An Investigation Into Different Text Representations to Train an Artificial Immune Network for Clustering Texts

Abstract

Extracting knowledge from text data is a complex task that is usually performed by first structuring the texts and then applying machine learning algorithms, or by using specific deep architectures capable of dealing directly with the raw text data. The traditional approach to structure texts is called Bag of Words (BoW) and consists of transforming each word in a document into a dimension (variable) in the structured data. Another approach uses grammatical classes to categorize the words and, thus, limit the dimension of the structured data to the number of grammatical categories. Another form of structuring text data for analysis is by using a distributed representation of words, sentences, or documents with methods like Word2Vec, Doc2Vec, and SBERT. This paper investigates four classes of text structuring methods to prepare documents for being clustered by an artificial immune system called aiNet. The goal is to assess the influence of each structuring method in the quality of the clustering obtained by the system and how methods that belong to the same type of representation differ from each other, for example both LIWC and MRC are considered grammarbased models but each one of them uses completely different dictionaries to generate its representation. By using internal clustering measures, our results showed that vector space models, on average, presented the best results for the datasets chosen, followed closely by the state of the art SBERT model, and MRC had the overall worst performance. We could also observe a consistency in the number of clusters generated by each representation and for each dataset, having SBERT as the model that presented a number of clusters closer to the original number of classes in the data.