A Novel Approach of Neural Topic Modelling for Document Clustering

摘要

Topic modelling is a text mining technique to discover common topics in a collection of documents. The proposed methodology of topic modelling used artificial neural networks to improve the clustering mechanism of similar documents by modelling probabilistic relations between the topics, documents and vocabulary. Currently, while topic modelling and clustering are considered to be manifestations of unsupervised learning, and neural networks on the other hand are used for supervised learning problems, Neural Topic Modelling reformulated topic modelling into a supervised learning task by defining an objective function whose loss function had to be minimized. Custom input embedding layers were designed in order to extract the semantic relationships between the words in the corpus, and the output of the model presented a topic probability distribution for each document. The documents with similar distributions were then bucketed together based on the criteria of meeting the threshold value of a simple distance based similarity metric, such as cosine similarity. The model was implemented using Keras with TensorFlow backend and the effectiveness of the clustering was validated on the IMDB Movie dataset and the News Aggregator dataset from UCI. On comparison with other commonly used clustering mechanisms in combination with traditional topic models, the proposed model delivered an improved Silhouette Co-efficient Score and Davies-Bouldin Index, along with an increased data handling capacity, thereby making the solution scalable.