View stunning SlideShares in full-screen with the new iOS app!Introducing SlideShare for AndroidExplore all your favorite topics in the SlideShare appGet the SlideShare app to Save for Later — even offline
View stunning SlideShares in full-screen with the new Android app!View stunning SlideShares in full-screen with the new iOS app!
A Spoken Dialogue System for Electronic Program Guide Information Access Seokhwan Kim, Cheongjae Lee, Sangkeun Jung, and Gary Geunbae Lee Pohang University of Science and Technology (POSTECH), South Korea ABSTRACT AUTOMATIC SPEECH SPOKEN LANGUAGE EPG DATABASE MANAGER In this paper, we present POSTECH Spoken Dialogue System RECOGNIZER UNDERSTANDING The main purpose of the EPG database manager is to build for Electronic Program Guide Information Access (POSSDS- a content database for the other modules in POSSDS-EPG EPG). POSSDS-EPG consists of automatic speech recognizer, To build the language model, the candidate utterances that The SLU module of POSSDS-EPG was constructed by a with minimal human effort. spoken language understanding, dialogue manager, system have high probability of being spoken by users are required. We concept spotting approach which aims to extract only the We chose an EPG website (http://www.epg.co.kr) dealing utterance generator, text-to-speech synthesizer, and EPG generate the candidate utterances automatically by using the essential information for predefined meaning representation with the information on Korean TV programs. The EPG database manager. Each module is designed and implemented dialogue examples in the existing example database and the slots. The semantic frame is made up of these slots including database manager builds a contents database from the to make an effective and practical spoken dialogue system. In retrieved result from the up-to-date EPG database. dialogue act, main action, and component slots for the EPG information on the website. particular, in order to reflect the up-to-date EPG information domain. An Existing Utterance We regarded the SLU problem as a classification problem, which is updated frequently and periodically, we applied a web- WEB PAGES I want to watch drama Hae-Sin around . which can be solved by statistical machine learning frame- mining technology to the EPG database manager, which builds [genre = drama], [program_name = Hae-Sin], [time = 9 pm] the content database based on automatically extracted works. To build a statistical model for the SLU problem, we Retrieved Results Contents Contents information from popular EPG websites. The automatically [genre = movie], [program_name = Monster], [time = 11 pm] should prepare the training corpus containing utterances that Filtering Tables generated content database is used by other modules in the [genre = sports], [program_name = Basketball], [time = 7 pm] have high probability of being spoken by users. We can easily Candidate Utterances create a training corpus by reusing the candidate utterances that Information Extracted system for building their own resources. Evaluations show that Extraction Information I want to watch movie Monster around . are used for building the language model in the speech our system performs EPG access task in high performance and I want to watch sports Basketball around . recognizer. can be managed with low cost. Building EPG DB DB POSSDS-EPG: POSTECH DIALOGUE MANAGER SYSTEM UTTERANCE EVALUATIONS SPOKEN DIALOGUE SYSTEM GENERATOR Manually Automatically Man To develop an effective and practical spoken dialogue system, Evaluation FOR EPG DOMAIN we proposed the situation-based dialogue management method TCR Managed System 0.76 aged System 0.72 The system utterance generator generates the literal sys-tem using dialogue examples. For the system utterance generation, STR 0.65 0.62 POSSDS-EPG consists of a set of appropriate modules that are utterances based on the system action tag and the utterance we automatically construct and index a dialogue example MRA 0.85 0.85designed to be connected to each other according to the order. The generating template. Each system action tag has at least one database from the dialogue corpus. The dialogue manager User Satisfaction 0.75 0.73overall system aims to output the synthesized spoken response utterance generating template which is constructed manually. TCR: User Perception of Task Completion Rate retrieves the best dialogue example for the current dialoguecorresponding to an input utterance spoken by the user.. The system utterance generating task is advanced by filling STR: Success Turn Rate situation, which includes a current user utterance, semantic slots in the template with proper values, such as retrieving MRA: Mean Recognition Accuracy frame and discourse history. From the retrieved result, the User Satisfaction = aTCR + bSTR + rMRA results from the EPG database, slot values in the semantic dialogue manager determines the system action tag from the User Utterance ASR Language frame, and constituents in the discourse history. MODEL pre-defined tag set. IMPLEMENTATION NLU SLU MODEL WEB Dialogue System Action Tag Inform_Channel Semantic Meta-Rules Frame User’s Utterance Corpus [program_name]은 [channel]에서 합니다. For DM Dialogue Automatic Utterance Template ( [program_name] eun [channel] e-seo hap-ni-da ) Manager User Semantic Discourse Indexing [program_name] is broadcasted on [channel]. Dialogue Intention Frame History Example DB System System Domain Slot Values [program_name = 해신, channel = KBS] Action Expert EPG DB Responses Dialogue Query Generation Meta-Rules System Response EPG DB Manager Example DB 해신은 KBS에서 합니다. For SRG Generator System Utterance ( Hae-Sin eun KBS e-seo hap-ni-da ) Utterance Similarity Retrieval Hae-Sin is broadcasted on KBS. Lexico-semantic Similarity Discourse history Similarity Best Dialogue Dialogue TTS System Utterance Examples Example Tie-breaking Overview of POSSDS-EPG System Architecture