Experiments on Unsupervised Learning for Extracting Relevant Fragments from Spoken Dialog Corpus

In: Proceedings of CoNLL-2000 and LLL-2000, pages 83-86, Lisbon, Portugal, 2000. Experiments on Unsupervised Learning for Extracting Relevant Fragments from Spoken Dialog Corpus Konstantin Biatov AT&T Bell-Labs Research 180 Park Avenue 07932-0971, Florham Park, N J, USA, Kbiatov@aol. tom Abstract In this paper are described experiments on un- supervised learning of the domain lexicon and relevant phrase fragments from a dialog cor- pus. Suggested approach is based on using do- main independent words for chunking and us- ing semantical predictional power of such words for clustering and automatic extraction phrase fragments relevant to dialog topics. 1 Introduction We are interested in rapid development of spo- ken dialog understanding systems. We present experiments on unsupervised learning of the domain lexicon and relevant phrase fragments from dialog corpus. Pereira (1993) described a method for auto- matically clustering words according to their distribution in particular syntactic context, for example verbs and direct objects of these verbs. By using preexisting concepts from Wordnet database Resnik (1998) described how to pre- dict words meaning from their distributional context. Both mentioned methods are fully unsupervised and are focused only on follow- ing word distribution. They describe the de- pendences between verb and noun as a direct object of the verb. A new method for gather- ing phrases into clusters was described by Arai (1999). This method uses following and preced- ing words distribution and call-types, associated with each utterance, but requires at the begin- ning labeling and transcribing a small number of the utterances. In contrast with the mentioned methodolo- gies, we are interested in finding a limited set of domain independent words (less than 1000) including prepositions, adverbs and adjectives and using these words for unsupervised clus- tering and automatic extraction of the relevant knowledge from dialog corpus. 2 Description of the algorithm There are four main steps in our approach. First step is to make automatically labeling and to chunk each sentence from spoken dia- log corpus into a set of short subphrases. We assume that in the spoken dialog a sentence consists of slightly related subphrases. In our experiment for labeling and chunking we use a relatively small set of domain independent words such as prepositions, determiners, arti- cles, modals and adverbs. For example articles: a, an, the; prepositions: in, with, about, un- der, for, of, to; determiners: some, many. The domain independent words are grouped in subvocabularies. For instance, subvocabulary <article> includes words a, an, the. Some subvocabularies include only one word. If a given sentence includes article A we'll replace it by the label (<article>A), article THE we'll replace by the label (<article>THE) and so on. Very important feature of our algorithm is that some of the words selected for tags can pre- dict the semantics of the followed words or sub- phrases. In all cases we could characterize this prediction as possibility. For example the word from predict semantics of the followed words or subphrases as a "start point", a "reason", a "source" or something else. For each of such …

Content truncated to 500 words