Measuring the Structural Importance through Rhetorical Structure Index

Proceedings of NAACL-HLT 2013, pages 783–788, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Measuring the Structural Importance through Rhetorical Structure Index Narine Kokhlikyan†, Alex Waibel† ‡, Yuqi Zhang†, Joy Ying Zhang‡ †Karlsruhe Institute of Technology Adenauerring 2 76131 Karlsruhe, Germany ‡ Carnegie Mellon University NASA Research Park, Bldg. 23 Moffett Field, CA 94035 narine.kokhlikyan@student.kit.edu, waibel@cs.cmu.edu, yuqi.zhang@kit.edu, joy.zhang@sv.cmu.edu Abstract In this paper, we propose a novel Rhetorical Structure Index (RSI) to measure the struc- tural importance of a word or a phrase. Un- like TF-IDF and other content-driven mea- surements, RSI identifies words or phrases that are structural cues in an unstructured doc- ument. We show structurally motivated fea- tures with high RSI values are more useful than content-driven features for applications such as segmenting unstructured lecture tran- scripts into meaningful segments. Experi- ments show that using RSI significantly im- proves the segmentation accuracy compared to TF-IDF, a traditional content-based feature weighting scheme. 1 Introduction Online learning, a new trend in distance learning, provides numerous lectures to students all over the world. More than 19,000 colleges offer thousands of free online lectures1. Starting from video record- ings of lectures which sometimes also come with the presentation material, a set of processes can be ap- plied to extract information from the unstructured data to assist students in browsing, searching and understanding the content of the lecture. These pro- cesses include automatic speech recognition (ASR) which converts the audio to text, lecture segmen- tation which inserts paragraph boundaries and adds section titles to the lecture transcriptions, automatic summarization that generates a short summary from 1http://www.thebestcolleges.org/ free-online-classes-and-course-lectures/ the full lecture, and lecture translation that translates the lecture from the original language to the native language of the student. The transcription of a lecture generated by the ASR system is a sequence of words which does not contain any structural information such as para- graph, section boundaries and section titles. Zhang et al. (2007; 2008; 2010) used acoustic and lin- guistic features for rhetorical structure detection and summarization. They showed that linguistic features such as TF-IDF are the most influential in segmenta- tion and summarization and that knowing the struc- ture of a lecture can significantly improve the perfor- mance of lecture summarization. Our experiments with a real-time lecture translation system also show that displaying the rolling translation results of a live lecture with proper paragraphing and inserted sec- tion titles makes it easier for students to grasp the key points during a lecture. In this paper, we apply existing algorithms, namely the Hidden Markov Model (HMM) (Gales and Young, 2007) to unstructured lecture transcrip- tion to infer the underlying structure for better lec- ture segmentation and summarization. HMM has been successfully applied in early works (van Mul- bregt et al., 1998; Sherman and Liu, 2008) for text segmentation, event tracking and boundary detec- tion. The focus of this work is to identify cue words and phrases that are good indicators of lec- ture structure. Intuitively, words and phrases such as “last …

Content truncated to 500 words