Computational Psycholinguistics Tokyo

プロジェクト名，担当者名: 対照言語学の観点から見た日本語の音声と文法 (文法研究班「名詞修飾表現」)
窪田悠介 (国立国語研究所理論・対照研究領域准教授); 大規模コーパスを利用した言語処理の計算心理言語学的研究
大関洋平 (早稲田大学理工学術院講師); コーパスアノテーションの拡張・統合・自動化に関する基礎研究 (係り受け班)
浅原正幸 (国立国語研究所コーパス開発センター教授)
開催期日: 2019年5月24日 (金) 11:00～13:00
開催場所: 早稲田大学西早稲田キャンパス 55号館 N棟 1階第1会議室 (東京都新宿区大久保3-4-1)
交通アクセス
関連サイト: Computational Psycholinguistics Tokyo

プログラム

11:00～12:00 "What can psycholinguistics and deep learning contribute to each other?" Tal LINZEN (Johns Hopkins University)

Deep learning systems with minimal or no explicit linguistic structure have recently proved to be surprisingly successful in language technologies. In this talk, I'll discuss ways in which psycholinguistics -- in particular, the study of humans' acquisition and comprehension of syntax -- can help guide these developments and simultaneously benefit from them. Illustrating one direction of this relationship, I will show how theories and experimental methods from psycholinguistics can be instrumental in identifying the remaining limitations of existing models and improving those models further. In the other direction, neural networks can be used to address classic questions in linguistics and psycholinguistics, in particular by (1) providing a platform for testing for the necessity and sufficiency of explicit structural biases in the acquisition of syntactic transformations, and (2) providing highly accurate and syntactically informed estimates of word predictability, which can serve to test theories that ascribe a central role to predictability in explaining human sentence processing.

12:00～13:00 "Subregular morphology: Structures, grammars, and learning" Jonathan RAWSKI (Stony Brook University)

This talk overviews recent advances in the computational nature of morphology. I will analyze the data structures and grammar expressivity necessary to characterize morphotactics, concatenative and nonconcatenative transformations, and provably efficient learning algorithms for these structured classes of grammars. I show how tradeoffs in data structures point to a unified view of memory requirements for morphological grammars, providing a clear way to test the cognitive representations characteristic of human morphological knowledge.