2020 Fiscal Year Research-status Report
Learning conversation agents for long-vision response generation in multi-round communications
| Project/Area Number |
19K20345
|
|---|
| Research Institution | The University of Tokushima |
|---|
Principal Investigator |
康 シン 徳島大学, 大学院社会産業理工学研究部(理工学域), 助教 (80777350)
|
|---|
| Project Period (FY) |
2019-04-01 – 2022-03-31
|
| Keywords | long-vision conversation |
|---|
| Outline of Annual Research Achievements |
The purpose of this research plan is to learn the long-vision conversation agents from scratch by allowing them to make sequential decisions in a multi-round communication. For these agents, the language ability, the self-concept identification ability, and the strategic conversation ability would be developed through three strategic communication games respectively.
In the second year, we collected the data of conversation for learning the self-concept identification ability for the conversation agents. The data was composed of two parts, which were the conversation turns collected from Twitter and the tags that should represent the meaning of the conversation. We employed a latent dirichlet allocation (LDA) model and a rule based method for the tag extraction. These tags were treated as the self-identities for the agents to learn.
The agents learned two neural networks which are a generator and a discriminator. The generator network was responsible to generate sequences of word tokens which represents its self-identity, that is, the tag given by the LDA model or the rule-based method, and the discriminator network was responsible to discriminate the identities of the other agents by analyzing the word tokens generated by those agents.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
The experiment is based on an inter-conversation between agents, which requires them to generate the reasonable and meaningful responses for each conversation turn. Because generation of the agent responses were purely based on a generative model with a generative adversarial learning framework, controlling the quality of responses was extremely difficult. The quality of agents' inter-conversations has shown significant defect compared to the quality of human conversations.
|
| Strategy for Future Research Activity |
In the future work, the quality of inter-conversation should be improved, while the strategic conversation ability of conversation agents should be learned.
To improve the conversation quality, we would consider replacing the generator with GPT-based models, the retrieval-based models, and other slot-based approaches.
To learn the strategic conversation ability of conversation agents, we will design a long-vision reward for the agents through a reinforcement learning framework. Three agents will be given two types of social identities, and they must learn to identify the similarity of each other and to avoid to be recognized as the distinct one, by following the conversation rules. Agents get rewarded separately based on their decisions made through out the reinforcement learning framework.
|
| Causes of Carryover |
理由：文房具が3月に納品となり、支払いが完了していないため。
使用計画：文房具の支払が4月に完了する予定である。
|
