Leveraging Social Annotation For Topic Language Model Adaptation

Social annotations such as Yahoo! Answers(1) already define broad coverage of hierarchical topic categories and include millions of documents annotated by web users. For topic language model (LM) adaptation, we present a novel approach of performing topic segmentation of LM training corpus via leveraging such social annotations, which may be more effective than unsupervised methods. Experimental results on the IWSLT-2011 TED ASR data sets demonstrate that we can achieve modest improvements when compared with the unsupervised methods such as clustering-based and LDA-based algorithms, e.g., the absolute WER improvement over the LDA-based method on the test set reaches 0.5 points.