2019 Fiscal Year Annual Research Report
Exploring Deep Neural Networks for Temporal Information Extraction
Publicly Offered Research
| Project Area | Chronogenesis: how the mind generates time |
|---|
| Project/Area Number |
19H05318
|
|---|
| Research Institution | Kyoto University |
|---|
Principal Investigator |
程 飛 京都大学, 情報学研究科, 特定助教 (70801570)
|
|---|
| Project Period (FY) |
2019-04-01 – 2021-03-31
|
| Keywords | Information Extraction / Temporal Information / Relation Extraction |
|---|
| Outline of Annual Research Achievements |
We proposed two work to tackle the temporal relation classification task. 1) we combine pre-trained BERT and syntactic information to improve the classification performance. The results show BERT significantly outperforms the existing recurrent neural networks and syntactic information benefits the model further. 2) we proposed a model adopting global entity and multi-task learning for temporal relation classification. The results show state-of-the-art performance, compared to the existing work.
Our work improves state-of-the-art temporal relation classification performance close to the human annotation agreements. It can potentially benefit a series of down-streaming applications, e.g. extracting event timeline from news articles.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
According to the research plan, in th begining of FY2019, we completed the preliminary step of making a survey of temporal information extraction, related work and datasets. We built a GPU server with two high-end Nvidia Titan RTX 24G graphic cards for accelerating the experiment speed.
We have conducted a serious of experiments for our propsed models. The results suggest we obtained state-of-the-art performance against the existing work. We published our work in two domestic conference (言語処理学会2020, 人工知能学会2020). We have written the English paper for submiting the latest internation conference (EMNLP2020).
|
| Strategy for Future Research Activity |
1)The current work suggests pre-trained BERT outperforms the existing recurrent neural networks by large margin. We plan to explore more task-specific solutions by incorporating temporal knowledge for the task of temporal relation classification.
2)Temporal information extraction is composed by a pipeline task setting of temporal mention recognition, modality classification and relation extraction. The known drawback of pipeline models is error propagation in the early stages. We plan to intoduce joint models for overcoming this issue.
|
