Producing Contextually Appropriate Intonation is an Information-State Based Dialogue System

Producing Contextually Appropriate Intonation in an Information-State Based Dialogue System Ivana Kruijff-Korbayova i Stina Ericsson2 Kepa J. Rodriguez' Elena Karagjosova l 'University of the Saarland, Germany� 2University of Gothenburg, Sweden fkorbay,kepa,elkal@coli.uni - sb.de� stinae@ling.gu.se Abstract Our goal is to improve the contextual appropriateness of spoken output in a dialogue system. We explore the use of the information state to determine the information structure of system utter- ances. We concentrate on the realiza- tion of information structure by into- nation. We present the results of eval- uating the contextual appropriateness of varied system output produced with a text-to-speech synthesis system that supports intonation annotation. 1 Introduction Most commercial spoken dialogue systems use carefully scripted dialogues. This has the advan- tage that the system output can be pre-recorded and have high quality. The disadvantage is lim- ited dialogue flexibility, as user-initiative must be restricted to ensure the dialogue adheres to the script. More flexible dialogues need dynamically produced output. As the range of possible system utterances grows pre-recording becomes infeasi- ble, and speech synthesis becomes necessary. One challenge for systems using synthesized speech is the generation of contextually appro- priate intonation. With dynamically produced output, the same sequence of words may appear in different contexts, possibly needing different intonation. For example, the intonation of an an- swer needs to correspond to the respective ques- tion: whereas in (1S) the nuclear intonation cen- ter has a "default" placement, in (2S) it does not. I (1) U: What is the status of the stove? S: The stove is switched ON. H* LL% (2) U: Which device is switched on? S: The STOVE is switched on. H*� LL% Contextually inappropriate intonation may have negative effect on intelligibility or even lead to confusion; for example, when (1U) is answered with (2S), or (2U) with (1S), a mismatch arises. The details of relating intonation and other as- pects of realization to context are still a research topic. In this paper, we concentrate on one func- tion of intonation, to realize information struc- ture (IS). We consider IS as a level of meaning that unifies a range of interacting contextually- dependent aspects of utterance realization, en- countered in various combinations within and across languages. Determining the IS of system utterances according to context, and producing the corresponding realizations, are thus impor- tant steps in generating natural system output. Outline We show how we improve the spoken output of an information-state based dialogue system by controlling intonation using IS. §2 summarizes related work on controlling in- tonation in context. §3 gives background on IS, 1 We print words bearing a pitch accent in SMALL CAPI- TALS and use the ToBI ("Tones, Breaks and Indices") nota- tion for intonation, cf. http://www.ling.ohio-state.edurtobi/ 227 §4 on the information-state approach to dialogue. §5 presents rules for determining IS from the in- formation state. §6 describes the generation of spoken output with contextually varied intona- tion in our system using two off-the-shelf speech synthesis systems. §7 presents evaluation results. We close with a summary …

Content truncated to 500 words