High Performance Segmentation of Spontaneous Speech Using Part of Speech and Trigger Word Information

High Performance Segmentation of Spontaneous Speech Using Part of Speech and Trigger Word Information Marsal Gavaldh Klaus Zechner Interactive Systems Labs. Interactive Systems Labs. Language Technologies Institute Comp. Ling. Program School of Computer Science Department of Philosophy Carnegie Mellon University Carnegie Mellon University Pittsburgh, PA 15213, USA Pittsburgh, PA 15213, USA marsal+@cs, cmu. edu zechner+©andrew, cmu. edu Gregory Aist Comp. Ling. Program Department of Philosophy Carnegie Mellon University Pittsburgh, PA 15213, USA aist+@andrew, cmu. edu Abstract We describe and experimentally evaluate an efficient method for automatically de- termining small clause boundaries in spon- taneous speech. Our method applies an ar- tificial neural network to information about part of speech and trigger words. We find that with a limited amount of data (less than 2500 words for the training set), a small sliding context window (+/-3 to- kens) and only two hidden units, the neural net performs extremely well on this task: less than 5% error rate and F-score (com- bined precision and recall) of over .85 on unseen data. These results prove to be better than those reported earlier using different approaches. 1 Introduction In the area of machine translation, one important in- terface is that between the speech recognizer and the parser. In the case of human-to-human dialogues, the speech recognizer's output is a sequence of turns (a contiguous segment of a single speaker's utter- ance) which in turn can consist of multiple clauses. Lavie et al. (1996) discuss that using smaller units rather than whole turns can greatly facilitate the task of the parser since it reduces the complexity of its input. The problem is thus how to correctly segment an utterance into clauses. The segmentation procedure described in Lavie et al. (1996) uses a combination of acoustic infor- mation, statistical calculation of boundary-trigrams, some highly indicative keywords and also some heuristics from the parser itself. Stolcke and Shriberg (1996) studied the relevance of several word-level features for segmentation per- formance on the Switchboard corpus (see Godfrey et al. (1992)). Their best results were achieved by using part of speech n-grams, enhanced by a couple of trigger words and biases. 12 Another, more acoustics-based approach for turn segmentation is reported in Takagi and Itahashi (1996). Palmer and Hearst (1994) used a neural network to find sentence boundaries in running text, i.e. to determine whether a period indicates end of sentence or end of abbreviation. The input to their network is a window of words centered around a period, where each word is encoded as a vector of 20 reals: 18 val- ues corresponding to the word's probabilistic mem- bership to each of 18 classes and 2 values represent- ing whether the word is capitalized and whether it follows a punctuation mark. Their best result of 98.5% accuracy was achieved with a context of 6 words and 2 hidden units. In this paper we bring their idea to the realm of speech and investigate the performance of a neural network on the task of turn segmentation using parts of …

Content truncated to 500 words