We explore a machine learning task, evidence recommendation (ER), the extraction of evidence from a source document to support an external claim. This task is an instance of the question answering machine learning task. We apply ER to academic publications because they cite other papers for the claims they make. Reading cited papers to corroborate claims is time-consuming and an automated ER tool could expedite it. Thus, we propose a methodology for collecting a dataset of academic papers and their references. We explore deep learning models for ER and achieve 77% accuracy with pairwise models and 75% pairwise accuracy with document-wise models.

One of the main research challenges nowadays in the context of Massive Open Online Courses (MOOCs) is the automation of the evaluation process of text-based assessments effectively. Text-based assessments, such as essay writing, have been proved to be better indicators of higher level of understanding than machine-scored assessments (E.g. Multiple Choice Questions). Nonetheless, due to the rapid growth of MOOCs, text-based evaluation has become a difficult task for human markers, creating the need of automated systems for grading. In this thesis, we focus on the automated short answer grading task (ASAG), which automatically assesses natural language answers to open-ended questions into correct and incorrect classes. We propose an ensemble supervised machine learning approach that relies on two types of classifiers: a response-based classifier, which centers around feature extraction from available responses, and a reference-based classifier which considers the relationships between responses, model answers and questions. For each classifier, we explored a set of features based on words and entities. For the response-based classifier, we tested and compared 5 features: traditional n-gram models, entity URIs (Uniform Resource Identifier) and entity mentions both extracted using a semantic annotation API, entity mention embeddings based on GloVe and entity URI embeddings extracted from Wikipedia. For the reference-based classifier, we explored fourteen features: cosine similarity between sentence embeddings from student answers and model answers, number of overlapping elements (words, entity URI, entity mention) between student answers and model answers or question text, Jaccard similarity coefficient between student answers and model answers or question text (based on words, entity URI or entity mentions) and a sentence embedding representation. We evaluated our classifiers on three datasets, two of which belong to the SemEval ASAG competition (Dzikovska et al., 2013). Our results show that, in general, reference-based features perform much better than response-based features in terms of accuracy and macro average f1-score. Within the reference-based approach, we observe that the use of S6 embedding representation, which considers question text, student and model answer, generated the best performing models. Nonetheless, their combination with other similarity features helped build more accurate classifiers. As for response-based classifiers, models based on traditional n-gram features remained the best models. Finally, we combined our best reference-based and response-based classifiers using an ensemble learning model. Our ensemble classifiers combining both approaches achieved the best results for one of the evaluation datasets, but underperformed on the remaining two. We also compared the best two classifiers with some of the main state-of-the-art results on the SemEval competition. Our final embedded meta-classifier outperformed the top-ranking result on the SemEval Beetle dataset and our top classifier on SemEval SciEntBank, trained on reference-based features, obtained the 2nd position. In conclusion, the reference-based approach, powered mainly by sentence level embeddings and other similarity features, proved to generate the most efficient models in two out of three datasets and the ensemble model was the best on the SemEval Beetle dataset.

by Ho, Kei Shiu Edward. / "June 1996." / Thesis (Ph.D.)--Chinese University of Hong Kong, 1996. / Includes bibliographical references (p. 186-193). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Mode of access: World Wide Web.

Many state-of-the-art results in natural language processing (NLP) are achieved with statistical models involving latent variables. Unfortunately, computational problems associated with such models (for instance, finding the optimal parameter values) are typically intractable, forcing practitioners to rely on heuristic methods without strong guarantees. While heuristics are often sufficient for empirical purposes, their de-emphasis on theoretical aspects has certain negative ramifications. First, it can impede the development of rigorous theoretical understanding which can generate new ideas and algorithms. Second, it can lead to black art solutions that are unreliable and difficult to reproduce. In this thesis, we argue that spectral methods---that is, methods that use singular value decomposition or other similar matrix or tensor factorization---can effectively remedy these negative ramifications. To this end, we develop spectral methods for two unsupervised language processing tasks. The first task is learning lexical representations from unannotated text (e.g., hierarchical clustering of a vocabulary). The second task is estimating parameters of latent-variable models used in NLP applications (e.g., for unsupervised part-of-speech tagging). We show that our spectral algorithms have the following advantages over previous methods: 1. The algorithms provide a new theoretical framework that is amenable to rigorous analysis. In particular, they are shown to be statistically consistent. 2. The algorithms are simple to implement, efficient, and scalable to large amounts of data. They also yield results that are competitive with the state-of-the-art.

Concept-to-text generation suffers from what can be called generation bottlenecks - aspects of the generated text which should change for different subject domains, and which are usually hard to obtain or require manual work. Some examples are domain-specific content, a type system, a dictionary, discourse style and lexical style. These bottlenecks have stifled attempts to create generation systems that are generic, or at least apply to a wide range of domains in non-trivial applications. This thesis is comprised of two parts. In the first, we propose data-driven solutions that automate obtaining the information and models required to solve some of these bottlenecks. Specifically, we present an approach to mining domain-specific paraphrasal templates from a simple text corpus; an approach to extracting a domain-specific taxonomic thesaurus from Wikipedia; and a novel document planning model which determines both ordering and discourse relations, and which can be extracted from a domain corpus. We evaluate each solution individually and independently from its ultimate use in generation, and show significant improvements in each. In the second part of the thesis, we describe a framework for creating generation systems that rely on these solutions, as well as on hybrid concept-to-text and text-to-text generation, and which can be automatically adapted to any domain using only a domain-specific corpus. We illustrate the breadth of applications that this framework applies to with three examples: biography generation and company description generation, which we use to evaluate the framework itself and the contribution of our solutions; and justification of machine learning predictions, a novel application which we evaluate in a task-based study to show its importance to users.

Automatic reading comprehension (RC) systems integrate various kinds of natural language processing (NLP) technologies to analyze a given passage and generate or extract answers in response to questions about the passage. Previous work applied a lot of NLP technologies including shallow syntactic analyses (e.g. base noun phrases), semantic analyses (e.g. named entities) and discourse analyses (e.g. pronoun referents) in the bag-of-words (BOW) matching approach. This thesis proposes a novel RC approach that integrates a set of NLP technologies in a maximum entropy (ME) framework to estimate candidate answer sentences' probabilities being answers. In contrast to previous RC approaches, which are in English-only, the presented RC approach is the first one for both English and Chinese, the two languages used by most people in the world. In order to support the evaluation of the bilingual RC systems, a parallel English and Chinese corpus is also designed and developed. Annotations deemed relevant to the RC task are also included in the corpus. In addition, useful NLP technologies are explored from a new perspective---referring the pedagogical guidelines of humans, reading skills are summarized and mapped to various NLP technologies. Practical NLP technologies, categorized as shallow syntactic analyses (i.e. part-of-speech tags, voices and tenses) and deep syntactic analyses (i.e. syntactic parse trees and dependency parse trees) are then selected for integration. The proposed approach is evaluated on an English corpus, namely Remedia and our bilingual corpus. The experimental results show that our approach significantly improves the RC results on both English and Chinese corpora. / Xu, Kui. / Adviser: Helen Mei-Ling Meng. / Source: Dissertation Abstracts International, Volume: 70-06, Section: B, page: 3618. / Thesis (Ph.D.)--Chinese University of Hong Kong, 2008. / Includes bibliographical references (leaves 132-141). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Electronic reproduction. [Ann Arbor, MI] : ProQuest Information and Learning, [200-] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstracts in English and Chinese. / School code: 1307.

Unsupervised approaches to learning the morphology of a language play an important role in computer processing of language from a practical and theoretical perspective, due their minimal reliance on manually produced linguistic resources and human annotation. Such approaches have been widely researched for the problem of concatenative affixation, but less attention has been paid to the intercalated (non-concatenative) morphology exhibited by Arabic and other Semitic languages. The aim of this research is to learn the root and pattern morphology of Arabic, with accuracy comparable to manually built morphological analysis systems. The approach is kept free from human supervision or manual parameter settings, assuming only that roots and patterns intertwine to form a word. Promising results were obtained by applying a technique adapted from previous work in concatenative morphology learning, which uses machine learning to determine relatedness between words. The output, with probabilistic relatedness values between words, was then used to rank all possible roots and patterns to form a lexicon. Analysis using trilateral roots resulted in correct root identification accuracy of approximately 86% for inflected words. Although the machine learning-based approach is effective, it is conceptually complex. So an alternative, simpler and computationally efficient approach was then devised to obtain morpheme scores based on comparative counts of roots and patterns. In this approach, root and pattern scores are defined in terms of each other in a mutually recursive relationship, converging to an optimized morpheme ranking. This technique gives slightly better accuracy while being conceptually simpler and more efficient. The approach, after further enhancements, was evaluated on a version of the Quranic Arabic Corpus, attaining a final accuracy of approximately 93%. A comparative evaluation shows this to be superior to two existing, well used manually built Arabic stemmers, thus demonstrating the practical feasibility of unsupervised learning of non-concatenative morphology.

Pang Chun Kiu. / Thesis (M.Phil.)--Chinese University of Hong Kong, 1997. / Includes bibliographical references (leaves 101-107). / Chapter 1 --- Introduction --- p.2 / Chapter 1.1 --- Chinese Indexing using Noun Phrases --- p.6 / Chapter 1.2 --- Objectives --- p.8 / Chapter 1.3 --- An Overview of the Thesis --- p.8 / Chapter 2 --- Background --- p.10 / Chapter 2.1 --- Technology Influences on Information Retrieval --- p.10 / Chapter 2.2 --- Related Work --- p.13 / Chapter 2.2.1 --- Statistical/Keyword Approaches --- p.13 / Chapter 2.2.2 --- Syntactical approaches --- p.15 / Chapter 2.2.3 --- Semantic approaches --- p.17 / Chapter 2.2.4 --- Noun Phrases Approach --- p.18 / Chapter 2.2.5 --- Chinese Information Retrieval --- p.20 / Chapter 2.3 --- Our Approach --- p.21 / Chapter 3 --- Chinese Noun Phrases --- p.23 / Chapter 3.1 --- Different types of Chinese Noun Phrases --- p.23 / Chapter 3.2 --- Ambiguous noun phrases --- p.27 / Chapter 3.2.1 --- Ambiguous English Noun Phrases --- p.27 / Chapter 3.2.2 --- Ambiguous Chinese Noun Phrases --- p.28 / Chapter 3.2.3 --- Statistical data on the three NPs --- p.33 / Chapter 4 --- Index Extraction from De-de Conj. NP --- p.35 / Chapter 4.1 --- Word Segmentation --- p.36 / Chapter 4.2 --- Part-of-speech tagging --- p.37 / Chapter 4.3 --- Noun Phrase Extraction --- p.37 / Chapter 4.4 --- The Chinese noun phrase partial parser --- p.38 / Chapter 4.5 --- Handling Parsing Ambiguity --- p.40 / Chapter 4.6 --- Index Building Strategy --- p.41 / Chapter 4.7 --- The cross-set generation rules --- p.44 / Chapter 4.8 --- Example 1: Indexing De-de NP --- p.46 / Chapter 4.9 --- Example 2: Indexing Conjunctive NP --- p.48 / Chapter 4.10 --- Experimental results and Discussion --- p.49 / Chapter 5 --- Indexing Compound Nouns --- p.52 / Chapter 5.1 --- Previous Researches on Compound Nouns --- p.53 / Chapter 5.2 --- Indexing two-term Compound Nouns --- p.55 / Chapter 5.2.1 --- About the thesaurus《同義詞詞林》 --- p.56 / Chapter 5.3 --- Indexing Compound Nouns of three or more terms --- p.58 / Chapter 5.4 --- Corpus learning approach --- p.59 / Chapter 5.4.1 --- An Example --- p.60 / Chapter 5.4.2 --- Experimental Setup --- p.63 / Chapter 5.4.3 --- An Experiment using the third level of the Cilin --- p.65 / Chapter 5.4.4 --- An Experiment using the second level of the Cilin --- p.66 / Chapter 5.5 --- Contextual Approach --- p.68 / Chapter 5.5.1 --- The algorithm --- p.69 / Chapter 5.5.2 --- An Illustrative Example --- p.71 / Chapter 5.5.3 --- Experiments on compound nouns --- p.72 / Chapter 5.5.4 --- Experiment I: Word Distance Based Extraction --- p.73 / Chapter 5.5.5 --- Experiment II: Semantic Class Based Extraction --- p.75 / Chapter 5.5.6 --- Experiments III: On different boundaries --- p.76 / Chapter 5.5.7 --- The Final Algorithm --- p.79 / Chapter 5.5.8 --- Experiments on other compounds --- p.82 / Chapter 5.5.9 --- Discussion --- p.83 / Chapter 6 --- Overall Effectiveness --- p.85 / Chapter 6.1 --- Illustrative Example for the Integrated Algorithm --- p.86 / Chapter 6.2 --- Experimental Setup --- p.90 / Chapter 6.3 --- Experimental Results & Discussion --- p.91 / Chapter 7 --- Conclusion --- p.95 / Chapter 7.1 --- Summary --- p.95 / Chapter 7.2 --- Contributions --- p.97 / Chapter 7.3 --- Future Directions --- p.98 / Chapter 7.3.1 --- Word-sense determination --- p.98 / Chapter 7.3.2 --- Hybrid approach for compound noun indexing --- p.99 / Chapter A --- Cross-set Generation Rules --- p.108 / Chapter B --- Tag set by Tsinghua University --- p.110 / Chapter C --- Noun Phrases Test Set --- p.113 / Chapter D --- Compound Nouns Test Set --- p.124 / Chapter D.l --- Three-term Compound Nouns --- p.125 / Chapter D.1.1 --- NVN --- p.125 / Chapter D.1.2 --- Other three-term compound nouns --- p.129 / Chapter D.2 --- Four-term Compound Nouns --- p.133 / Chapter D.3 --- Five-term and six-term Compound Nouns --- p.134

Wang Ying. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2003. / Includes bibliographical references (leaves 102-104). / Abstracts in English and Chinese. / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Overview --- p.1 / Chapter 1.2 --- Thesis Goals --- p.3 / Chapter 1.3 --- Thesis Outline --- p.3 / Chapter 2 --- Background --- p.4 / Chapter 2.1 --- Speech Recognition --- p.4 / Chapter 2.2 --- Natural Language Understanding --- p.6 / Chapter 2.2.1 --- Rule-based Approach --- p.7 / Chapter 2.2.2 --- Corpus-based Approach --- p.7 / Chapter 2.3 --- Integration of Speech Recognition with NLU --- p.8 / Chapter 2.3.1 --- Word Graph --- p.9 / Chapter 2.3.2 --- N-best List --- p.9 / Chapter 2.4 --- The ATIS Domain --- p.10 / Chapter 2.5 --- Chapter Summary --- p.14 / Chapter 3 --- Generation of Speech Recognition Hypotheses --- p.15 / Chapter 3.1 --- Grammar Development for the OpenSpeech Recognizer --- p.16 / Chapter 3.2 --- Generation of Speech Recognition Hypotheses --- p.22 / Chapter 3.3 --- Evaluation of Speech Recognition Hypotheses --- p.24 / Chapter 3.3.1 --- Recognition Accuracy --- p.24 / Chapter 3.3.2 --- Concept Accuracy --- p.28 / Chapter 3.4 --- Results and Analysis --- p.33 / Chapter 3.5 --- Chapter Summary --- p.38 / Chapter 4 --- Belief Networks for NLU --- p.40 / Chapter 4.1 --- Problem Formulation --- p.40 / Chapter 4.2 --- The Original NLU Framework --- p.41 / Chapter 4.2.1 --- Semantic Tagging --- p.41 / Chapter 4.2.2 --- Concept Selection --- p.42 / Chapter 4.2.3 --- Bayesian Inference --- p.43 / Chapter 4.2.4 --- Thresholding --- p.44 / Chapter 4.2.5 --- Goal Identification --- p.45 / Chapter 4.3 --- Evaluation Method of Goal Identification Performance --- p.45 / Chapter 4.4 --- Baseline Result --- p.48 / Chapter 4.5 --- Chapter Summary --- p.50 / Chapter 5 --- The Effects of Recognition Errors on NLU --- p.51 / Chapter 5.1 --- Experiments --- p.51 / Chapter 5.1.1 --- Perfect Case´ؤThe Use of Transcripts --- p.53 / Chapter 5.1.2 --- Train on Recognition Hypotheses --- p.53 / Chapter 5.1.3 --- Test on Recognition Hypotheses --- p.55 / Chapter 5.1.4 --- Train and Test on Recognition Hypotheses --- p.56 / Chapter 5.2 --- Analysis of Results --- p.60 / Chapter 5.3 --- Chapter Summary --- p.67 / Chapter 6 --- The Use of Multiple Speech Recognition Hypotheses for NLU --- p.69 / Chapter 6.1 --- The Extended NLU Framework --- p.76 / Chapter 6.1.1 --- Semantic Tagging --- p.76 / Chapter 6.1.2 --- Recognition Confidence Score Normalization --- p.77 / Chapter 6.1.3 --- Concept Selection --- p.79 / Chapter 6.1.4 --- Bayesian Inference --- p.80 / Chapter 6.1.5 --- Combination with Confidence Scores --- p.81 / Chapter 6.1.6 --- Thresholding --- p.84 / Chapter 6.1.7 --- Goal Identification --- p.84 / Chapter 6.2 --- Experiments --- p.86 / Chapter 6.2.1 --- The Use of First Best Recognition Hypothesis --- p.86 / Chapter 6.2.2 --- Train on Multiple Recognition Hypotheses --- p.86 / Chapter 6.2.3 --- Test on Multiple Recognition Hypotheses --- p.87 / Chapter 6.2.4 --- Train and Test on Multiple Recognition Hypotheses --- p.88 / Chapter 6.3 --- Significance Testing --- p.90 / Chapter 6.4 --- Result Analysis --- p.91 / Chapter 6.5 --- Chapter Summary --- p.97 / Chapter 7 --- Conclusions and Future Work --- p.98 / Chapter 7.1 --- Conclusions --- p.98 / Chapter 7.2 --- Contribution --- p.99 / Chapter 7.3 --- Future Work --- p.100 / Bibliography --- p.102 / Chapter A --- Speech Recognition Hypotheses Distribution --- p.105 / Chapter B --- Recognition Errors in Three Kinds of Queries --- p.107 / Chapter C --- The Effects of Recognition Errors in N-Best list on NLU --- p.114 / Chapter D --- Training on Multiple Recognition Hypotheses --- p.117 / Chapter E --- Testing on Multiple Recognition Hypotheses --- p.132 / Chapter F --- Hand-designed Grammar For ATIS --- p.139

Press releases serve as a major influence on public opinion of a politician, since they are a primary means of communicating with the public and directing discussion. Thus, the public’s ability to digest them is an important factor for politicians to consider. This study employs several well-studied measures of linguistic complexity and proposes a new one to examine whether politicians change their language to become more or less difficult to parse in different situations. This study uses 27,500 press releases from the US Senate between 2004–2008 and examines election cycles and natural disasters, namely hurricanes, as situations where politicians’ language may change. We calculate the syntactic complexity measures clauses per sentence, T-unit length, and complex-T ratio, as well as the Automated Readability Index and Flesch Reading Ease of each press release. We also propose a proof-of-concept measure called logical complexity to find if classical Boolean logic can be applied as a practical linguistic complexity measure. We find that language becomes more complex in coastal senators’ press releases concerning hurricanes, but see no significant change for those in election cycles. Our measure shows similar results to the well-established ones, showing that logical complexity is a useful lens for measuring linguistic complexity.