2019 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 first year, we have developed the basic language ability for the agents. Through a competing game of two communication agents, one agent is able to describe a given word with a simple natural language sentence. The generated sentences are syntactically and semantically reasonable, and can be well understood by human beings. And the other agent is able to read these generated sentence and to guess what the originally given word is.
The model for word-to-sentence interpretation and the model for sentence-to-word induction are based on two deep neural networks, respectively. These two models are trained side-by-side through a competing game, in which a correct round of interpretation-induction renders a positive feed back to encourage the models to make similar decisions in the future games while an incorrect round renders a negative feed back to encourage them to try other options in the future.
The developed language ability is found as the basis for the agents to describe their opinions for the other agents to infer those opinions from either similar agents or human beings.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The experiment of language ability development for long-vision conversation agents is done smoothly. The results suggest that the agents are able to interpret word into syntactically correct and semantically reasonable natural language sentences while at the same time are able to induce such sentences backward into words successfully.
|
| Strategy for Future Research Activity |
To learn long-vision conversation agents, we will continue to develop a self-concept identification ability and a strategic conversation ability for the agents as proposed in the research plan.
Specifically for this year, we will focus on developing the self-concept identification ability for the agents, by asking them to categorize everyone’s social identities and to recognize the commonality in those social identities as their self-concepts. The already learned encoder and decoder networks in the last year will be integrated into a new model with a categorization layer on top, in order to interpret and induce the identities of themselves and the others for a successful self-concept recognition.
|
Research Products
(29 results)
