Abstract: There is lack of research data on real application environment for interactive question-answering research. This paper collected a large number of online shopping customer service dialogue records as real application environment interactive question-answering corpus. First, the online customer service dialogue records were statistics and analysis. Then 174 groups service dialogues were randomly selected. Those dialogues were annotated and statistics on unnormal language, question relevance and question answer matching phenomena. The annotation and statistics results show that: high frequent dialogue sentences reached to large proportion, 15% of high frequent customer dialogue sentences covered 45% of all data customer sent out; 50% of dialogue sentences contained unnormal language phenomena; Anaphora relevance, omission relevance and common word sequences are the three most important features for judging relevance of client questions; more than 60% of service dialogue sentences are cross matching question answers pairs, and more than 50% of matching question answers pairs are recessive matching.

Key words: corpus annotation, interactive question answering, corpus analysis, customer service dialogue