Paper Review 10: “Chitty-Chitty-Chat Bot”: Deep Learning for Conversational AI

In this post, the paper “ “Chitty-Chitty-Chat Bot”: Deep Learning for Conversational AI ” is summarized.

Link to paper: https://www.ijcai.org/proceedings/2018/0778.pdf

Yan, R., 2018, “Key-Value Retrieval Networks for Task-Oriented Dialogue,” In IJCAI, pages 5520–5526

This paper is actually a survey paper. We think that it might be a good idea to review such a paper for the last entry of our blog as a summary of previous papers we have reviewed so far. Although this paper highlights overview of methods as well as problem formulation and data collection, here we only mention approaches to bots. The writers divide methods into 3 parts. These are retrieval-based and generation-based conversational systems and combination of them, which is ensemble approaches.

In retrieval-based systems, given a user input utterance as the query, the system searches for candidate responses by matching metrics. The core of retrieval-based conversational systems is formulated as a matching problem between the query utterance and the candidate responses.

Another way to build a conversational system is to use language generation techniques. In these approaches, a response is generated from a model, not retrieved from a repository, and thus it cannot be guaranteed to be a legitimate natural language text at all times. With deep learning techniques applied, generation-based systems are greatly developed. In general, the conversational system applies the sequence-to-sequence generation manner. In general, the framework consists of an encoder-decoder framework using the sequence-to-sequence model.

Finally, a feasible solution is to utilize the advantage of both systems. Given a query utterance, an optimized system will find candidate responses from both the retrieval-based and the generation-based system, merge the candidates from both systems together, and output the best response in the merged list.