학술논문
UA-HGAT: Uncertainty-aware Heterogeneous Graph Attention Network for Short Text Classification
Document Type
Conference
Author
Source
2022 IEEE International Conferences on Internet of Things (iThings) and IEEE Green Computing & Communications (GreenCom) and IEEE Cyber, Physical & Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on Cybermatics (Cybermatics) ITHINGS-GREENCOM-CPSCOM-SMARTDATA-CYBERMATICS Internet of Things (iThings) and IEEE Green Computing & Communications (GreenCom) and IEEE Cyber, Physical & Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on Cybermatics (Cybermatics), 2022 IEEE International Conferences. :495-500 Aug, 2022
Subject
Language
Abstract
During the last decade, deep learning has played a significant role in improving text classification which is widely applied in question answering systems, comment extraction, sentiment analysis, etc. Considering that few labeled training data are available, semi-supervised short text classification needs to be studied urgently. Existing works focus on long texts and cannot achieve satisfactory performance on short texts due to the sparsity, non-standardization, and insufficient labeled data. In this paper, we propose a novel uncertainty-aware heterogeneous graph attention network for semi-supervised short text classification (UA-HGAT), which makes full use of limited labeled data and large amounts of unlabeled data via exploiting an uncertainty-aware mechanism. In the model, entity description from Wikipedia is introduced to empower short texts, and word-level information is acquired by Word2Vec. The entity description and word-level information are combined with topic information to address the issue of short text sparseness. In particular, we introduce uncertainty-aware mechanism and negative sample learning to relieve the erroneous high confidence predictions from poorly calibrated models, so that it can offer enough pseudo-label samples for the model training. Extensive experiments show that our model significantly outperforms the SOTA methods on AGNews and Snippets datasets.