2021 Fiscal Year Research-status Report
Measuring Group Interaction in Online Discussions and Application to Autonomous Agent Deliberation
| Project/Area Number |
20K11936
|
|---|
| Research Institution | Kyoto University |
|---|
Principal Investigator |
Rafik Hadfi 京都大学, 情報学研究科, 特定助教 (30867495)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
伊藤 孝行 名古屋工業大学, 工学(系)研究科(研究院), 教授 (50333555)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | Online Debates / Agent Deliberation / Predictive Deliberation / Conversational Agents / Augmented Democracy / Mutual Information / Interdependence / Natural Language |
|---|
| Outline of Annual Research Achievements |
The results in the second year consist in the application of new deliberative metrics to real-world discussion datasets with the focus on argumentative discussions (debates). This is rendered possible by building new algorithms that use NLP techniques (Deep Neural Networks, classifiers), added to the previously developed information theoretic methods to account for argumentative textual content (problems, solutions, and arguments). The result is a well-principled framework that can 1) measure deliberation in a debate, 2) predict the evolution of a deliberation, and 3) optimize the deliberation by maximizing an objective function. This framework was published in international top conferences (AAMAS, JSAI). These results allowed also for the development of methods to detect polarization in online debates. These results were introduced in invited talks and, panel discussions (MediaX, KICSS). The adopted approach and its implementation were introduced and taught at tutorials in international conferences (AAMAS, PRICAI). Another ramification is showing the interplay between the deliberation metrics and quality of discussions. It was found that such metrics correlated with textual readability. This finding allows for the evaluation of discussions without semantics. Other results include the development of a game theoretic model of fairness to assess interaction between self-interested agents in a particular domain. This will be used to quantify the fairness of agent-based deliberations and is a key element in assessing the trustworthiness and ethics behind the proposed approach.
|
| 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 project at this stage is progressing steadily and consistently due to the fact that the framework for the quantification of deliberation was tested successfully on several datasets from real-world debates and does corroborate the findings of many social science on deliberation. The algorithms were refined by integrating several components that were developed in the previous phase as well as new components such as the use of GPT language models to estimate information measures. This solid foundation is making it easy to apply the method to several scenarios and datasets, possibly with the help of artificial agent mediation. The use of agent mediation mechanism itself is currently being developed consistently.
|
| Strategy for Future Research Activity |
The plan for the next phase is to continue integrating the deliberation metrics in algorithms used by conversational artificial agents. These agents will later be tested in social experiments to see how well they could modulate the discussions by optimizing the previously defined deliberative objective functions. The hypothesis here, is that such agent mediated deliberation will increase the quality of the debates when humans are interacting with the agents. This phase requires running several controlled social experiments to evaluate the agents in real settings. Another task is to evaluate the fairness of the deliberation using the game theoretic models developed previously. To this end, the utility models will be transcribed from purely economic domains (game theoretic) into linguistic domains (NLP) defined over argumentative utterances.
|
| Causes of Carryover |
The amount was not spent due to the focus on research tasks, which did not require hiring subjects or purchasing devises. Most of the work required my existing computers, books, and consisted only in conducting computational simulations and experiments. However, the amount will be used within the next fiscal year to firstly pay for the fees of publications of three international journal articles that illustrate the essence of the results of this project. Furthermore, with the lifted travel restrictions, I will attend one international conference to present the results of the Deliberative Conventional Multiagent System, which is at the core of this project. Finally, I am expecting to use the rest of the fund to pay for social experiments with human subjects to test the impact of the agents in real-world social settings. This social experiment requires surveying, recruiting, and mentoring the subjects on using an online platform to interact with the conversational agents.
|
Research Products
(7 results)
