Techniques based on artificial neural networks represent the current state-of-the-art in machine learning due to the availability of improved hardware and large data sets. Here we employ doc2vec, an unsupervised neural network, to capture the semantic content of text messages sent by adolescents during high school, and encode this semantic content as numeric vectors. These vectors effectively condense the text message data into highly leverageable inputs to a logistic regression classifier in a matter of hours, as compared to the tedious and often quite lengthy task of manually coding data. Using our machine learning approach, we are able to train a logistic regression model to predict adolescents' engagement in substance abuse during distinct life phases with accuracy ranging from 76.5% to 88.1%. We show the effects of grade level and text message aggregation strategy on the efficacy of document embedding generation with doc2vec. Additional examination of the vectorizations for specific terms extracted from the text message data adds quantitative depth to this analysis. We demonstrate the ability of the method used herein to overcome traditional natural language processing concerns related to unconventional orthography. These results suggest that the approach described in this thesis is a competitive and efficient alternative to existing methodologies for predicting substance abuse behaviors. This work reveals the potential for the application of machine learning-based manipulation of text messaging data to development of automatic intervention strategies against substance abuse and other adolescent challenges.

Identifying sentiment in a low-resource language is essential for understanding opinions internationally and for responding to the urgent needs of locals affected by disaster incidents in different world regions. While tools and resources for recognizing sentiment in high-resource languages are plentiful, determining the most effective methods for achieving this task in a low-resource language which lacks annotated data is still an open research question. Most existing approaches for cross-lingual sentiment analysis to date have relied on high-resource machine translation systems, large amounts of parallel data, or resources only available for Indo-European languages. This work presents methods, resources, and strategies for identifying sentiment cross-lingually in a low-resource language. We introduce a cross-lingual sentiment model which can be trained on a high-resource language and applied directly to a low-resource language. The model offers the feature of lexicalizing the training data using a bilingual dictionary, but can perform well without any translation into the target language. Through an extensive experimental analysis, evaluated on 17 target languages, we show that the model performs well with bilingual word vectors pre-trained on an appropriate translation corpus. We compare in-genre and in-domain parallel corpora, out-of-domain parallel corpora, in-domain comparable corpora, and monolingual corpora, and show that a relatively small, in-domain parallel corpus works best as a transfer medium if it is available. We describe the conditions under which other resources and embedding generation methods are successful, and these include our strategies for leveraging in-domain comparable corpora for cross-lingual sentiment analysis. To enhance the ability of the cross-lingual model to identify sentiment in the target language, we present new feature representations for sentiment analysis that are incorporated in the cross-lingual model: bilingual sentiment embeddings that are used to create bilingual sentiment scores, and a method for updating the sentiment embeddings during training by lexicalization of the target language. This feature configuration works best for the largest number of target languages in both untargeted and targeted cross-lingual sentiment experiments. The cross-lingual model is studied further by evaluating the role of the source language, which has traditionally been assumed to be English. We build cross-lingual models using 15 source languages, including two non-European and non-Indo-European source languages: Arabic and Chinese. We show that language families play an important role in the performance of the model, as does the morphological complexity of the source language. In the last part of the work, we focus on sentiment analysis towards targets. We study Arabic as a representative morphologically complex language and develop models and morphological representation features for identifying entity targets and sentiment expressed towards them in Arabic open-domain text. Finally, we adapt our cross-lingual sentiment models for the detection of sentiment towards targets. Through cross-lingual experiments on Arabic and English, we demonstrate that our findings regarding resources, features, and language also hold true for the transfer of targeted sentiment.

English-language news articles are no longer necessarily the best source of information. The Web allows information to spread more quickly and travel farther: first-person accounts of breaking news events pop up on social media, and foreign-language news articles are accessible to, if not immediately understandable by, English-speaking users. This thesis focuses on developing automatic summarization techniques for these new sources of information. We focus on summarizing two specific new sources of information: personal narratives, first-person accounts of exciting or unusual events that are readily found in blog entries and other social media posts, and non-English documents, which must first be translated into English, often introducing translation errors that complicate the summarization process. Personal narratives are a very new area of interest in natural language processing research, and they present two key challenges for summarization. First, unlike many news articles, whose lead sentences serve as summaries of the most important ideas in the articles, personal narratives provide no such shortcuts for determining where important information occurs in within them; second, personal narratives are written informally and colloquially, and unlike news articles, they are rarely edited, so they require heavier editing and rewriting during the summarization process. Non-English documents, whether news or narrative, present yet another source of difficulty on top of any challenges inherent to their genre: they must be translated into English, potentially introducing translation errors and disfluencies that must be identified and corrected during summarization. The bulk of this thesis is dedicated to addressing the challenges of summarizing personal narratives found on the Web. We develop a two-stage summarization system for personal narrative that first extracts sentences containing important content and then rewrites those sentences into summary-appropriate forms. Our content extraction system is inspired by contextualist narrative theory, using changes in writing style throughout a narrative to detect sentences containing important information; it outperforms both graph-based and neural network approaches to sentence extraction for this genre. Our paraphrasing system rewrites the extracted sentences into shorter, standalone summary sentences, learning to mimic the paraphrasing choices of human summarizers more closely than can traditional lexicon- or translation-based paraphrasing approaches. We conclude with a chapter dedicated to summarizing non-English documents written in low-resource languages – documents that would otherwise be unreadable for English-speaking users. We develop a cross-lingual summarization system that performs even heavier editing and rewriting than does our personal narrative paraphrasing system; we create and train on large amounts of synthetic errorful translations of foreign-language documents. Our approach produces fluent English summaries from disdisfluent translations of non-English documents, and it generalizes across languages.

This thesis principally addresses some problems in genetic programming (GP) and grammar-guided genetic programming (GGGP) arising from the lack of operators able to make small and bounded changes on both genotype and phenotype space. It proposes a new and flexible representation for genetic programming, using a state-of-the-art formalism from natural language processing, Tree Adjoining Grammars (TAGs). It demonstrates that the new TAG-based representation possesses two important properties: non-fixed arity and locality. The former facilitates the design of new operators, including some which are bio-inspired, and others able to make small and bounded changes. The latter ensures that bounded changes in genotype space are reflected in bounded changes in phenotype space. With these two properties, the thesis shows how some well-known difficulties in standard GP and GGGP tree-based representations can be solved in the new representation. These difficulties have been previously attributed to the treebased nature of the representations; since TAG representation is also tree-based, it has enabled a more precise delineation of the causes of the difficulties. Building on the new representation, a new grammar guided GP system known as TAG3P has been developed, and shown to be competitive with other GP and GGGP systems. A new schema theorem, explaining the behaviour of TAG3P on syntactically constrained domains, is derived. Finally, the thesis proposes a new method for understanding performance differences between GP representations requiring different ways to bound the search space, eliminating the effects of the bounds through multi-objective approaches.

Linguistic patterns have been used widely in shallow methods to develop numerous NLP applications. Approaches for acquiring linguistic patterns can be broadly categorised into three groups: supervised learning, unsupervised learning and manual methods. In supervised learning approaches, a large annotated training corpus is required for the learning algorithms to achieve decent results. However, annotated corpora are expensive to obtain and usually available only for established tasks. Unsupervised learning approaches usually start with a few seed examples and gather some statistics based on a large unannotated corpus to detect new examples that are similar to the seed ones. Most of these approaches either populate lexicons for predefined patterns or learn new patterns for extracting general factual information; hence they are applicable to only a limited number of tasks. Manually creating linguistic patterns has the advantage of utilising an expert's knowledge to overcome the scarcity of annotated data. In tasks with no annotated data available, the manual way seems to be the only choice. One typical problem that occurs with manual approaches is that the combination of multiple patterns, possibly being used at different stages of processing, often causes unintended side effects. Existing approaches, however, do not focus on the practical problem of acquiring those patterns but rather on how to use linguistic patterns for processing text. A systematic way to support the process of manually acquiring linguistic patterns in an efficient manner is long overdue. This thesis presents KAFTIE, an incremental knowledge acquisition framework that strongly supports experts in creating linguistic patterns manually for various NLP tasks. KAFTIE addresses difficulties in manually constructing knowledge bases of linguistic patterns, or rules in general, often faced in existing approaches by: (1) offering a systematic way to create new patterns while ensuring they are consistent; (2) alleviating the difficulty in choosing the right level of generality when creating a new pattern; (3) suggesting how existing patterns can be modified to improve the knowledge base's performance; (4) making the effort in creating a new pattern, or modifying an existing pattern, independent of the knowledge base's size. KAFTIE, therefore, makes it possible for experts to efficiently build large knowledge bases for complex tasks. This thesis also presents the KAFDIS framework for discourse processing using new representation formalisms: the level-of-detail tree and the discourse structure graph.

Skyline has been proposed as an important operator for many applications, such as multi-criteria decision making, data mining and visualization, and user-preference queries. Due to its importance, skyline and its computation have received considerable attention from database research community recently. All the existing techniques, however, focus on the conventional databases. They are not applicable to online computation environment, such as data stream. In addition, the existing studies consider efficiency of skyline computation only, while the fundamental problem on the semantics of skylines still remains open. In this thesis, we study three problems of skyline computation: (1) online computing skyline over data stream; (2) skyline cube computation and its analysis; and (3) top-k most representative skyline. To tackle the problem of online skyline computation, we develop a novel framework which converts more expensive multiple dimensional skyline computation to stabbing queries in 1-dimensional space. Based on this framework, a rigorous theoretical analysis of the time complexity of online skyline computation is provided. Then, efficient algorithms are proposed to support ad hoc and continuous skyline queries over data stream. Inspired by the idea of data cube, we propose a novel concept of skyline cube which consists of skylines of all possible non-empty subsets of a given full space. We identify the unique sharing strategies for skyline cube computation and develop two efficient algorithms which compute skyline cube in a bottom-up and top-down manner, respectively. Finally, a theoretical framework to answer the question about semantics of skyline and analysis of multidimensional subspace skyline are presented. Motived by the fact that the full skyline may be less informative because it generally consists of a large number of skyline points, we proposed a novel skyline operator -- top-k most representative skyline. The top-k most representative skyline operator selects the k skyline points so that the number of data points, which are dominated by at least one of these k skyline points, is maximized. To compute top-k most representative skyline, two efficient algorithms and their theoretical analysis are presented.

This thesis develops a new method for predicting probabilities in a statistical parser so that more sophisticated probabilistic grammars can be used. A statistical parser uses a probabilistic grammar derived from a training corpus of hand-parsed sentences. The grammar is represented as a set of constructions - in a simple case these might be context-free rules. The probability of each construction in the grammar is then estimated by counting its relative frequency in the corpus. A crucial problem when building a probabilistic grammar is to select an appropriate level of granularity for describing the constructions being learned. The more constructions we include in our grammar, the more sophisticated a model of the language we produce. However, if too many different constructions are included, then our corpus is unlikely to contain reliable information about the relative frequency of many constructions. In existing statistical parsers two main approaches have been taken to choosing an appropriate granularity. In a non-lexicalised parser constructions are specified as structures involving particular parts-of-speech, thereby abstracting over individual words. Thus, in the training corpus two syntactic structures involving the same parts-of-speech but different words would be treated as two instances of the same event. In a lexicalised grammar the assumption is that the individual words in a sentence carry information about its syntactic analysis over and above what is carried by its part-of-speech tags. Lexicalised grammars have the potential to provide extremely detailed syntactic analyses; however, Zipf�s law makes it hard for such grammars to be learned. In this thesis, we propose a method for optimising the trade-off between informative and learnable constructions in statistical parsing. We implement a grammar which works at a level of granularity in between single words and parts-of-speech, by grouping words together using unsupervised clustering based on bigram statistics. We begin by implementing a statistical parser to serve as the basis for our experiments. The parser, based on that of Michael Collins (1999), contains a number of new features of general interest. We then implement a model of word clustering, which we believe is the first to deliver vector-based word representations for an arbitrarily large lexicon. Finally, we describe a series of experiments in which the statistical parser is trained using categories based on these word representations.

Personal assistants need to allow the user to interact with the system in a flexible and adaptive way such as through spoken language dialogue. This research is aimed at achieving robust and effective dialogue management in such applications. We focus on an application, the Smart Personal Assistant (SPA), in which the user can use a variety of devices to interact with a collection of personal assistants, each specializing in a task domain. The current implementation of the SPA contains an e-mail management agent and a calendar agent that the user can interact with through a spoken dialogue and a graphical interface on PDAs. The user-system interaction is handled by a Dialogue Manager agent. We propose an agent-based approach that makes use of a BDI agent architecture for dialogue modelling and control. The Dialogue Manager agent of the SPA acts as the central point for maintaining coherent user-system interaction and coordinating the activities of the assistants. The dialogue model consists of a set of complex but modular plans for handling communicative goals. The dialogue control flow emerges automatically as the result of the agent???s plan selection by the BDI interpreter. In addition the Dialogue Manager maintains the conversational context, the domainspecific knowledge and the user model in its internal beliefs. We also consider the problem of dialogue adaptation in such agent-based dialogue systems. We present a novel way of integrating learning into a BDI architecture so that the agent can learn to select the most suitable plan among those applicable in the current context. This enables the Dialogue Manager agent to tailor its responses according to the conversational context and the user???s physical context, devices and preferences. Finally, we report the evaluation results, which indicate the robustness and effectiveness of the dialogue model in handling a range of users.

Thesis (Ph.D.)--University of Delaware, 2007. / Principal faculty advisors: Lori L. Pollock and Vijay K. Shanker, Dept. of Computer & Information Sciences. Includes bibliographical references.

Skyline has been proposed as an important operator for many applications, such as multi-criteria decision making, data mining and visualization, and user-preference queries. Due to its importance, skyline and its computation have received considerable attention from database research community recently. All the existing techniques, however, focus on the conventional databases. They are not applicable to online computation environment, such as data stream. In addition, the existing studies consider efficiency of skyline computation only, while the fundamental problem on the semantics of skylines still remains open. In this thesis, we study three problems of skyline computation: (1) online computing skyline over data stream; (2) skyline cube computation and its analysis; and (3) top-k most representative skyline. To tackle the problem of online skyline computation, we develop a novel framework which converts more expensive multiple dimensional skyline computation to stabbing queries in 1-dimensional space. Based on this framework, a rigorous theoretical analysis of the time complexity of online skyline computation is provided. Then, efficient algorithms are proposed to support ad hoc and continuous skyline queries over data stream. Inspired by the idea of data cube, we propose a novel concept of skyline cube which consists of skylines of all possible non-empty subsets of a given full space. We identify the unique sharing strategies for skyline cube computation and develop two efficient algorithms which compute skyline cube in a bottom-up and top-down manner, respectively. Finally, a theoretical framework to answer the question about semantics of skyline and analysis of multidimensional subspace skyline are presented. Motived by the fact that the full skyline may be less informative because it generally consists of a large number of skyline points, we proposed a novel skyline operator -- top-k most representative skyline. The top-k most representative skyline operator selects the k skyline points so that the number of data points, which are dominated by at least one of these k skyline points, is maximized. To compute top-k most representative skyline, two efficient algorithms and their theoretical analysis are presented.