Previous work in social network analysis (SNA) has modeled the existence of links from one entity to another, but not the language content or topics on those links. We present the Author-Recipient-Topic (ART) model for social network analysis, which learns topic distributions based on the direction-sensitive messages sent between entities. The model builds on Latent Dirichlet Allocation (LDA) and the Author-Topic (AT) model, adding the key attribute that distribution over topics is conditioned distinctly on both the sender and recipient—steering the discovery of topics according to the relationships between people. We give results on both the Enron email corpus and a researcher’s email archive, providing evidence not only that clearly relevant topics are discovered, but that the ART model better predicts people’s roles. 1 Introduction and Related Work Social network analysis (SNA) is the study of mathematical models for interactions among people, organizations and groups. With the recent availability of large datasets of human

This paper presents preliminary results on the detection of cultural differences from people’s experiences in various countries from two perspectives: tourists and locals. Our approach is to develop probabilistic models that would provide a good framework for such studies. Thus, we propose here a new model, ccLDA, which extends over the Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and crosscollection mixture (ccMix) (Zhai et al., 2004) models on blogs and forums. We also provide a qualitative and quantitative analysis of the model on the cross-cultural data. 1

We present a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing which handles fractional counts. We address the scalability issue to large training corpora via bootstrapping of bigram LSA from unigram LSA. For LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM via marginal adaptation and linear interpolation respectively. Experimental results on the Mandarin RT04 test set show that applying unigram and bigram LSA together yields 6%–8 % relative perplexity reduction and 2.5 % relative character error rate reduction which is statistically significant compared to applying only unigram LSA. On the large-scale evaluation on Arabic, word error rate reduction from bigram LSA is statistically significant compared to the unadapted baseline. 1

Permission is herewith granted to Dalhousie University to circulate and to have copied for non-commercial purposes, at its discretion, the above title upon the request of individuals or institutions. Signature of Author The author reserves other publication rights, and neither the thesis nor extensive extracts from it may be printed or otherwise reproduced without the author’s written permission. The author attests that permission has been obtained for the use of any copyrighted material appearing in the thesis (other than brief excerpts requiring only proper acknowledgement in scholarly writing) and that all such use is clearly acknowledged. iii I dedicate this to my family, Zahra and Taha for their love, help, and patience

as to style and content by:

Previous research has shown that topics can be automatically discovered in a software project’s source code. Topics are collections of words that co-occur frequently in a text collection and are discovered using topic models such as latent Dirichlet allocation (LDA). Tracking how topics evolve, i.e., grow and spread, over time is useful for supporting software maintenance, comprehension, and re-engineering activities. The evolution of topics is typically recovered by applying LDA to all versions of a project’s source code at once, followed by post processing to map topics across versions. Although this technique works well in applications where each version of the data is completely different, for example in the analysis of conference proceedings, the technique does not work well with source code, which typically changes only incrementally and contains significant duplication across versions. In this paper, we present a new approach, called DiffLDA, for automatically mining topic evolution in source code. The approach addresses LDA’s sensitivity to document duplication by operating on the differences between versions of a source code document, resulting in a more accurate, finer-grained representation of topic evolution. We validate our approach through case studies on simulated data and two open source projects. 1

Given a movie comment, does it contain a spoiler? A spoiler is a comment that, when disclosed, would ruin a surprise or reveal an important plot detail. We study automatic methods to detect comments and reviews that contain spoilers and apply them to reviews from the IMDB (Internet Movie Database) website. We develop topic models, based on Latent Dirichlet Allocation (LDA), but using linguistic dependency information in place of simple features from bag of words (BOW) representations. Experimental results demonstrate the effectiveness of our technique over four movie-comment datasets of different scales. 1

Most text mining methods are based on representing documents using a vector space model, commonly known as a bag of word model, where each document is modeled as a linear vector representing the occurrence of independent words in the text corpus. It is well known that using this vector-based representation, important information, such as semantic relationship among concepts, is lost. This paper proposes a novel text representation model called ConceptLink graph. The ConceptLink graph does not only represent the content of the document, but also captures some of its underlying semantic structure in terms of the relationships among concepts. The ConceptLink graph is constructed in two main stages. First, we find a set of concepts by clustering conceptually related terms using the self-organizing map method. Secondly, by mapping each document’s content to concept, we generate a graph of concepts based on the occurrences of concepts using a singular value decomposition technique. The ConceptLink graph will overcome the keyword independence limitation in the vector space model to take advantage of the implicit concept relationships exhibit in all natural language texts. As an information-rich text representation model, the ConceptLink graph will advance text mining technology beyond feature-based to structure-based knowledge discovery. We will illustrate the ConceptLink graph method using samples generated from benchmark text mining dataset.

The growing number of statistical topic models led to the need to better evaluate their output. Traditional evaluation means estimate the model’s fitness to unseen data. It has recently been proven than the output of human judgment can greatly differ from these measures. Thus the need for methods that better emulate human judgment is stringent. In this paper we present a system that computes the conceptual relevance of individual topics from a given model on the basis of information drawn from a given concept hierarchy, in this case WordNet. The notion of conceptual relevance is regarded as the ability to attribute a concept to each topic and separate words related to the topic from the unrelated ones based on that concept. In multiple experiments we prove the correlation between the automatic evaluation method and the answers received from human evaluators, for various corpora and difficulty levels. By changing the evaluation focus from a statistical one to a conceptual one we were able to detect which topics are conceptually meaningful and rank them accordingly. 1

ii CONTENTS