This thesis explores a range of approaches to the task of recognising textual entailment (RTE), i.e. determining whether one text snippet entails another, for Arabic, where we are faced with an exceptional level of lexical and structural ambiguity. To the best of our knowledge, this is the first attempt to carry out this task for Arabic. Tree edit distance (TED) has been widely used as a component of natural language processing (NLP) systems that attempt to achieve the goal above, with the distance between pairs of dependency trees being taken as a measure of the likelihood that one entails the other. Such a technique relies on having accurate linguistic analyses. Obtaining such analyses for Arabic is notoriously difficult. To overcome these problems we have investigated strategies for improving tagging and parsing depending on system combination techniques. These strategies lead to substantially better performance than any of the contributing tools. We describe also a semi-automatic technique for creating a first dataset for RTE for Arabic using an extension of the ‘headline-lead paragraph’ technique because there are, again to the best of our knowledge, no such datasets available. We sketch the difficulties inherent in volunteer annotators-based judgment, and describe a regime to ameliorate some of these. The major contribution of this thesis is the introduction of two ways of improving the standard TED: (i) we present a novel approach, extended TED (ETED), for extending the standard TED algorithm for calculating the distance between two trees by allowing operations to apply to subtrees, rather than just to single nodes. This leads to useful improvements over the performance of the standard TED for determining entailment. The key here is that subtrees tend to correspond to single information units. By treating operations on subtrees as less costly than the corresponding set of individual node operations, ETED concentrates on entire information units, which are a more appropriate granularity than individual words for considering entailment relations; and (ii) we use the artificial bee colony (ABC) algorithm to automatically estimate the cost of edit operations for single nodes and subtrees and to determine thresholds, since assigning an appropriate cost to each edit operation manually can become a tricky task.The current findings are encouraging. These extensions can substantially affect the F-score and accuracy and achieve a better RTE model when compared with a number of string-based algorithms and the standard TED approaches. The relative performance of the standard techniques on our Arabic test set replicates the results reported for these techniques for English test sets. We have also applied ETED with ABC to the English RTE2 test set, where it again outperforms the standard TED.

Defending models against Natural Language Processing adversarial attacks is a challenge because of the discrete nature of the text dataset. However, given the variety of Natural Language Processing applications, it is important to make text processing models more robust and secure. This paper aims to develop techniques that will help text processing models such as BERT to combat adversarial samples that contain misspellings. These developed models are more robust than off the shelf spelling checkers.

<p>In this thesis, I deal with the task of computationally processing stories with a focus on multidisciplinary ends, specifically in Digital Humanities and Cultural Analytics. In the process, I collect, clean, investigate, and predict from two datasets. The first is a dataset of 2,302 open-source literary works categorized by the time period they are set in. These works were all collected from Project Gutenberg. The classification of the time period in which the work is set was discovered by collecting and inspecting Library of Congress subject classifications, Wikipedia Categories, and literary factsheets from SparkNotes. The second is a dataset of 6,991 open-source literary works categorized by the hierarchical location the work is set in; these labels were constructed from Library of Congress subject classifications and SparkNotes factsheets. These datasets are the first of their kind and can help move forward an understanding of 1) the presentation of settings in stories and 2) the effect the settings have on our understanding of the stories.</p>

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

What do you think of when you hear the word chatbot? A helpful assistant when booking flight tickets? Maybe a frustrating encounter with a company’s customer support, or smart technologies that will eventually take over your job? The field of chatbots is under constant development and bots are more and more taking a place in our everyday life, but how well do they really understand us humans?  The objective of this thesis is to investigate how capable two open source chatbots are in understanding human language when given input containing spelling errors, synonyms or faulty syntax. The study will further investigate if the bots get better at identifying what the user’s intention is when supplied with more training data to base their analysis on.  Two different chatbot frameworks, Botpress and Rasa, were consulted to execute this experiment. The two bots were created with basic configurations and trained using the same data. The chatbots underwent three rounds of training and testing, where they were given additional training and asked control questions to see if they managed to interpret the correct intent. All tests were documented and scores were calculated to create comparable data. The results from these tests showed that both chatbots performed well when it came to simpler spelling errors and syntax variations. Their understanding of more complex spelling errors were lower in the first testing phase but increased with more training data. Synonyms followed a similar pattern, but showed a minor tendency towards becoming overconfident and producing incorrect results with a high confidence in the last phase. The scores pointed to both chatbots getting better at understanding the input when receiving additional training. In conclusion, both chatbots showed signs of understanding language variations when given minimal training, but got significantly better results when provided with more data. The potential to create a bot with a substantial understanding of human language is evident with these results, even for developers who are previously not experienced with creating chatbots, also taking into consideration the vast possibilities to customise your chatbot.

Measuring semantic similarity is a task needed in many Natural Language Processing (NLP) applications. For example, in Machine Translation evaluation, semantic similarity is used to assess the quality of the machine translation output by measuring the degree of equivalence between a reference translation and the machine translation output. The problem of semantic similarity (Corley and Mihalcea, 2005) is de ned as measuring and recognising semantic relations between two texts. Semantic similarity covers di erent types of semantic relations, mainly bidirectional and directional. This thesis proposes new methods to address the limitations of existing work on both types of semantic relations. Recognising Textual Entailment (RTE) is a directional relation where a text T entails the hypothesis H (entailment pair) if the meaning of H can be inferred from the meaning of T (Dagan and Glickman, 2005; Dagan et al., 2013). Most of the RTE methods rely on machine learning algorithms. de Marne e et al. (2006) propose a multi-stage architecture where a rst stage determines an alignment between the T-H pairs to be followed by an entailment decision stage. A limitation of such approaches is that instead of recognising a non-entailment, an alignment that ts an optimisation criterion will be returned, but the alignment by itself is a poor predictor for iii non-entailment. We propose an RTE method following a multi-stage architecture, where both stages are based on semantic representations. Furthermore, instead of using simple similarity metrics to predict the entailment decision, we use a Markov Logic Network (MLN). The MLN is based on rich relational features extracted from the output of the predicate-argument alignment structures between T-H pairs. This MLN learns to reward pairs with similar predicates and similar arguments, and penalise pairs otherwise. The proposed methods show promising results. A source of errors was found to be the alignment step, which has low coverage. However, we show that when an alignment is found, the relational features improve the nal entailment decision. The task of Semantic Textual Similarity (STS) (Agirre et al., 2012) is de- ned as measuring the degree of bidirectional semantic equivalence between a pair of texts. The STS evaluation campaigns use datasets that consist of pairs of texts from NLP tasks such as Paraphrasing and Machine Translation evaluation. Methods for STS are commonly based on computing similarity metrics between the pair of sentences, where the similarity scores are used as features to train regression algorithms. Existing methods for STS achieve high performances over certain tasks, but poor results over others, particularly on unknown (surprise) tasks. Our solution to alleviate this unbalanced performances is to model STS in the context of Multi-task Learning using Gaussian Processes (MTL-GP) ( Alvarez et al., 2012) and state-of-the-art iv STS features ( Sari c et al., 2012). We show that the MTL-GP outperforms previous work on the same datasets.

When computers generate text, they have to consider how to describe the entities mentioned in the text. This situation becomes more difficult when the audience is unknown, as it is not clear what information is available to the addressees. This thesis investigates generation of descriptions in situations when an algorithm does not have a precise model of addressee's knowledge. This thesis starts with the collection and analysis of a corpus of descriptions of famous people. The analysis of the corpus revealed a number of useful patterns, which informed the remainder of this thesis. One of the difficult questions is how to choose information that helps addressees identify the described person. This thesis introduces a corpus-based method for determining which properties are more likely to be known by the addressees, and a probability-based method to identify properties that are distinguishing. One of the patterns observed in the collected corpus is the inclusion of multiple properties each of which uniquely identifies the referent. This thesis introduces a novel corpus-based method for determining how many properties to include in a description. Finally, a number of algorithms that leverage the findings of the corpus analysis and their computational implementation are proposed and tested in an evaluation involving human participants. The proposed algorithms outperformed the Incremental Algorithm in terms of numbers of correctly identified referents and in terms of providing a better mental image of the referent. The main contributions of this thesis are: (1) a corpus-based analysis of descriptions produced for an unknown audience; (2) a computational heuristic for estimating what information is likely to be known to addressees; and (3) algorithms that can generate referring expressions that benefit addressees without having an explicit model of addressee's knowledge.

The question of whether a machine can reproduce human intelligence is older than modern computation, but has received a great deal of attention since the first digital computers emerged decades ago. Language understanding, a hallmark of human intelligence, has been the focus of a great deal of work in Artificial Intelligence (AI). In 1950, mathematician Alan Turing proposed a kind of game, or test, to evaluate the intelligence of a machine by assessing its ability to understand written natural language. But nearly sixty years after Turing proposed his test of machine intelligence—pose questions to a machine and a person without seeing either, and try to determine which is the machine—no system has passed the Turing Test, and the question of whether a machine can understand natural language cannot yet be answered. The present investigation is, firstly, an attempt to advance the state of the art in natural language understanding by building a machine whose input is English natural language and whose output is a set of assertions that represent answers to certain questions posed about the content of the input. The machine we explore here, in other words, should pass a simplified version of the Turing Test and by doing so help clarify and expand on our understanding of the machine intelligence. Toward this goal, we explore a constraint framework for partial solutions to the Turing Test, propose a problem whose solution would constitute a significant advance in natural language processing, and design and implement a system adequate for addressing the problem proposed. The fully implemented system finds primary specific events and their locations in monologue discourse using a hierarchical classification approach, and as such provides answers to questions of central importance in the interpretation of discourse. / text

The central problem of the dissertation concerns the possibility of a distinction between truth-evaluable and non-truth-evaluable utterances of a natural language. The class of truth-evaluable utterances includes assertions, con. ectures and other kinds of speech act susceptible of truth evaluation. The class of non-truth-evaluable utterances includes commands, exhortations, wishes i.e. utterances not evaluated as being true or false. The problem is placed in the context of radical interpretation theory and it shown that it is a substantial problem of Davidson‘s early theory of radical interpret at ion. I consider the possibility of distinguishing between locutionary and illocutionary act in uttering a sentence and its significance in the present project. I discuss the suggestion that the mood of the verb of the sentence signifies the required distinction between truth-evaluable utterances and non-truth-evaluable ones. I argue that no criterion for the distinction based on the mood of the verb is adequate. The solution that I propose to the problem of classifylng truth-evaluable utterances appeals to mental states. The view that grounds this line of inquiry is that the truth-evaluability of an utterance is a characteristic of it exclusively relevant to the doxastic dimension of the speaker’s mind. I discuss the constraints that the nature of radical interpretation puts upon the way we construe the notion of belief. I propose that a possible classification of mental states into doxastic and non-doxastic that would result in a classification of utterances into truth-evaluable and non-truthevaluable ones can be given by an elaborated version of a decision theoretic scheme. I suggest that a decision theoretic scheme based on a decision theory that, like Savage’s theory, grants independence axioms is a better candidate to offer a solution to the central problem of the dissertation than a scheme based on a non- standard decision theory such as Richard Jeffrey’s. I conclude by showing that the proposal I make satisfies the constraints I have considered and that it can be accommodated by a radical interpretation theory.

This thesis describes the development and evaluation of new automated Question Answering (QA) methods tailored to clinical comparison questions that give clinicians a rank-ordered list of MEDLINE® abstracts targeted to natural language clinical drug comparison questions (e.g. ”Have any studies directly compared the effects of Pioglitazone and Rosiglitazone on the liver?”). Three corpora were created to develop and evaluate a new QA system for clinical comparison questions called RetroRank. RetroRank takes the clinician’s plain text question as input, processes it and outputs a rank-ordered list of potential answer candidates, i.e. MEDLINE® abstracts, that is reordered using new post-retrieval ranking strategies to ensure the most topically-relevant abstracts are displayed as high in the result set as possible. RetroRank achieves a significant improvement over the PubMed recency baseline and performs equal to or better than previous approaches to post-retrieval ranking relying on query frames and annotated data such as the approach by Demner-Fushman and Lin (2007). The performance of RetroRank shows that it is possible to successfully use natural language input and a fully automated approach to obtain answers to clinical drug comparison questions. This thesis also introduces two new evaluation corpora of clinical comparison questions with “gold standard” references that are freely available and are a valuable resource for future research in medical QA.