Incorporating Prior Knowledge by Selective Context Features to Enhance Topic Coherence

Abstract

Latent variable model has been widely used to extract topics from document collections, but this unsupervised mode makes it difficult to interpret the topics covered by the constructed cluster due to no artificial tags. Considering the influence of preprocessing on subsequent data mining, this study extracts news feature words from CSCP word segmentation method and named entity identification (NER). For NER, names of people, places, organizations are extracted by syntax rules and verified by Wikipedia. In terms of CSCP, the probability of continuing characters, the distribution probabilities and numbers of links before and after the character are used to constantly merge the unit words, extract the important narrative words effectively, and then conduct Unigram, Compounds and Mixture word processing. The corpus of words obtained after word processing by CSCP and NER were used as LDA prior knowledge. Finally, the topic coherence of NER and CSCP is evaluated by UMass Topic Coherence Measurement. The experimental results show that Compounds are of specific meaning; Mixtures represents its diverse scope; Unigrams and NER are relatively short, while NER can accurately represent the important features of news content, the topic is more cohesive. In terms of efficiency, NER-LDA takes the longest, while it had the highest degree of topic coherence.