The last decade has witnessed the emergence of a new breed of human computer interfaces that combines several human language technologies to enable information access and transactional processing using spoken dialogue. In this paper, I discuss my view on the research issues involved in the development of such interfaces, describe the recent work done in this area at the MIT Laboratory for Computer Science, and outline some of the unmet research challenges, including the need to work in real domains, spoken language generation, and portability across domains and languages.

Interactive spoken dialogue provides many new challenges for natural language understanding systems. One of the most critical challenges is simply determining the speaker’s intended utterances: both segmenting a speaker’s turn into utterances and determining the intended words in each utterance. Even assuming perfect word recognition, the latter problem is complicated by the occurrence of speech repairs, which occur where speakers go back and change (or repeat) something they just said. The words that are replaced or repeated are no longer part of the intended utterance, and so need to be identified. Segmenting turns and resolving repairs are strongly intertwined with a third task: identifying discourse markers. Because of the interactions, and interactions with POS tagging and speech recognition, we need to address these tasks together and early on in the processing stream. This paper presents a statistical language model in which we redefine the speech recognition problem so that it includes the identification of POS tags, discourse markers, speech repairs and intonational phrases. By solving these simultaneously, we obtain better results on each task than addressing them separately. Our model is able to identify 72 % of turn-internal intonational boundaries with a precision of 71%, 97 % of discourse markers with 96 % precision, and detect and correct 66 % of repairs with 74 % precision.

In this paper a new, tree-trellis based fast search for finding the N best sentence hypotheses in continuous speech recognition is proposed. The search consists of two parts: a forward, time-synchronous, trellis search and a backward, time asynchronous, tree search. In the first module the well known Viterbi algorithm is used for finding the best hypothesis and for preparing a map of all partial paths scores time synchronously. In the second module a tree search is used to grow partial paths backward and time asynchronously. Each partial path in the backward tree search is rank ordered in a stack by the corresponding full path score, which is computed by adding the partial path score with the best possible score of the remaining path obtained from the trellis path map. In each path growing cycle, the current best partial path, which is at the top of the stack, is extended by one arc (word). The new tree-trellis search is different from the traditional time synchronous Viterbi search in its ability for finding not just the best but the N-best paths of different word content. The new search is also different from the A * algorithm, or the stack algorithm, in its capability for providing an exact, full path score estimate of any given partial (i.e., incomplete) path before its completion. When compared with the best candidate Viterbi search, the search complexities for finding the N-best strings are rather low, i.e., only a fraction more computation is needed.

Peter Heeman was born October 22, 1963, and much to his dismay his parents had already moved away from Toronto. Instead he was born in London Ontario, where he grew up on a strawberry farm. He attended the University of Waterloo where he re-ceived a Bachelors of Mathematics with a joint degree in Pure Mathematics and Com-puter Science in the spring of 1987. After working two years for a software engineering company, which supposedly used artificial intelligence techniques to automate COBOL and CICS programming, Peter was ready for a change. What better way to wipe the slate clear than by going to graduate school at the University of Toronto, but not without first spending the sum-mer in Europe. After spending two months in countries where he couldn’t speak the language, Peter became fascinated by language, and so decided to give computational linguistics a try.

Currently, most approaches to speech recognition are frame-based in that they represent speech as a temporal sequence of feature vectors. Although these approaches have been successful, they cannot easily incorporate complex modeling strategies that may further improve speech recognition performance. In contrast, segment-based approaches represent speech as a temporal graph of feature vectors and facilitate the incorporation of a wide range of modeling strategies. However, difficulties in segmentbased recognition have impeded the realization of potential advantages in modeling. This thesis

We developed a faster search algorithm that avoids the use of the N-Best paradigm until after more powerful knowledge sources have been used. We found, however, that there was little or no decrease in word errors. We then showed that the use of the N-Best paradigm is still essential for the use of still more powerful knowledge sources, and for several other purposes that are outlined in the paper. 1.

ABSTRACT 2 Speaker-independent speech recognition technology has made significant progress from the days of isolated word recognition. Today, state-of-the-art systems are capable of performing large vocabulary continuous speech recognition (LVCSR) on audio streams derived from complex information sources such as broadcast news and two-way telephone dialogs. A significant contribution to this advancement in technology is the development of search techniques that find suboptimal but accurate solutions in problems involving large search spaces and extremely complex statistical models. Moreover, these search strategies are capable of dynamically integrating information from a number of diverse knowledge sources to determine the correct word hypothesis, and limit the scope of the search by using a hierarchical search strategy. We refer to this problem as the decoding or search problem. This paper describes the complexity associated with decoding using hierarchical representations for linguistic and acoustic knowledge sources. An extensible object-oriented decoder available in the public domain, that leverages current state-of-the-art technology is described to illustrate these concepts. This decoder supports efficient handling of acoustic models for cross-word contextdependent phones, multiple pronunciations of words using lexical trees, and rescoring of word graphs based on N-gram language models in a single pass. It employs a state-of-the-art Viterbistyle dynamic programming algorithm, and is equipped with several heuristic pruning criteria to minimize the consumption of computational resources while maintaining good accuracy.

This paper presents a new technique for overcoming several types of speech recognition errors by post-processing the output of a continuous speech recognizer. The post-processor output contains fewer errors, thereby making interpretation by higher-level modules, such as a parser, in a speech understanding system more reliable. The primary advantage to the post-processing approach over existing approaches for overcoming SR errors lies in its ability to introduce options that are not available in the SR module's output. This work provides evidence for the claim that a modern continuous speech recognizer can be used successfully in "black-box" fashion for robustly interpreting spontaneous utterances in a dialogue with a human.

Information extraction systems incorporate multiple stages of linguistic analysis. Although errors are typically compounded from stage to stage, it is possible to reduce the errors in one stage by harnessing the results of the other stages. We demonstrate this by using the results of coreference analysis and relation extraction to reduce the errors produced by a Chinese name tagger. We use an N-best approach to generate multiple hypotheses and have them re-ranked by subsequent stages of processing. We obtained thereby a reduction of 24 % in spurious and incorrect name tags, and a reduction of 14 % in missed tags. 1

problem in a weighted automaton. This problem arises commonly in speech recognition applications when a ranked list of unique recognizer hypotheses is desired. We believe this is the first n-best algorithm to remove redundant hypotheses before rather than after the n-best determination. We give a detailed description of the algorithm and demonstrate its correctness. We report experimental results showing its efficiency and practicality even for large n in a 40; 000-word vocabulary North American Business News (NAB) task. In particular, we show that 1000-best generation in this task requires negligible added time over recognizer lattice generation.