Logo image
整合文本多層次表達與嵌入演講屬性之表徵學習於強健候用校長演講自動化評分系統
Thesis

整合文本多層次表達與嵌入演講屬性之表徵學習於強健候用校長演講自動化評分系統

黃文俞
Masters, 國立清華大學, 電機工程學系所
2016

Abstract

行為訊號處理 教育研究 口頭演講 自然語言處理 演講屬性 主題模型 behavior signal processing educational research oral presentation natural language processing self-defined attribute tag topic model
Among increasing needs of domain-aware computational models that can perform large-scale assessment like domain experts, the development of automatic oral presentation assessment system is important for education researchers. In this work, we extend the previous audiovisual framework on pre-service school principals’ 3-minute long impromptu speech using lexical information as additional modality. We aim at exploring effective feature set for text and enhancing the performance of lexical modality by manual tagging information. First, we utilize multi-level feature extraction approach, which consists of distributed representations and word categories, to derive features from the transcripts in the 2014 National Academy for Educational Research (NAER) oral presentation database, and improve the result of lexical modality from Spearman correlation of 0.378 to 0.493. Furthermore, inspired by folksonomy, we propose to enhance lexical feature by using a self-defined attribute tags of speech transcripts. Therefore, we carry out two different experiments: Exp I) considering the tags as other labels and employing multi-label learning, and Exp II) feature inspired by tags and topic modeling. After incorporating the two methods, the improved system obtains Spearman correlation of 0.574. Our experiment demonstrates the concept of self-defined attribute tags has capability to enrich lexical modality and improve system.

Metrics

1 Record Views

Details

Logo image