Title: System Design, Evaluation and Applications of Domain Term Extraction from Engineering Videos
Abstract: Understanding the meanings of domain-specific terms is essential to academic success in college-level STEM courses.However, it can be challenging for students to obtain correct spellings and precise definitions of domain-specific terms from lecture videos, given the limited lecture time, rarity of the terms, and possibly confusing pronunciations.To provide accurate speech-to-text transcription, and enable students to search for domain-specific terms and obtain term definitions in real-time, we designed, implemented, and evaluated the PhraseHinter tool, a text analytics pipeline that efficiently extracts domain-specific terms from engineering educational videos.The tool is lightweight and adaptable to online instruction platforms.In our approach, a series of key scenes are initially extracted from a lecture video using a novel scene detection algorithm.The algorithm employs a support vector machine to classify image differences based on pixel, face, and text similarity information [2].A domain corpus is built by using the optical character recognition (OCR) technique to extract text from the scenes.A sequence of text-cleaning algorithms is applied to the domain corpus to filter out invalid characters, punctuation, and stop words.Frequent phrases are identified using standard text mining algorithms including PrefixSpan [15].Using the TF-IDF metric [16], we compare the cleaned corpus to the background corpus to determine domain-specific terms and phrases.The proposed PhraseHinter tool has been successfully integrated into ClassTranscribe [4,11,3,19,2,10], a web-based video lecture platform, for multiple purposes: 1) Improve the Microsoft Azure Speech-to-Text accuracy by preparing a list of domainspecific terms with high confidence of occurrence in the audio, 2) Provide the input for the glossary tool, another text analytics service in ClassTranscribe that automatically generates the explanation for the domain-specific terms, and, currently in progress, 3) Provide search capability in order to locate the moments in the video when a domainspecific term is visually presented.In this paper, we evaluate the performance and accuracy of the PhraseHinter system based on a representative corpus of videos from different engineering disciplines with domain-specific terms and phrases correctly pre-identified.We share the evaluation dataset to the education community for further research.In addition, we present the source code and provide guidance for instructors who would like to adopt the tool.