Construction of bilingual multimodal corpora of referring expressions in collaborative problem solving

This paper presents on-going work on constructing bilingual multimodal corpora of referring expressions in collaborative problem solving for English and Japanese. The corpora were collected from dialogues in which two participants collaboratively solved Tangram puzzles with a puzzle simulator. Extra-linguistic information such as operations on puzzle pieces, mouse cursor position and piece positions were recorded in synchronisation with utterances. The speech data was transcribed and time-aligned with the extra-linguistic information. Referring expressions in utterances that refer to puzzle pieces were annotated in terms of their spans, their referents and their other attributes. The Japanese corpus has already been completed, but the English counterpart is still undergoing annotation. We have conducted a preliminary comparative analysis of both corpora, mainly with respect to task completion time, task success rates and attributes of referring expressions. These corpora showed significant differences in task completion time and success rate.