Response Sentence Modification Using a Sentence Vector for a Flexible Response Generation of Retrieval-based Dialogue Systems

We proposed a sentence modification method that makes retrieval-based dialogue response generation more flexible. A retrieval-based dialogue model is simple and robust, but it cannot output a response sentence that is not prepared in the example-response database. On the other hand, a generation-based dialogue model can flexibly generate the response, but the developer cannot control how the system responds to the input. The proposed model is based on the retrieval-based model, and the idea is that the model modifies the response sentence according to the difference between the example sentence and the input sentence. To realize the proposed method, we developed a new Transformer-based autoencoder model to calculate sentence vectors that contain enough information to reconstruct the input sentence. The proposed autoencoder, m-Query, can recover sentences much more accurately than the conventional model. Next, we trained the sentence modification model using a dialogue corpus and conducted an experiment to generate response sentences. As a result, it was shown that the proposed model generated modified responses reflecting the difference between the input and example sentences, such as subject or tense.