Improving biterm topic model with word embeddings

As one of the fundamental information extraction methods, topic model has been widely used in text clustering, information recommendation and other text analysis tasks. Conventional topic models mainly utilize word co-occurrence information in texts for topic inference. However, it is usually hard to extract a group of words that are semantically coherent and have competent representation ability when the models applied into short texts. It is because the feature space of the short texts is too sparse to provide enough co-occurrence information for topic inference. The continuous development of word embeddings brings new representation of words and more effective measurement of word semantic similarity from concept perspective. In this study, we first mine word co-occurrence patterns (i.e., biterms) from short text corpus and then calculate biterm frequency and semantic similarity between its two words. The result shows that a biterm with higher frequency or semantic similarity usually has more similar words in the corpus. Based on the result, we develop a novel probabilistic topic model, named Noise Biterm Topic Model with Word Embeddings (NBTMWE). NBTMWE extends the Biterm Topic Model (BTM) by introducing a noise topic with prior knowledge of frequency and semantic similarity of biterm. NBTMWE shows the following advantages compared with BTM: (1) It can distinguish meaningful latent topics from a noise topic which consists of some common-used words that appear in many texts of the dataset; (2) It can promote a biterm's semantically related words to the same topic during the sampling process via generalized Polya Urn (GPU) model. Using auxiliary word embeddings trained from a large scale of corpus, we report the results testing on two short text datasets (i.e., Sina Weibo and Web Snippets). Quantitatively, NBTMWE outperforms the state-of-the-art models in terms of coherence, topic word similarity and classification accuracy. Qualitatively, each of the topics generated by NBTMWE contains more semantically similar words and shows superior intelligibility.