All areas of language and speech technology must deal, in one way or another, with real text. Real text is messy: many things one nds in text | numbers, abbreviations, dates, currency amounts, acronyms . . . | are not standard words in that one cannot nd their properties by looking them up in a dictionary or deriving them morphologically from words that are in a dictionary, nor can one nd their pronunciation by an application of \letter-to-sound" rules. For many applications, such non-standard words | NSW's | need to be normalized, or in other words converted into standard words. Since the correct normalization of a given token often depends upon both the local context and the type (genre) of text one is dealing with, \text-normalization" is in general a very hard problem. Typical technology for text-normalization mostly involves sets of ad hoc rules tuned to handle one or two genres of text (often newspaper-style text), with the expected result that the techniques, do...

Since Shannon's original experiment in 1951, several methods have been applied to the problem of determining the entropy of English text. These methods were based either on prediction by human subjects, or on computer-implemented parametric models for the data, of a certain Markov order. We ask why computer-based experiments almost always yield much higher entropy estimates than the ones produced by humans. We argue that there are two main reasons for this discrepancy. First, the long-range correlations of English text are not captured by Markovian models and, second, computerbased models only take advantage of the text statistics without being able to "understand" the contextual structure and the semantics of the given text. The second question we address is what does the "entropy" of a text say about the author's literary style. In particular, is there an intuitive notion of "complexity of style" that is captured by the entropy? We present preliminary results based on a non-parametric entropy estimation algorithm that o er partial answers to these questions. These results indicate that taking long-range correlations into account significantly improves the entropy estimates. We get an estimate of 1.77 bits-per-character for a onemillion-character sample taken from Jane Austen's works. Also comparing the estimates obtained from several di erent texts provides some insight into the interpretation of the notion of "entropy" when applied to English text rather than to random processes, and the relationship between the entropy and the "literary complexity" of an author's style. Advantages of this entropy estimation method are that it does not require prior training, it is uniformly good over different styles and languages, and it seems to converge reasonably fast.

An automatic word classification system has been designed which processes word unigram and bigram frequency statistics extracted from a corpus of natural language utterances. The system implements a type of simulated annealing which employs an average class mutual information metric. Resulting classifications are hierarchical, allowing variable class granularity. Words are represented as structural tags --- unique n-bit numbers the most significant bit-patterns of which incorporate class information. Therefore, access to a structural tag immediately provides access to all classification levels for the corresponding word. The classification system has successfully revealed some of the structure of two natural languages, from the phonemic to the semantic level. The system has been favourably compared --- directly and indirectly --- with other word classification systems. Class based interpolated language models have been constructed to exploit the extra information supplied by structural...

The focus of this thesis proposal is to improve the ability of a computational system to understand spoken utterances in a dialogue with a human. Available computational methods for word recognition do not perform as well on spontaneous speech as we would hope. Even a state of the art recognizer achieves slightly worse than 70% word accuracy on (nearly) spontaneous speech in a conversation about a specific problem. To address this problem, I will explore novel methods for post-processing the output of a speech recognizer in order to correct errors. I adopt statistical techniques for modeling the noisy channel from the speaker to the listener in order to correct some of the errors introduced there. The statistical model accounts for frequent errors such as simple word/word confusions and short phrasal problems (one-to-many word substitutionsand many-to-one word concatenations). To use the model, a search algorithm is required to find the most likely correction of a given word sequence ...

We describe a simple variant of the interpolated Markov model with nonemitting state transitions and prove that it is strictly more powerful than any Markov model. More importantly, the non-emitting model outperforms the classic interpolated model on natural language texts under a wide range of experimental conditions, with only a modest increase in computational requirements. The non-emitting model is also much less prone to overfitting.

A speech recognizer is a device that translates speech into text. Many current speech recognizers contain two components, an acoustic model and a statistical language model. The acoustic model indicates how likely it is that a certain word corresponds to a part of the acoustic signal (e.g. the speech). The statistical language model indicates how likely it is that a certain word will be spoken next, given the words recognized so far. Even though the acoustic model might for example not be able to decide between the acoustically similar words "peach" and "teach", the statistical language model can indicate that the word "peach" is more likely if the previously recognized words are "He ate the". Current speech recognizers perform well on constrained tasks, but the goal of continuous, speaker independent speech recognition in potentially noisy environments with a very large vocabulary has not been reached so far. How can statistical language models be improved so that more complex tasks c...

Web documents are typically associated with many text streams, including the body, the title and the URL that are determined by the authors, and the anchor text or search queries used by others to refer to the documents. Through a systematic large scale analysis on their cross entropy, we show that these text streams appear to be composed in different language styles, and hence warrant respective language models to properly describe their properties. We propose a language modeling approach to Web document retrieval in which each document is characterized by a mixture model with components corresponding to the various text streams associated with the document. Immediate issues for such a mixture model arise as all the text streams are not always present for the documents, and they do not share the same lexicon, making it challenging to properly combine the statistics from the mixture components. To address these issues, we introduce an “openvocabulary” smoothing technique so that all the component language models have the same cardinality and their scores can simply be linearly combined. To ensure that the approach can cope with Web scale applications, the model training algorithm is designed to require no labeled data and can be fully automated with few heuristics and no empirical parameter tunings. The evaluation on Web document ranking tasks shows that the component language models indeed have varying degrees of capabilities as predicted by the cross-entropy analysis, and the combined mixture model outperforms the state-of-the-art BM25F based system.

This thesis presents the results of a study on information extraction from unrestricted Turkish text using statistical language processing methods. We have successfully applied statistical methods using both the lexical and morphological information to the following tasks: The Turkish Text Deasciifier task aims to convert the ASCII characters in a Turkish text, into the corresponding non-ASCII Turkish characters (i.e., "fi", ";5", "g", "", "", '5", and their upper cases).

Abstract. The paper concerns the use of multiple viewpoint representation schemes for prediction with statistical models of monophonic music. We present an experimental comparison of the performance of two techniques for combining predictions within the multiple viewpoint framework. The results demonstrate that a new technique based on a weighted geometric mean outperforms existing techniques. This finding is discussed in terms of previous research in machine learning. 1

We propose a unified neural network architecture and learning algorithm that can be applied to various natural language processing tasks including part-of-speech tagging, chunking, named entity recognition, and semantic role labeling. This versatility is achieved by trying to avoid task-specific engineering and therefore disregarding a lot of prior knowledge. Instead of exploiting man-made input features carefully optimized for each task, our system learns internal representations on the basis of vast amounts of mostly unlabeled training data. This work is then used as a basis for building a freely available tagging system with good performance and minimal computational requirements.