Abstract: The special features of microblog text cause difficulties in building microblog user interest portrait. To address the problem, an ensemble algorithm based on NWD-Bi-LSTM-XGBoost is proposed. Firstly, a new word discovery algorithm from the perspective of support is raised to deal with the informality of microblog text, exploring the ubiquitous internet phrases and achieving more accurate word segmentation and semantic understanding. Then, a Simhash algorithm is introduced to mitigate the information overload of microblog text. To improve the feature sparsity caused by microblog text’s conciseness, bidirectional long short-term memory networks are used to extract semantic features. Finally, the XGBoost model is trained by combining the static features of microblog users with the semantic features of the blog text for constructing the multi-granularity microblog user interest portrait efficiently. The experimental results show that the macro-average F1 score and AUC value of coarse-granularity (primary) interest tag model are up to 83.6% and 79.7% and that of fine-granularity (secondary) interest tag model are 70.4% and 63.6%, respectively. Compared with other benchmark models, the macro-average F1 score and AUC value of the models increase by 3%~5% due to ensemble of the NWD algorithm, which is superior to the existing new word discovery methods.

Key words: new word discovery, bidirectional long short-term memory, extreme Gradient Boosting, multi-granularity, microblog user interest portrait