Abstract: Aimed at the difficulties of semantic representation of long news text,an enhanced document feature representation is constructed based on Doc2Vec embedding and word vector weighting.Enhanced features from the specific parts-of-speech contents on the head and tail of the document are extracted by the method of DV-sim or DV-tfidf.These features are then combined with doc2vec to form a new global representation.DV-sim uses the similarity between feature words and doc2vec vectors to obtain word weight from the semantic point of view,and DV-tfidf uses term frequency inverse document frequency to obtain word weight from the word frequency statistics point of view.Then the HDBSCAN algorithm is applied to cluster topics on the Thucnews and Sogou datasets.Compared with the Doc2Vec vector,the noise number on the two datasets reduces by 60.82% and 60.63%,the accuracy improves by 12.14% and 20.58%,and the F1-score increases by 15.61% and 11.58%,respectively,with DV-sim.The noise number on the two datasets reduces by 15.20% and 59.55%,the accuracy improves by 10.85% and 17.93%,and the F1-score increases by 15.60% and 9.21%,respectively,with DV-tfidf.Experiments show that both DV-sim and DV-tfidf can improve the performance of topic clustering,and the enhancement feature based on semantics is better than that based on word frequency.DV-sim has also been effectively applied in topic clustering of excellent female character reports.

Key words: Topic clustering, Text representation, Doc2Vec, Word embedding, HDBSCAN