Character Mining: Machine Comprehension on Multiparty Dialogues

   May 23, 2017, 4PM

302- 309-1

 

Abstract:

 

This seminar presents the Character Mining project that extracts and infers various information about individual characters in multiparty dialogues. The long-term goal of this project is to develop a machine comprehension system that understands human conversations. Currently, this project focuses on three tasks, character identification, emotion detection, and crossdomain document retrieval. Character identification is an entity linking task that identifies each mention referring to a human (e.g., she, mom) as a certain character in the dialogues. We introduce the agglomerative convolutional neural network model that gives the F1 score of 86.76% and the accuracy of 95.30% for this task. Emotion detection classifies each utterance in the dialogues to one of seven emotions, that are sad, mad, scared, powerful, peaceful, joyful, and neutral. We introduce the sequence-based convolutional neural network model that shows the accuracies of 37.51% and 53.76% for fine and coarse-grained emotions, respectively. Finally, cross-domain document retrieval is challenged, where the target documents are dialogues, and the queries are sentences describing events in those dialogues. We suggest a structure reranking model to improve the initial ranking from a search engine by utilizing syntactic and semantic structures, which shows an over 4% improvement