Short text topic modelling using local and global word-context semantic correlation.

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Short texts from social media and e-commerce lack contextual information, posing challenges for traditional topic modeling.
• Existing methods like Latent Dirichlet Allocation struggle with the sparsity and noise inherent in short documents.

Purpose of the Study:

• To propose a new topic modeling model, G_SeaNMF (Gensim_SeaNMF), for short texts.
• To enhance word-context semantic relationships by integrating local and global word embeddings.

Main Methods:

• Incorporated the Semantics-assisted Non-negative Matrix Factorization (SeaNMF) with the Gensim word2vec model.
• Explored topic modeling techniques including Dirichlet Multinomial Mixture (DMM), self-aggregation, and global word co-occurrence.
• Evaluated model performance on diverse real-world datasets (Search Snippet, Biomedicine, Pascal Flickr, Tweet, TagMyNews).

Main Results:

• The G_SeaNMF model effectively improves the semantic relationship between words and their contexts in short texts.
• Empirical evaluations demonstrated superior cluster coherence compared to other methods.
• Combining local and global word embeddings yielded more relevant words per topic.

Conclusions:

• G_SeaNMF offers a robust solution for topic modeling in sparse, short-text environments.
• The integration of word embeddings significantly enhances the quality of topic discovery.
• This approach is beneficial for analyzing large volumes of user-generated content.