Chats and Chunks: Annotation and Analysis of Multiparty Long Casual Conversations

Chats and Chunks: Annotation and Analysis of Multiparty Long Casual Conversations Emer Gilmartin, Carl Vogel, Nick Campbell Trinity College Dublin Ireland {gilmare, vogel, nick}@tcd.ie Abstract Casual talk or social conversation is a fundamental form of spoken interaction. Corpora of casual talk often comprise relatively short dyadic conversations, although research into such talk has found longer multiparty interaction to be very common. This genre of spoken interaction is attracting more interest with attempts to build more friendly and natural spoken dialog systems. To study longer multiparty casual talk, we have assembled a collection of conversations from three existing corpora. We describe the collection, organization, and annotation of structural chat and chunk phases in these conversations. we then review our preliminary results, noting significant differences in the distribution of overlap, laughter and disfluency in chat and chunk phases, and finding that chunk dominates as conversations get longer. We outline our continuing work on gaining greater understanding of this genre of spoken interaction, with implications for the design of spoken dialog systems. Keywords: casual conversation, multimodal corpus, annotation 1. Introduction Many of the practicalities of life are managed through spo- ken interaction. Such interaction is task-based or practical, facilitating activities such as service encounters (buying a pizza), workplace projects (meetings), or education (tutori- als, discussions). People also talk for social reasons – such interactional talk or casual conversation happens whenever people congregate, from short greetings at bus stops to in- teractions such as dinner party conversations lasting sev- eral hours. Although casual conversation is ubiquitous, it is not as well studied as practical talk. We investigate the structure of such talk, particularly in multiparty long (c. 1 hour) conversations. Dialog research has advanced greatly through studies of corpora of relevant speech inter- action. In recent years, several high quality corpora have been made available and have underpinned a wide range of research. However, there remain very few collections of longer stretches of multiparty casual talk, and although the use of existing corpora of multiparty talk in scenarios such as meetings is convenient, it is unclear that results would generalise to casual talk. We have gathered a set of six 3 to 5 party conversations, drawn from three corpora, which have been manually segmented, transcribed, and an- notated structurally into conversational phases - interactive chat phases and chunk phases where one speaker domi- nates. We have performed several analyses of the structure of these phases in terms of the occurrence of speech, laugh- ter, and silence, and in the distribution of disfluencies and overlap. The motivation for this work is two-fold. First, the scientific goal of greater understanding of this fundamental human behaviour, and second, to better model conversation in human-machine interaction, particularly in dialog sys- tems which engage in social, companionable talk. Below, we briefly review theories of casual conversation, describe the collection and annotation of the data used, and report on our current and future work towards a fuller account of the structure of multiparty casual conversation. 2. Casual Conversation Studies of …

Content truncated to 500 words