Modeling Concept Dependencies in a Scientific Corpus

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 866–875, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Modeling Concept Dependencies in a Scientific Corpus Jonathan Gordon, Linhong Zhu, Aram Galstyan, Prem Natarajan, and Gully Burns USC Information Sciences Institute Marina del Rey, CA, USA {jgordon, linhong, galstyan, pnataraj, burns}@isi.edu Abstract Our goal is to generate reading lists for stu- dents that help them optimally learn techni- cal material. Existing retrieval algorithms return items directly relevant to a query but do not return results to help users read about the concepts supporting their query. This is because the dependency structure of concepts that must be understood before reading material pertaining to a given query is never considered. Here we formulate an information-theoretic view of concept de- pendency and present methods to construct a “concept graph” automatically from a text corpus. We perform the first human evalu- ation of concept dependency edges (to be published as open data), and the results ver- ify the feasibility of automatic approaches for inferring concepts and their dependency relations. This result can support search ca- pabilities that may be tuned to help users learn a subject rather than retrieve docu- ments based on a single query. 1 Introduction Corpora of technical documents, such as the ACL Anthology, are valuable for learners, but it can be difficult to find the most appropriate documents to read in order to learn about a concept. This problem is made more complicated by the need to trace the ideas back to those that need to be learned first (e.g., before you can learn about Markov logic networks, you should understand first-order logic and probability). That is, a crucial question when learning a new subject is “What do I need to know before I start reading about this?” To answer this question, learners typically rely on the guidance of domain experts, who can devise pedagogically valuable reading lists that order doc- Automatic Speech Recognition (ASR) with HMMs Noisy Channel Model Viterbi Decoding for ASR Training ASR Parameters Viterbi Algorithm Dynamic Programming Decoding/ Search Problem HMMs Markov Chains HMM Pronunciation Lexicon Iterative Parameter Estimation with EM Gaussian Acoustic Model Discrete Fourier Transforms Gaussian Mixture Models Phonemes N-gram Language Model Figure 1: A human-authored concept graph excerpt, showing possible concepts related to automatic speech recognition and their concept dependencies. uments to progress from prerequisite to target con- cepts. Thus, it is desirable to have a model where each concept is linked to the prerequisite concepts it depends upon – a concept graph. A manually constructed concept graph excerpt related to au- tomatic speech recognition is shown in Figure 1. The dependency relation between two concepts is interpreted as whether understanding one concept would help a learner understand the other. Representing a scientific corpus in this way can improve tasks such as curriculum plan- ning (Yang et al., 2015), automatic reading list generation (Jardine, 2014), and improving educa- tion quality (Rouly et al., 2015). Motivated by the importance …

Content truncated to 500 words