Towards Continuous Dialogue Corpus Creation: writing to corpus and generating from it

Towards Continuous Dialogue Corpus Creation: writing to corpus and generating from it Andrei Malchanau1, Volha Petukhova1 and Harry Bunt2 1Spoken Language Systems Group, Saarland University, Germany 2Tilburg Center for Communication and Cognition, Tilburg University, The Netherlands {andrei.malchanau,v.petukhova}@lsv.uni-saarland.de; harry.bunt@uvt.nl Abstract This paper describes a method to create dialogue corpora annotated with interoperable semantic information. The corpus development is performed following the ISO linguistic annotation framework and primary data encoding initiatives. The Continuous Dialogue Corpus Creation (D3C) methodology is proposed, where a corpus is used as a shared repository for analysis and modelling of interactive dialogue behaviour, and for implementation, integration and evaluation of dialogue system components. All these activities are supported by the use of ISO standard data models including annotation schemes, encoding formats, tools, and architectures. Standards also facilitate practical work in dialogue system implementation, deployment, evaluation and re-training, and enabling automatic generation of adequate system behaviour from the data. The proposed methodology is applied to the data-driven design of two multimodal interactive applications - the Virtual Negotiation Coach, used for the training of metacognitive skills in a multi-issue bargaining setting, and the Virtual Debate Coach, used for the training of debate skills in political contexts. Keywords: dialogue resources, interoperable semantic annotations, international standards, dialogue system design 1. Introduction A steadily growing interest can be observed in data-driven modelling of phenomena related to natural language, vi- sion, behavioural and organizational processes. Data have become essential to advance the state of the art in many ar- eas including the development of spoken (multimodal) di- alogue systems. Conversational applications such as Ap- ple’s Siri, Microsoft’s Cortana and Google Now became successful and robust partly due to the amount of real user data available to their developers. The most recent trend in dialogue system design involves end-to-end dialogue sys- tems using neural network models trained on previously collected dialogue data, without any detailed specification of dialogue states (Wen et al., 2017; Bayer et al., 2017). This requires large amounts of data to cover a reasonable number of possible dialogue states and participant actions. Dialogue data have often been collected in Wizard-of-Oz experiments (Dahlb¨ack et al., 1993), where the dialogue system is replaced by a human Wizard who simulates the system’s behaviour according to a pre-defined script. An alternative is to use simulated users. With good user modelling, a dialogue system could be rapidly prototyped and evaluated. Simulated data sets are, however, rather scarce (Schatzmann et al., 2006). Resources for data-driven learning of task-oriented systems are also collected with existing systems (Bennett and Rud- nicky, 2002; Henderson et al., 2014). For example, the Di- alPort project addresses the need for dialogue resources by offering a portal connected to different existing dialogue systems (Lee et al., 2017). Learning algorithms have also been proposed to train a di- alogue system online. System behaviour is initially learned from a minimal number of dialogues and is then optimized as more data arrives (Daubigney et al., 2012). As a data col- lection strategy this approach may not be really …

Content truncated to 500 words