Ce travail de thèse propose une méthodologie visant la résolution de certains problèmes de classification, notamment ceux concernant la classification séquentielle en tâches de Traitement Automatique des Langues. Afin d'améliorer les résultats de la tâche de classification, nous proposons l'utilisation d'une approche itérative basée sur l'erreur, qui intègre, dans le processus d'apprentissage, des connaissances d'un expert représentées sous la forme de "règles de connaissance". Nous avons appliqué la méthodologie à deux tâches liées à la détection de la spéculation ("hedging") dans la littérature scientifique: la détection de segments textuels spéculatifs ("hedge cue identification") et la détection de la couverture de ces segments ("hedge cue scope detection"). Les résultats son prometteurs: pour la première tâche, nous avons amélioré le F-score de la baseline de 2,5 points en intégrant des données sur la co-occurrence de segments spéculatifs. Concernant la deuxième tâche, l'intégration d'information syntaxique et des règles pour l'élagage syntaxique ont permis d'améliorer les résultats de la classification de 0,712 à 0,835 (F-score). Par rapport aux méthodes de l'état de l'art, les résultats sont très bons et ils suggèrent que l'approche consistant à améliorer les classifieurs basées seulement sur des erreurs commises dans un corpus, peut être également appliquée à d'autres tâches similaires. Qui plus est, ce travail de thèse propose un schéma de classes permettant de représenter l'analyse d'une phrase dans une structure unique qui intègre les résultats de différentes analyses linguistiques. Cela permet de mieux gérer le processus itératif d'amélioration du classifieur, dans lequel différents ensembles d'attributs d'apprentissage sont utilisés à chaque itération. Nous proposons également de stocker les attributs dans un modèle relationnel au lieu des structures textuelles classiques, afin de faciliter l'analyse et la manipulation des données apprises.

While linguistic skill is a hallmark of humanity, the increasing volume of linguistic data each of us faces is causing individual and societal problems — ‘information overload’ is a commonly discussed condition. Tasks such as finding the most appropriate information online, understanding the contents of a personal email repository, and translating documents from another language are now commonplace. These tasks need not cause stress and feelings of overload: the human intellectual capacity is not the problem. Rather, the computational interfaces to linguistic data are problematic — there exists a Linguistic Visualization Divide in the current state-of-the-art. Through five design studies, this dissertation combines sophisticated natural language processing algorithms with information visualization techniques grounded in evidence of human visuospatial capabilities. The first design study, Uncertainty Lattices, augments real-time computermediated communication, such as cross-language instant messaging chat and automatic speech recognition. By providing explicit indications of algorithmic confidence, the visualization enables informed decisions about the quality of computational outputs. Two design studies explore the space of content analysis. DocuBurst is an interactive visualization of document content, which spatially organizes words using an expert-created ontology. Broadening from single documents to document collections, Parallel Tag Clouds combine keyword extraction and coordinated visualizations to provide comparative overviews across subsets of a faceted text corpus. Finally, two studies address visualization for natural language processing research. The Bubble Sets visualization draws secondary set relations around arbitrary collections of items, such as a linguistic parse tree. From this design study we propose a theory of spatial rights to consider when assigning visual encodings to data. Expanding considerations of spatial rights, we present a formalism to organize the variety of approaches to coordinated and linked visualization, and introduce VisLink, a new method to relate and explore multiple 2d visualizations in 3d space. Intervisualization connections allow for cross-visualization queries and support high level comparison between visualizations. From the design studies we distill challenges common to visualizing language data, including maintaining legibility, supporting detailed reading, addressing data scale challenges, and managing problems arising from semantic ambiguity.

Neologisms, or newly-coined words, pose problems for natural language processing (NLP) systems. Due to the recency of their coinage, neologisms are typically not listed in computational lexicons---dictionary-like resources that many NLP applications depend on. Therefore when a neologism is encountered in a text being processed, the performance of an NLP system will likely suffer due to the missing word-level information. Identifying and documenting the usage of neologisms is also a challenge in lexicography, the making of dictionaries. The traditional approach to these tasks has been to manually read a lot of text. However, due to the vast quantities of text being produced nowadays, particularly in electronic media such as blogs, it is no longer possible to manually analyze it all in search of neologisms. Methods for automatically identifying and inferring syntactic and semantic properties of neologisms would therefore address problems encountered in both natural language processing and lexicography. Because neologisms are typically infrequent due to their recent addition to the language, approaches to automatically learning word-level information relying on statistical distributional information are in many cases inappropriate. Moreover, neologisms occur in many domains and genres, and therefore approaches relying on domain-specific resources are also inappropriate. The hypothesis of this thesis is that knowledge about etymology---including word formation processes and types of semantic change---can be exploited for the acquisition of aspects of the syntax and semantics of neologisms. Evidence supporting this hypothesis is found in three case studies: lexical blends (e.g., "webisode" a blend of "web" and "episode"), text messaging forms (e.g., "any1" for "anyone"), and ameliorations and pejorations (e.g., the use of "sick" to mean `excellent', an amelioration). Moreover, this thesis presents the first computational work on lexical blends and ameliorations and pejorations, and the first unsupervised approach to text message normalization.

With a growing amount of subjective content distributed across the Web, there is a need for a domain-independent information retrieval system that would support ad hoc retrieval of documents expressing opinions on a specific topic of the user’s query. While the research area of opinion detection and sentiment analysis has received much attention in the recent years, little research has been done on identifying subjective content targeted at a specific topic, i.e. expressing topical opinion. This thesis presents a novel method for ad hoc retrieval of documents which contain subjective content on the topic of the query. Documents are ranked by the likelihood each document expresses an opinion on a query term, approximated as the likelihood any occurrence of the query term is modified by a subjective adjective. Domain-independent user-based evaluation of the proposed methods was conducted, and shows statistically significant gains over Google ranking as the baseline.

This thesis investigates how word sense disambiguation affects lexical chains, as well as proposing an improved model for lexical chaining in which word sense disambiguation is performed prior to lexical chaining. A lexical chain is a set of words from a document that are related in meaning. Lexical chains can be used to identify the dominant topics in a document, as well as where changes in topic occur. This makes them useful for applications such as topic segmentation and document summarization. <br /><br /> However, polysemous words are an inherent problem for algorithms that find lexical chains as the intended meaning of a polysemous word must be determined before its semantic relations to other words can be determined. For example, the word "bank" should only be placed in a chain with "money" if in the context of the document "bank" refers to a place that deals with money, rather than a river bank. The process by which the intended senses of polysemous words are determined is word sense disambiguation. To date, lexical chaining algorithms have performed word sense disambiguation as part of the overall process building lexical chains. Because the intended senses of polysemous words must be determined before words can be properly chained, we propose that word sense disambiguation should be performed before lexical chaining occurs. Furthermore, if word sense disambiguation is performed prior to lexical chaining, then it can be done with any available disambiguation method, without regard to how lexical chains will be built afterwards. Therefore, the most accurate available method for word sense disambiguation should be applied prior to the creation of lexical chains. <br /><br /> We perform an experiment to demonstrate the validity of the proposed model. We compare the lexical chains produced in two cases: <ol> <li>Lexical chaining is performed as normal on a corpus of documents that has not been disambiguated. </li> <li>Lexical chaining is performed on the same corpus, but all the words have been correctly disambiguated beforehand. </li></ol> We show that the lexical chains created in the second case are more correct than the chains created in the first. This result demonstrates that accurate word sense disambiguation performed prior to the creation of lexical chains does lead to better lexical chains being produced, confirming that our model for lexical chaining is an improvement upon previous approaches.

A <em>lexical analogy</em> is two pairs of words (<em>w</em>₁, <em>w</em>₂) and (<em>w</em>₃, <em>w</em>₄) such that the relation between <em>w</em>₁ and <em>w</em>₂ is identical or similar to the relation between <em>w</em>₃ and <em>w</em>₄. For example, (<em>abbreviation</em>, <em>word</em>) forms a lexical analogy with (<em>abstract</em>, <em>report</em>), because in both cases the former is a shortened version of the latter. Lexical analogies are of theoretic interest because they represent a second order similarity measure: <em>relational similarity</em>. Lexical analogies are also of practical importance in many applications, including text-understanding and learning ontological relations. <BR> <BR> This thesis presents a novel system that generates lexical analogies from a corpus of text documents. The system is motivated by a well-established theory of analogy-making, and views lexical analogy generation as a series of three processes: identifying pairs of words that are semantically related, finding clues to characterize their relations, and generating lexical analogies by matching pairs of words with similar relations. The system uses a <em>dependency grammar</em> to characterize semantic relations, and applies machine learning techniques to determine their similarities. Empirical evaluation shows that the system performs remarkably well, generating lexical analogies at a precision of over 90%.

This study seeks to explore and develop innovative methods for the extraction of semantic knowledge from unlabelled written English documents and the representation of this knowledge using a formal mathematical expression to facilitate its use in practical applications. The first method developed in this research focuses on semantic information extraction. To perform this task, the study introduces a natural language processing (NLP) method designed to extract information-rich keywords from English sentences. The method involves initially learning a set of rules that guide the extraction of keywords from parts of sentences. Once this learning stage is completed, the method can be used to extract the keywords from complete sentences by pairing these sentences to the most similar sequence of rules. The key innovation in this method is the use of a part-of-speech hierarchy. By raising words to increasingly general grammatical categories in this hierarchy, the system can compare rules, compute the degree of similarity between them, and learn new rules. The second method developed in this study addresses the problem of knowledge representation. This method processes triplets of keywords through several successive steps to represent information contained in the triplets using possibility distributions. These distributions represent the possibility of a topic given a particular triplet of keywords. Using this methodology, the information contained in the natural language triplets can be quantified and represented in a mathematical format, which can be easily used in a number of applications, such as document classifiers. In further extensions to the research, a theoretical justification and mathematical development for both methods are provided, and examples are given to illustrate these notions. Sample applications are also developed based on these methods, and the experimental results generated through these implementations are expounded and thoroughly analyzed to confirm that the methods are reliable in practice.

This thesis investigates coreference and its resolution within the domain of health education. Coreference is the relationship between two linguistic expressions that refer to the same real-world entity, and resolution involves identifying this relationship among sets of referring expressions. The coreference resolution task is considered among the most difficult of problems in Artificial Intelligence; in some cases, resolution is impossible even for humans. For example, "she" in the sentence "Lynn called Jennifer while she was on vacation" is genuinely ambiguous: the vacationer could be either Lynn or Jennifer. <br/><br/> There are three primary motivations for this thesis. The first is that health education has never before been studied in this context. So far, the vast majority of coreference research has focused on news. Secondly, achieving domain-independent resolution is unlikely without understanding the extent to which coreference varies across different genres. Finally, coreference pervades language and is an essential part of coherent discourse. Its effective use is a key component of easy-to-understand health education materials, where readability is paramount. <br/><br/> No suitable corpus of health education materials existed, so our first step was to create one. The comprehensive analysis of this corpus, which required manual annotation of coreference, confirmed our hypothesis that the coreference used in health education differs substantially from that in previously studied domains. This analysis was then used to shape the design of a knowledge-lean algorithm for resolving coreference. This algorithm performed surprisingly well on this corpus, e.g., successfully resolving over 85% of all pronouns when evaluated on unseen data. <br/><br/> Despite the importance of coreferentially annotated corpora, only a handful are known to exist, likely because of the difficulty and cost of reliably annotating coreference. The paucity of genres represented in these existing annotated corpora creates an implicit bias in domain-independent coreference resolution. In an effort to address these issues, we plan to make our health education corpus available to the wider research community, hopefully encouraging a broader focus in the future.

Open source projects do have requirements; they are, however, mostly informal, text descriptions found in requests, forums, and other correspondence. Understanding such requirements provides insight into the nature of open source projects. Unfortunately, manual analysis of natural language requirements is time-consuming, and for large projects, error-prone. Automated analysis of natural language requirements, even partial, will be of great benefit. Towards that end, I describe the design and validation of an automated natural language requirements classifier for open source software development projects. I compare two strategies for recognizing requirements in open forums of software features. The results suggest that classifying text at the forum post aggregation and sentence aggregation levels may be effective. Initial results suggest that it can reduce the effort required to analyze requirements of open source software development projects. Software development organizations and communities currently employ a large number of software development techniques and methodologies. This implied complexity is also enhanced by a wide range of software project types and development environments. The resulting lack of consistency in the software development domain leads to one important challenge that researchers encounter while exploring this area: specificity. This results in an increased difficulty of maintaining a consistent unit of measure or analysis approach while exploring a wide variety of software development projects and environments. The problem of specificity is more prominently exhibited in an area of software development characterized by a dynamic evolution, a unique development environment, and a relatively young history of research when compared to traditional software development: the open-source domain. While performing research on open source and the associated communities of developers, one can notice the same challenge of specificity being present in requirements engineering research as in the case of closed-source software development. Whether research is aimed at performing longitudinal or cross-sectional analyses, or attempts to link requirements to other aspects of software development projects and their management, specificity calls for a flexible analysis tool capable of adapting to the needs and specifics of the explored context. This dissertation covers the design, implementation, and evaluation of a model, a method, and a software tool comprising a flexible software development analysis framework. These design artifacts use a rule-based natural language processing approach and are built to meet the specifics of a requirements-based analysis of software development projects in the open-source domain. This research follows the principles of design science research as defined by Hevner et. al. and includes stages of problem awareness, suggestion, development, evaluation, and results and conclusion (Hevner et al. 2004; Vaishnavi and Kuechler 2007). The long-term goal of the research stream stemming from this dissertation is to propose a flexible, customizable, requirements-based natural language processing software analysis framework which can be adapted to meet the research needs of multiple different types of domains or different categories of analyses.

With a growing amount of subjective content distributed across the Web, there is a need for a domain-independent information retrieval system that would support ad hoc retrieval of documents expressing opinions on a specific topic of the user’s query. While the research area of opinion detection and sentiment analysis has received much attention in the recent years, little research has been done on identifying subjective content targeted at a specific topic, i.e. expressing topical opinion. This thesis presents a novel method for ad hoc retrieval of documents which contain subjective content on the topic of the query. Documents are ranked by the likelihood each document expresses an opinion on a query term, approximated as the likelihood any occurrence of the query term is modified by a subjective adjective. Domain-independent user-based evaluation of the proposed methods was conducted, and shows statistically significant gains over Google ranking as the baseline.