Spoken Document Summarization
Document Structure
|
Heterogeneous Information
|
Question Answering
In retrieval-based question answering (QA) system, the key steps involve query formulation, document retrieval, document re-ranking and answer generation. In this paper, we used tree-structured conditional random fields (CRF) leveraging parse tree, where phrase can be easily introduced into query, to formulate query. After formulating query, we use GOOGLE search engine to get relevant documents/webpages, and then re-ranking these documents using two-layer random walk with speech recognition information. The basic answer generation part includes answer type detection using binary SVM and answer generation using KNN.
- Publication: Interspeech 2014 as full paper. [pdf] [poster]
Slide and Lecture Alignment
Nowadays more and more online learning video help people learn from the Internet. To help people learning by video more efficiently, we propose structure support vector machine (SVM) to align the audio signal and slides. We first apply unsupervised method to get some pseudo alignment labels, and then use them as training data to train structure SVM model.
|