desired answers. It is further shown that this formalism can be used for robust question answering using the concept network and WordNet synonym, hypernym, hyponym, and antonym relationships. This formalism was implemented in the Semantic Extractor (SEMEX) research tool and was tested against the factoid questions from the 2005 Text Retrieval Conference (TREC), which operated upon the AQUAINT corpus of newswire documents. After adjusting for the limitations of the tool and the document set, correct answers were found for approximately fifty percent of the questions analyzed, which compares favorably with other question answering systems.

While large language models like GPT-3 have achieved impressive results in the zero-, one-, and few-shot settings, they still significantly underperform on some tasks relative to the state of the art (SOTA). For many tasks it would be useful to have answer options explicitly listed out in a multiple choice format, decreasing computational cost and allowing the model to reason about the relative merits of possible answers. We argue that the reason this hasn't helped models like GPT-3 close the gap with the SOTA is that these models struggle with symbol binding - associating each answer option with a symbol that represents it. To ameliorate this situation we introduce index prompting, a way of leveraging language models trained on code to successfully answer multiple choice formatted questions. When used with the OpenAI Codex model, our method improves accuracy by about 18% on average in the few-shot setting relative to GPT-3 across 8 datasets representing 4 common NLP tasks. It also achieves a new single-model state of the art on ANLI R3, ARC (Easy), and StoryCloze, suggesting that GPT-3's latent "understanding" has been previously underestimated.

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第24929号 / 情博第840号 / 新制||情||140(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)特定教授 黒橋 禎夫, 教授 河原 達也, 教授 西野 恒 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM