Work

Learning Label Relationships Using Deep Neural Networks for Hierarchical Text Classification

Year: 2023
Type: dissertation
Abstract: Hierarchical text classification has been receiving increasing attention due to its vast range of applications in real-world natural language processing tasks. With the recent advances in deep learning, deep learning-based approaches achieved state-of-the-art hierarchical text classification performance. While existing approaches focus on exploiting the label hierarchy or modeling implicit label relationships, only a few studies integrated these two concepts. This thesis proposes a graph attention capsule network for hierarchical text classification (GACaps-HTC), a deep learning-based approach designed to capture both the explicit hierarchy and latent label relationships. A graph attention network is employed in the proposed approach for fusing information on the label hierarchy into a textual representation, while a capsule network is employed to understand the latent label relationships and infer classification probabilities. The proposed approach is optimized using a loss term designed to address the innate label imbalance issue of the task and post-processed using various methods specified for hierarchical text classification. Results of the experiments conducted on two benchmark datasets demonstrate that the proposed approach outperformed previous state-of-the-art approaches and ablation studies show that each component in the GACaps-HTC played a part in enhancing the performance. \n \nFurthermore, this thesis proposes a semantic-aware dynamic routing algorithm, a new dynamic routing algorithm that initializes and updates a capsule network's coupling coefficients using semantic representations of labels. As a coupling coefficient of a pair of capsules indicates how similar their information is, the coefficient is initialized from the similarity of semantic representations corresponding to the capsules' labels. Experiment results show that the proposed algorithm outperformed other methods that inject semantic information of labels and GACaps-HTC with semantic-aware dynamic routing algorithm reached faster convergence compared to GACaps-HTC with conventional dynamic routing algorithm. \n \nFinally, this thesis investigates another use case of GACaps-HTC by employing the model for aspect category sentiment analysis that can be formulated as hierarchical text classification. Experiments were conducted on four sentiment analysis datasets, and the results show that the proposed approach performs well on not only the semantic analysis of a document but also sentiment analysis.
Author: 방진현
Language: Korean
Open Access status: green
OpenAlex ID: W6983053880
Cites: 142
Cited by:
Related to: 10