LE TRINDIKIT A toolkit for building and experimenting with dialogue move engines and systems, based on the information state approach
LE TRINDIKIT Architecture based on information states Modules (dialogue move engine, input, interpretation, generation, output etc.) access the information state Resources (databases, lexicons, domain knowledge etc.) are hooked up to the information state
LE Features TRINDIKIT enables rapid prototyping of simple systems, as well as advanced research systems TRINDIKIT’s modularity provides simple and efficient methods for reconfigurationModularity also encourages reusability Explicit information state datastructure makes systems more transparent Update rules provide an intuitive way of formalising theories in a way which can be used by a system
LE TRINDIKIT architecture
LE A library of datatype definitions (records, DRSs, sets, stacks etc.) in terms of conditions and operations which provide the basic building blocks for formulating information state update rules Methods for accessing the information state, implemented as a syntax for update rules Methods and tools for visualising the information state debugging facilities Generic WWW interface* * planned for TRINDIKIT 2.0 TRINDIKIT components
LE TRINDIKIT components (cont’d) A language for defining update algorithms used by TRINDIKIT modules to coordinate update rule application A language for defining basic control structure, allowing both serial and asynchronous* systems A library of basic ready-made modules for input/output (incl. speech), interpretation, generation, handling dialogue plans etc.; GUI* for increased usability and overview, including tools for building systems Resource interfaces, e.g. to hook up grammars, SQL databases, WWW agents etc.
LE Building a system Type of information state (DRS, record, set of propositions, frame,...) A set of dialogue moves Information state update rules, including rules for integrating and selecting moves DME Module algorithm(s), including control algorithm Resources: databases, grammars, plan libraries etc., or external modules (optionally) additional modules, datatypes
LE Building a system (cont’d) TRINDIKIT dialogue theory (IS, rules, moves etc,) domain knowledge (resources) generic system domain-specific system
LE Systems GoDiS and IMDiS – information state based on Questions Under Discussion MIDAS – DRS information state, first-order reasoning EDIS – information state based on Poesio&Traum Autoroute – information state based on Conversational Game Theory
LE An experimental dialogue system built using the TRINDIKIT GoDiS
LE GoDiS architecture
LE Information state based Ginzburg’s notion of Questions Under Discussion (QUD) Dialogue plans to drive dialogue Information-seeking dialogue QUD and dialogue plans provide the right level of complexity for information-seeking dialogue; simpler then general reasoning and planning, but more versatile then frame-filling and finite automata GoDiS features
LE PRIVATE =PLAN = AGENDA = { findout(?return) } SHARED = findout(? x.month(x)) findout(? x.class(x)) respond(? x.price(x)) COM = dest(paris) transport(plane) task(get_price_info) QUD = LM = { ask(sys, x.origin(x)) } BEL = { } TMP = (*mirrors SHARED*) Sample GoDiS information state
LE integrateAnswer Before an answer can be integrated by the system, it must be matched to a question on QUD pre: eff: in( SHARED.LM, answer(usr, A)) fst( SHARED.QUD, Q) relevant_answer(Q, A) pop( SHARED.QUD ) reduce(Q, A, P) add( SHARED.COM, P) Sample update rule
LE Adapted for travel agency and autoroute domains (also being adapted to cinema booking, and to function as dialogue interface to a handheld computer and a mobile phone) Lexicons for english and swedish; lexicon can be switched dynamically Question and task accommodation to enable mixed initative; influenced by Lewis’ (1979) Focus intonation based on information state contents GoDiS features (cont’d)
LE GoDiS has been hooked up to IBM’s ViaVoice, an off-the-shelf speech synthesiser Default intonation gives incoherent impression, mainly because of presuppositions failures Chu & Nakatani (2000) put focus on features; this is insufficient in e.g. the travel agency domain Focus and intonation
LE GoDiS can control focus intonation based on the contents of the information state Currently implemented: a simple version of Pulman’s theory of parallellism; focus depends on what has been established in the dialogue In the near future: also use QUD to control focus intonation Focus and intonation (cont’d)