Foreword : This blog records two works that I am more interested in in the dialogue generation related work of TNNLS journals in the past three years. First, I would like to share with you the way to accurately search for papers, and then briefly introduce the main ideas of the two works.
Table of contents
1. Precise search method for papers
-
In the search interface of the Web of Science website , select the search scope by restricting conditions such as the subject of the paper, publication name and publication year:
-
Refine the search results based on information such as article popularity, publication year, and article type:
After a precise search of the natural language generation work of the TNNLS journal in the past three years, it was found that there were not many works in this field in the journal. After extensively reading the works in the search results, I selected the following two interesting works to introduce my ideas.
2. Introduction to the paper
2.1 Two-stage dialog generation for user personality maintenance and reply diversity
- A Novel Two-Stage Generation Framework for Promoting the Persona-Consistency and Diversity of Responses in Neural Dialog Systems
- Submission information: See the publication in August 2021; mid-draft in August 2021; submission in February 2021
- The main idea: Design a two-stage dialogue generation network; first generate a set containing diverse dialogues, and then pass a personality consistency check module to correct the part of the generated dialogue set that does not conform to the user's personality, and finally generate diversity Dialogue that coexists with individuality.
2.2 Personalized dialogue generation based on multi-task learning and reinforcement learning
- Multitask Learning and Reinforcement Learning for Personalized Dialog Generation: An Empirical Study
- Submission information: See the publication in 2021; mid-draft in February 2020; submission in April 2018
- The main idea: use a combination of multi-task learning and reinforcement learning to generate personalized dialogue.
Add a user personality recognition module to assist in the generation of personalized dialogue; in reinforcement learning, design 3 kinds of reward measures (Q-learning, ploicy gradient, actor-critic).
References