The labeling of language resources is a time consuming task, whether aided by machine learning or not. Much of the prior work in this area has focused on accelerating human annotation in the context of machine learning, yielding a variety of active learning approaches. Most of these attempt to lead an annotator to label the items which are most likely to improve the quality of an automated, machine learning-based model. These active learning approaches seek to understand the effect of item selection on the machine learning model, but give significantly less emphasis to the effect of item selection on the human annotator. In this work, we consider a sentiment labeling task where existing, traditional active learning seems to have little or no value. We focus instead on the human annotator by ordering the items for better annotator efficiency.

Statistical machine translation (SMT) requires large quantities of bitexts (i.e., bilingual parallel corpora) as training data to yield good quality translations. While obtaining a large amount of training data is critical, the similarity between training and test data also has a significant impact on SMT performance. Many SMT studies define data similarity in terms of domain-overlap, and domains are defined to be synonymous with data sources. Consequently, the SMT community has focused on domain adaptation techniques that augment small (in-domain) datasets with large datasets from other sources (hence, out-of-domain, per the definition). However, many training datasets consist of topically diverse data, and not all data contained in a single dataset are useful for translations of a specific target task. In this study, we propose a new perspective on data quality and topical similarity to enhance SMT performance. Using our data adaptation approach called topic adaptation, we select topically suitable training data corresponding to test data in order to produce better translations. We propose three topic adaptation approaches for the SMT process and investigate the effectiveness in both idealized and realistic settings using large parallel corpora. We measure performance of SMT systems trained on topically similar data and their effectiveness based on BLEU, the widely-used objective SMT performance metric. We show that topic adaptation approaches outperform baseline systems (0.3 – 3 BLEU points) when data selection parameters are carefully determined.

This study describes a series of case studies on topic management patterns of five children (ages 5 to 10 years) with language impairment. The children participated in semi-structured topic tasks that assessed conversational abilities on topics that were verbally introduced and topics that were introduced both verbally and with an object. Although there was considerable variability among participants, the children generally responded to most introductions by acknowledging and maintaining the topic. With the exception of one child, however, the children in this study demonstrated immature topic manipulation patterns that could be expected to have negative social ramifications.

The availability of large-scale data on the Web motivates the development of automatic algorithms to analyze topics and to identify relationships between topics. Various approaches have been proposed in the literature. Most focus on specific topics, mainly those representing people, with little attention to topics of other kinds. They are also less flexible in how they represent topics. In this thesis we study existing methods as well as describe a different approach, based on profiles, for representing topics. A Topic Profile is analogous to a synopsis of a topic and consists of different types of features. Profiles are flexible to allow different combinations of features to be emphasized and are extensible to support new features to be incorporated without having to change the underlying logic. More generally, topic profiles provide an abstract framework that can be used to create different types of concrete representations for topics. Different options regarding the number of documents considered for a topic or types of features extracted can be decided based on requirements of the problem as well as the characteristics of the data. Topic profiles also provide a framework to explore relationships between topics. We compare different methods for building profiles and evaluate them in terms of their information content and their ability to predict relationships between topics. We contribute new methods in term weighting and for identifying relevant text segments in web documents. In this thesis, we present an application of our profile-based approach to explore social networks of US senators generated from web data and compare with networks generated from voting data. We consider both general networks as well as issue-specific networks. We also apply topic profiles for identifying and ranking experts given topics of interest, as part of the 2007 TREC Expert Search task. Overall, our results show that topic profiles provide a strong foundation for exploring different topics and for mining relationships between topics using web data. Our approach can be applied to a wide range of web knowledge discovery problems, in contrast to existing approaches that are mostly designed for specific problems.

No description available.

No description available.

Thesis (Ph. D.)--University of Washington, 2006. / Vita. Includes bibliographical references (p. 148-153).

This article takes stock of the basic notions of Information Structure (IS). It first provides a general characterization of IS — following Chafe (1976) — within a communicative model of Common Ground(CG), which distinguishes between CG content and CG management. IS is concerned with those features of language that concern the local CG. Second, this paper defines and discusses the notions of Focus (as indicating alternatives) and its various uses, Givenness (as indicating that a denotation is already present in the CG), and Topic (as specifying what a statement is about). It also proposes a new notion, Delimitation, which comprises contrastive topics and frame setters, and indicates that the current conversational move does not entirely satisfy the local communicative needs. It also points out that rhetorical structuring partly belongs to IS.

While the Information Structure (IS) is most naturally interpreted as 'structure of information', some may argue that it is structure of something else, and others may object to the use of the word 'structure'. This paper focuses on the question of whether the informational component can have structural properties such that it can be called 'structure'. The preliminary conclusion is that, although there are some vague indications of structurehood in it, it is perhaps better understood to be a representation that encodes a finite set of information-based partitions, rather than structure.

The rapid growth in the number of documents available to various end users from around the world has led to a greatly increased need for machine understanding of their topics, as well as for automatic grouping of related documents. This constitutes one of the main current challenges in text mining. We introduce in this thesis a novel approach for identifying document topics. In this approach, we try to utilize human background knowledge to help us to automatically find the best matching topic for input documents. There are several applications for this task. For example, it can be used to improve the relevancy of search engine results by categorizing the search results according to their general topic. It can also give users the ability to choose the domain which is most relevant to their needs. It can also be used for an application like a news publisher, where we want to automatically assign each news article to one of the predefined news main topics. In order to achieve this, we need to extract background knowledge in a form appropriate to this task. The thesis contributions can be summarized into two main modules. In the first module, we introduce a new approach to extract background knowledge from a human knowledge source, in the form of a knowledge repository, and store it in a well-structured and organized form, namely an ontology. We define the methodology of identifying ontological concepts, as well as defining the relations between these concepts. We use the ontology to infer the semantic similarity between documents, as well as to identify their topics. We apply our proposed approach using perhaps the best-known of the knowledge repositories, namely Wikipedia. The second module of this dissertation defines the framework for automatic document topic identification (ADTI). We present a new approach that utilizes the knowledge stored in the created ontology to automatically find the best matching topics for input documents, without the need for a training process such as in document classification. We compare ADTI to other text mining tasks by conducting several experiments to compare the performance of ADTI and its competitors, namely document clustering and document classification. Results show that our document topic identification approach outperforms several document clustering techniques. They show also that while ADTI does not require training, it nevertheless shows competitive performance with one of the state-of-the-art methods for document classification.