In this article, we present and evaluate a method for training a statistical partof-speech tagger on data from written language and then adapting it to the requirements of tagging a corpus of transcribed spoken language, in our case spoken Swedish. This is currently a significant problem for many research groups working with spoken language, since the availability of tagged training data from spoken language is still very limited for most languages. The overall accuracy of the tagger developed for spoken Swedish is quite respectable, varying from 95% to 97 % depending on the tagset used. In conclusion, we argue that the method presented here gives good tagging accuracy with relatively little effort.

Most text mining tasks, including clustering and topic detection, are based on statistical methods that treat text as bags of words. Semantics in the text is largely ignored in the mining process, and mining results often have low interpretability. One particular challenge faced by such approaches lies in short text understanding, as short texts lack enough content from which statistical conclusions can be drawn easily. In this paper, we improve text understanding by using a probabilistic knowledgebase that is as rich as our mental world in terms of the concepts (of worldly facts) it contains. We then develop a Bayesian inference mechanism to conceptualize words and short text. We conducted comprehensive experiments on conceptualizing textual terms, and clustering short pieces of text such as Twitter messages. Compared to purely statistical methods such as latent semantic topic modeling or methods that use existing knowledgebases (e.g., WordNet, Freebase and Wikipedia), our approach brings significant improvements in short text understanding as reflected by the clustering accuracy.

In this thesis, we investigate using the Language Modeling approach for ad hoc Information Retrieval as a theoretically principled framework for encoding contextual evidence. Using context to improve retrieval performance is a current challenge within the discipline and presents a major challenge to the research community. The Language Modeling approach provides a natural and intuitive means of encoding the context as-sociated with a document. However, the Language Modeling approach also represents a change to the way probability theory is applied in ad hoc Information Retrieval and makes several assumptions for its application [112, 113, 57, 96]. We consider these assumptions and study them in detail during the course of this thesis. Central to the assumptions is the key implication that better retrieval performance can be obtained through developing better representation of the documents. We posit that the context associated with a document will enable the development of such representations-context based document models. This premise relies upon the explicit and implicit assumptions of the Language Modeling approach being valid, which have, up until now,

OF THE DISSERTATION Model Checking for Incomplete High Dimensional Categorical Data by Ming-Yi Hu Doctor of Philosophy in Statistics University of California, Los Angeles, 1999 Professor Thomas R. Belin, Co-chair Professor Robert I. Jennrich, Co-chair Categorical data are often arranged in a contingency table and summarized by a loglinear model. A standard approach for comparing two competing models is to calculate twice the discrepancy between maximized loglikelihoods, which follows a 2 distribution asymptotically. But when data are sparse, the 2 approximation may be questionable. xii As an alternative to a large-sample approximation to the reference distribution, we implement the framework introduced by Rubin (1984) for finding the posterior predictive check (PPC) distribution. The PPC distribution represents the conditional probability of a future value of a test statistic based on the information given by observed data along with model specifications, which can se...

Estimating the parameters of sparse multinomial distributions is an important component of many statistical learning tasks. Recent approaches have used uncertainty over the vocabulary of symbols in a multinomial distribution as a means of accounting for sparsity.

Motivation: The MHC superfamily (MhcSF) consists of immune system MHC class I (MHC-I) proteins, along with proteins with a MHC-I-like structure that are involved in a large variety of biological processes. Beta2-microglobulin (B2M) noncovalent binding to MHCI proteins is required for their surface expression and function, while MHC-I-like proteins interact, or not, with B2M. This study was designed to predict B2M binding (or non-binding) of newly identified MhcSF proteins, in order to decipher their function, understand the molecular recognition mechanisms, and identify deleterious mutations. IMGT standardization of MhcSF protein domains provides a unique numbering of the multiple alignment positions, and conditions to develop such predictive tool.

The rapid growth of communication technologies and the invention of set-top-box (STB) and personal digital recorder (PDR) have enabled today's television to receive and store tremendous programs. The abundance of TV programs precipitates a need for personalization tools to help people obtain programs that they really want to watch. User preference learning plays an important role in TV program personalization. In this paper, we introduce a hybrid user preference learning approach for TV program personalization. The learning architecture is designed to integrate multiple learning sources for preference learning, which are explicit input/modification, user viewing history, and user real-time feedback. Among those, learning from user viewing history and learning from user real-time feedback are described in detail. The experimental results proved that the hybrid learning approach outperforms the learning method merely adopting user real-time feedback.

Most text mining tasks, including clustering and topic detection, are based on statistical methods that treat text as bags of words. Semantics in the text is largely ignored in the mining process, and mining results often have low interpretability. One particular challenge faced by such approaches lies in short text understanding, as short texts lack enough content from which statistical conclusions can be drawn easily. In this paper, we improve text understanding by using a probabilistic knowledgebase that is as rich as our mental world in terms of the concepts (of worldly facts) it contains. We then develop a Bayesian inference mechanism to conceptualize words and short text. We conducted comprehensive experiments on conceptualizing textual terms, and clustering short pieces of text such as Twitter messages. Compared to purely statistical methods such as latent semantic topic modeling or methods that use existing knowledgebases (e.g., WordNet, Freebase and Wikipedia), our approach brings significant improvements in short text understanding as reflected by the clustering accuracy. 1

We present a computational framework for stochastically modeling dyad interaction chronograms. The framework’s most novel feature is the capacity for incremental learning and forgetting. To showcase its flexibility, we design experiments answering four concrete questions about the systematics of spoken interaction. The results show that: (1) individuals are clearly affected by one another; (2) there is individual variation in interaction strategy; (3) strategies wander in time rather than converge; and (4) individuals exhibit similarity with their interlocutors. We expect the proposed framework to be capable of answering many such questions with little additional effort. Index Terms: interaction, chronogram modeling, turn-taking, incremental learning.

In this paper we present experiments using Random Indexing for “query expansion ” in Word Sense Disambiguation. Random Indexing is an efficient, scalable and incremental latent semantic indexing method somewhat akin to LSA, and has in these experiments shown promising results on a small test set for Swedish with an accuracy up to 80 % with relatively little training data. We also compare it to results obtained when applying a Naïve Bayes classifier to the same training and data sets, retrieving a maximum accuracy of 56%. 1