Auswahl der wissenschaftlichen Literatur zum Thema „Data Domains“

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2009.<br>This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.<br>Cataloged from student submitted PDF version of thesis.<br>Includes bibliographical references (p. 99-109).<br>With the increasing focus of speech recognition and natural language processing applications on domains with limited amount of in-domain training data, enhanced system performance often relies on approaches involving model adaptation and combination. In such domains, language models are often constructed by interpolating component models trained from partially matched corpora. Instead of simple linear interpolation, we introduce a generalized linear interpolation technique that computes context-dependent mixture weights from features that correlate with the component confidence and relevance for each n-gram context. Since the n-grams from partially matched corpora may not be of equal relevance to the target domain, we propose an n-gram weighting scheme to adjust the component n-gram probabilities based on features derived from readily available corpus segmentation and metadata to de-emphasize out-of-domain n-grams. In scenarios without any matched data for a development set, we examine unsupervised and active learning techniques for tuning the interpolation and weighting parameters. Results on a lecture transcription task using the proposed generalized linear interpolation and n-gram weighting techniques yield up to a 1.4% absolute word error rate reduction over a linearly interpolated baseline language model. As more sophisticated models are only as useful as they are practical, we developed the MIT Language Modeling (MITLM) toolkit, designed for efficient iterative parameter optimization, and released it to the research community.<br>(cont.) With a compact vector-based n-gram data structure and optimized algorithm implementations, the toolkit not only improves the running time of common tasks by up to 40x, but also enables the efficient parameter tuning for language modeling techniques that were previously deemed impractical.<br>by Bo-June (Paul) Hsu.<br>Ph.D.

The emergence of advanced communication techniques has increased availability of large collection of data in electronic form in a number of application domains including healthcare, e- business, and e-learning. Everyday a large amount of records are stored electronically. However, finding useful information from such a large data collection is a challenging issue. Data mining technology aims automatically extracting hidden knowledge from large data repositories exploiting sophisticated algorithms. The hidden knowledge in the electronic data may be potentially utilized to facilitate the procedures, productivity, and reliability of several application domains. The PhD activity has been focused on novel and effective data mining approaches to tackle the complex data coming from two main application domains: Healthcare data analysis and Textual data analysis. The research activity, in the context of healthcare data, addressed the application of different data mining techniques to discover valuable knowledge from real exam-log data of patients. In particular, efforts have been devoted to the extraction of medical pathways, which can be exploited to analyze the actual treatments followed by patients. The derived knowledge not only provides useful information to deal with the treatment procedures but may also play an important role in future predictions of potential patient risks associated with medical treatments. The research effort in textual data analysis is twofold. On the one hand, a novel approach to discovery of succinct summaries of large document collections has been proposed. On the other hand, the suitability of an established descriptive data mining to support domain experts in making decisions has been investigated. Both research activities are focused on adopting widely exploratory data mining techniques to textual data analysis, which require overcoming intrinsic limitations for traditional algorithms for handling textual documents efficiently and effectively.

Temporal logics are a very popular family of logical languages, used to specify properties of abstracted systems. In the last few years, many extensions of temporal logics have been proposed, in order to address the need to express more than just abstract properties. In our work we study temporal logics extended by local constraints, which allow to express quantitative properties on data values from an arbitrary relational structure called the concrete domain. An example of concrete domain can be (Z, <, =), where the integers are considered as a relational structure over the binary order relation and the equality relation. Formulas of temporal logics with constraints are evaluated on data-words or data-trees, in which each node or position is labeled by a vector of data from the concrete domain. We call the constraints local because they can only compare values at a fixed distance inside such models. Several positive results regarding the satisfiability of LTL (linear temporal logic) with constraints over the integers have been established in the past years, while the corresponding results for branching time logics were only partial. In this work we prove that satisfiability of CTL* (computation tree logic) with constraints over the integers is decidable and also lift this result to ECTL*, a proper extension of CTL*. We also consider other classes of concrete domains, particularly ones that are \"tree-like\". We consider semi-linear orders, ordinal trees and trees of a fixed height, and prove decidability in this framework as well. At the same time we prove that our method cannot be applied in the case of the infinite binary tree or the infinitely branching infinite tree. We also look into extending the expressiveness of our logic adding non-local constraints, and find that this leads to undecidability of the satisfiability problem, even on very simple domains like (Z, <, =). We then find a way to restrict the power of the non-local constraints to regain decidability.

I assume that the reader of this thesis is reasonably familiar with artificial neural network (ANN) methods in computer science, including the multi-layer perceptron (ML?) and the backpropagation training method. I have not needed to use any difficult or esoteric mathematics; the major mathematical concept encountered in the thesis is the multiset (which is easy to grasp for anyone familiar with set theory). Certain chapters also make use of the notions of partial derivatives. inner products in arbitrary vector spaces, and metrics.

This thesis presents a novel method for performing multi-agent behaviour recognition without requiring large training corpora. The reduced need for data means that robust probabilistic recognition can be performed within domains where annotated datasets are traditionally unavailable (e.g. surveillance, defence). Human behaviours are composed from sequences of underlying activities that can be used as salient features. We do not assume that the exact temporal ordering of such features is necessary, so can represent behaviours using an unordered “bag-of-features”. A weak temporal ordering is imposed during inference to match behaviours to observations and replaces the learnt model parameters used by competing methods. Our three-tier architecture comprises low-level video tracking, event analysis and high-level inference. High-level inference is performed using a new, cascading extension of the Rao-Blackwellised Particle Filter. Behaviours are recognised at multiple levels of abstraction and can contain a mixture of solo and multiagent behaviour. We validate our framework using the PETS 2006 video surveillance dataset and our own video sequences, in addition to a large corpus of simulated data. We achieve a mean recognition precision of 96.4% on the simulated data and 89.3% on the combined video data. Our “bag-of-features” framework is able to detect when behaviours terminate and accurately explains agent behaviour despite significant quantities of low-level classification errors in the input, and can even detect agents who change their behaviour.