YJTI at the NTCIR-13 STC Japanese Subtask

In this paper, we describe our participation in the NTCIR-13 STC Japanese Subtask, in which we develop systems with the retrieval-based method. To retrieve reply texts for a given comment text, our system generates vector representations of both the comment and candidate replies by a 3-layer LSTM-RNN and evaluate distances between the comment vector and the candidate reply vectors, selecting the top-k nearest reply vectors and returning the corresponding reply texts. We also take Theme and Genre into consideration to decide the final ranking. In preparation of the candidate reply texts, we utilize all the comments and replies in the training set of Yahoo! News comments data. Our two runs are based on two different LSTM-RNN models, one trained over Twitter conversation data and the other mainly trained over Yahoo! Chiebukuro QA data. Each dataset has no less than 60 million pairs of text, and we aim to show how effective these combinations of large-scale datasets and largescale neural models are for developing dialog systems. In addition, we had an assumption that the model of Twitter conversation data would outperform that of Yahoo! Chiebukuro QA data as the task domain seemed to be more similar to conversations in microblog services than social question answering, but the reported results revealed that it was not the case.