2018 Fiscal Year Research-status Report
Development of a System for Collecting Context Data for Large-Scale Inverse Reinforcement Learning
| Project/Area Number |
17K00295
|
|---|
| Research Institution | Hokkaido University |
|---|
Principal Investigator |
RZEPKA Rafal 北海道大学, 情報科学研究科, 助教 (80396316)
|
|---|
| Project Period (FY) |
2017-04-01 – 2021-03-31
|
| Keywords | context processing / implicit knowledge / semantic chains / text generation |
|---|
| Outline of Annual Research Achievements |
The second year of the grant realization was concentrated on a) developing methods for the new knowledge generation in order to allow machines to read context which is not explicitly shown in Japanese text b) testing automatically retrieved knowledge and contexts with machine learning methods on English (motivation discovery) and Chinese (sentiment analysis).
a) In the past, researchers calculated the relevance between two events appearing in the same document using distributed representations for machine learning. Since such methods handle only explicitly written information, it is difficult for a machine to, e.g. automatically answer questions about context. To tackle this problem, a method for generating event chains using relevance filtering and for augmenting implicit knowledge as places as an example was created. Temporal events were created and their implicit relationship by relevance using a distributed representation of words was calculated. Likewise, relevant places were also added. Experiments were performed to show the temporal and semantic naturalness of the knowledge completion process needed to enrich machine learning.
b) Various experiments on the influence of context were performed and their results were presented during national and international conferences. The main findings were that 1) emoticons are one of the means which can enrich context and provide additional information helping to estimate affect beyond positivity and negativity; 2) that computer can learn a motivation degree of an advice depending on the context of the request for this advice.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
I have discovered that existing corpora give a vast source of knowledge, however, the machines lack an ability to fill in missing parts of context as places and agents, which is well visible in Japanese language. For this reason, I needed to work on a new algorithm to deal with this phenomenon which slightly delayed the process of generating Shankian scripts. Nevertheless, planned work on contextual understanding of emotions and motivations generated more results as expected, therefore I assess this year progress as acceptedly paced in general.
|
| Strategy for Future Research Activity |
I will continue developing the knowledge acquisition system that adds implicit knowledge about given situations. As stated in the research plan, I will broaden tests of the acquired knowledge beyond the sentiment and experiment with vectors for deeper semantic representations to discover further dependencies between places, actors and acts to allow understanding short utterances by extending assumptions about the context. I plan to show how, through interaction with users, a machine can narrow the enormous data of possible interpretations. This approach should lead to higher explainability of AI systems, especially in the automatic moral judgement scenarios.
|
