2013 Fiscal Year Annual Research Report
あらゆる句の正規化：事実の抽出と発見のための大規模テキスト解析
| Project/Area Number |
13F03041
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
鶴岡 慶雅 東京大学, 大学院工学系研究科, 准教授
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
STENETORP Pontus 東京大学, 大学院工学系研究科, 外国人特別研究員
|
|---|
| Keywords | 深層学習 / 句の表現 / ジョイントモデリング / 意味解析 / 構文解析 / ベクトル空間モデル |
|---|
| Research Abstract |
Over the last year we have been working on the problem of creating representations of phrases. For this we have followed two approaches. First, a more traditional approach of creating the representation while analysing the syntactic structure of the sentence (Dependency Parsing). Work which Dr. Stenetorp recently presented. Secondly, a more unconventional approach of avoiding any reliance on syntactic structure and not just learn the representation but also the structure jointly. The latter is work Dr. Stenetorp started during his visit to Stanford earlier this year and we so far have promising results, being able to achieve performance on-par with models that utilise manually annotated structures. What the models have in common is that they draw upon recent advances in Deep Learning and allow us to learn a representation jointly with an end task such as semantic category disambiguation or phrase normalisation.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We have made good progress towards the goal of classifying phrases and assigning them an appropriate semantic category. Both of the two approaches we have worked on are showing promise and the joint Dependency Parsing approach has been picked up by several other external research groups, among them the NLP group at the Nara Institute of Science and Technology.
|
| Strategy for Future Research Activity |
We are currently working on two publications, one extension of Dr Stenetorp's work from last year on jointly composing a phrase representation while performing Dependency Parsing and also wrapping up the collaborative work with Stanford. Once these publications have been cleared we expect that we can move on working with both styles of phrasal representations and use them as the basis for our joint phrase normalisation system.
|
