Few-Shot Text Classification with Edge-Labeling Graph Neural Network-Based Prototypical Network

Proceedings of the 28th International Conference on Computational Linguistics, pages 5547–5552 Barcelona, Spain (Online), December 8-13, 2020 5547 Few-Shot Text Classification with Edge-Labeling Graph Neural Network-Based Prototypical Network Chen Lyu Peking University chenlyu@pku.edu.cn Weijie Liu Peking University dataliu@pku.edu.cn Ping Wang Peking University pwang@pku.edu.cn Abstract In this paper, we propose a new few-shot text classification method. Compared with supervised learning methods which require a large corpus of labeled documents, our method aims to make it possible to classify unlabeled text with few labeled data. To achieve this goal, we take advantage of advanced pre-trained language model to extract the semantic features of each document. Fur- thermore, we utilize an edge-labeling graph neural network to implicitly models the intra-cluster similarity and the inter-cluster dissimilarity of the documents. Finally, we take the results of the graph neural network as the input of a prototypical network to classify the unlabeled texts. We verify the effectiveness of our method on a sentiment analysis dataset and a relation classification dataset and achieve the state-of-the-art performance on both tasks. 1 Introduction In recent years, deep learning has achieved great success in many fields such as computer vision, speech recognition and natural language processing. However, to train a deep learning model usually requires a large scale of manually annotated data, which greatly limits the practicality and scalability of the model. In the text classification task, the problem of greed for large amounts of labeled data also exists. Text classification is an important task in natural language processing field. Over the years, many text classification methods have been proposed, varying from CNN based methods (Kim, 2014; Johnson and Zhang, 2015; Zhang et al., 2015) to RNN, including LSTM and GRU-based methods (Zhou et al., 2015; Yang et al., 2016). But all the above supervised learning methods require a large corpus of labeled data, making these models hindered in practical application. In recent years, some scholars have contributed to solving the few-shot text classification problem. Han (2018) proposed a dataset FewRel and tested the performance of several typical few-shot learning models proposed in recent years on this dataset. Yu (2018) proposed an adaptive metric learning model, which can automatically determine the best weighted combination of a set of metrics obtained from a meta-learning process for a newly arrived few-shot text classification task. Gao (2019) proposed a pro- totypical network model (Snell et al., 2017) based on a hybrid attention mechanism, which has a better performance for noisy data scenes. Sun (2019) improved prototypical network by adopting hierarchical attention mechanism, which is applied in feature level, word level and instance level to enhance the ex- pressive ability of semantic space. Geng (2019) applied the dynamic routing algorithm in meta-learning and proposed an induction network, which achieves a better generalization ability on different few-shot text classification tasks. We believe that previous few-shot text classification methods generally have two flaws. First, all the methods mentioned above adopt a GloVe word embedding combined with CNN or RNN structure for text embedding instead of more …

Content truncated to 500 words