Short Text Categorization using World Knowledge

The content of the World Wide Web is drastically multiplying, and thus the amount of available online text data is increasing every day.
Today, many users contribute to this massive global network via online platforms by sharing information in the form of a short text. Such an immense amount of data covers subjects from all the existing domains (e.g., Sports, Economy, Biology, etc.). Further, manually processing such data is beyond human capabilities. As a result, Natural Language Processing (NLP) tasks, which aim to automatically analyze and process natural language documents have gained significant attention. Among these tasks, due to its application in various domains, text categorization has become one of the most fundamental and crucial tasks.

However, the standard text categorization models face major challenges while performing short text categorization, due to the unique characteristics of short texts, i.e., insufficient text length, sparsity, ambiguity, etc. In other words, the conventional approaches provide substandard performance, when they are directly applied to the short text categorization task. Furthermore, in the case of short text, the standard feature extraction techniques such as bag-of-words suffer from limited contextual information. ... mehrHence, it is essential to enhance the text representations with an external knowledge source. Moreover, the traditional models require a significant amount of manually labeled data and obtaining labeled data is a costly and time-consuming task. Therefore, although recently proposed supervised methods, especially, deep neural network approaches have demonstrated notable performance, the requirement of the labeled data remains the main bottleneck of these approaches.

In this thesis, we investigate the main research question of how to perform \textit{short text categorization} effectively \textit{without requiring any labeled data} using knowledge bases as an external source. In this regard, novel short text categorization models, namely, Knowledge-Based Short Text Categorization (KBSTC) and Weakly Supervised Short Text Categorization using World Knowledge (WESSTEC) have been introduced and evaluated in this thesis. The models do not require any hand-labeled data to perform short text categorization, instead, they leverage the semantic similarity between the short texts and the predefined categories. To quantify such semantic similarity, the low dimensional representation of entities and categories have been learned by exploiting a large knowledge base. To achieve that a novel entity and category embedding model has also been proposed in this thesis. The extensive experiments have been conducted to assess the performance of the proposed short text categorization models and the embedding model on several standard benchmark datasets.