2019 Fiscal Year Annual Research Report
A scalable privacy-preserving information retrieval system based on federated optimization, on-device intelligence and semantic matching
| Project/Area Number |
19H04215
|
|---|
| Research Institution | University of Tsukuba |
|---|
Principal Investigator |
于 海涛 筑波大学, 図書館情報メディア系, 助教 (30751052)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
Adam Jatowt 京都大学, 情報学研究科, 特定准教授 (00415861)
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 – 2024-03-31
|
| Keywords | Semantic Matching / Learning-To-Rank / Metric Optimisation |
|---|
| Outline of Annual Research Achievements |
This year we focused on developing semantic matching models based on deep neural networks. First, we explored how to directly optimise the evaluation metrics for ranking. A novel framework for direct optimization of IR metrics (e.g., nDCG and MAP) has been proposed. Second, we explored how to develop new matching methods in an adversarial way, and new matching approaches are proposed. The aforementioned methods have been extensively evaluated using benchmark collections and submitted to top international conferences (SIGIR2020 and CIKM2020). Moreover, we plan to disseminate the source codes via the project website, encouraging the entire community to improve this research towards new stages. Overall, the project has been smoothly conducted step by step according to the research proposal.
|
| 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 fact that our planned research is being conduced smoothly can be demonstrated as follow: First, we have proposed a number of novel semantic matching methods based on deep neural networks. Based on a series of experiments against standard datasets, the experiments have demonstrated their effectiveness for information retrieval. This echoes our proposal well. Second, we initiated an open-source project which includes not only our proposed semantic matching methods but also many other representative approaches within the same learning-to-rank framework. Thanks to this, it becomes easy for us to incorporate our future work. Moreover, we can get valuable feedbacks from world-wide researchers who are interested in our project.
|
| Strategy for Future Research Activity |
For the future work, our first major objective is developing an effective on-device intent-detection model. The first task is the algorithm design for query understanding by incorporating the recent techniques, especially the on-device methods based on deep neural networks. The factors, such as model size and computation cost, will be taken into account. The second task will be the detailed implementation of the on-device intent-detection model. The factor of robustness across different kinds of devices, such as mobile phones and laptops will be considered. The final task is to design a series of reasonable experiments including user studies in order to test the efficiency and effectiveness of the proposed intent-detection model.
|
