Language Information Access Technology team ( https://aip.riken.jp/labs/goalorient_tech/lang_inf_access_tech/ ) at RIKEN AIP
Speaker 1: (10 min) Satoshi Sekine
Title: Overview of Language Information Access Technology Team
We will introduce the research activities at our team.
Speaker 2(20 min.): Satoshi Sekine
Title: SHINRA project: overview and SHINRA2021-ML task (English)
SHINRA is a project to build structured knowledge base based on Wikipedia. We conducted two major tasks, extracting attribute values from Japanese Wikipedia articles, and categorize Wikipedia pages in 30 languages. We conducted these tasks under the scheme of “Resource by Collaborative Contribution (RbCC)”, where we call participants to the task for the evaluation projects, and the outputs are used to create the resource by collaborative manner. We are also planning to hold three tasks in 2021.
Speaker 3(20 min.): Daisuke Kawahara
Title: Acquiring Wide-coverage Semantic Frame Knowledge using the Wisdom of Crowds and Artificial Intelligence
To achieve natural language understanding by computers, semantic frame knowledge is essential. This knowledge is description of the world using “frames”, which represent a structure of an event, relation, or object with its participants. We are developing a method for acquiring wide-coverage semantic frame knowledge using the wisdom of crowds and natural language processing technologies. In this talk, we will introduce the overview of this project and the progress so far.
Speaker 4(20 min.): Shuhei Kurita
Title: Generative Language-Grounded Policy in Vision-and-Language Navigation Vision-and-language navigation (VLN) is a task to develop an agent that receives textual instruction and navigates in photorealistic 3D environments. Existing work for VLN uses neural cross-attention modules to combine the textual and visual information. However, it is still unclear that the cross-attention is enough to represent the relation between the cross-modal information. We utilize the vision-and-action-conditioned language modeling in the navigation and visualize how our model integrates the vision-and-language information with the language modeling scores.
Speaker 5(20 min.): Kouta Nakayama
Title: Reduce of target data to be predicted in resource construction shared-task using active sampling. The resource construction shared task is building resources using the prediction results submitted by the participants. However, in this task, participants need to make predictions on a massive amount of unlabeled data, which is an obstacle to participation. In this study, we propose a method that combines pre-distillation ensemble and active learning to build resources while reducing participants’ burden.
Speaker 6: (20 min.):関根聡 (in Japanese; same talk as Speaker 2)
Title: 森羅プロジェクト：概要とSHINRA2021-MLタスクの紹介 (日本語）
|日時||2021/02/17(水) 15:00 - 17:00|