[Li, T; Li, Tao; Wei, Yingying] Hunan Univ Sci & Technol, Sch Comp Sci & Engn, Xiangtan 411201, Hunan, Peoples R China.

[Wang, Ze; Li, Jianbin] Cent South Univ Forestry & Technol, Sch Comp & Informat Engn, Changsha 410004, Hunan, Peoples R China.

[Li, T ] H

Hunan Univ Sci & Technol, Sch Comp Sci & Engn, Xiangtan 411201, Hunan, Peoples R China.

Short texts typically refer to texts of relatively short length;generally not exceeding 160 characters. Short text classification (Chakraborty & Singh;2022) is a form of text mining technology and a significant research direction in natural language processing (NLP) (Hirschberg & Manning;2015). With the rapid development of the internet;a large volume of information in the form of short texts such as news tags;reviews;instant messages;and tweets has emerged on online platforms. These short texts contain a wealth of helpful information that contributes to establishing services in various fields;such as news classification;social media (Kateb & Kalita;sentiment analysis (Balomenos et al.;e-commerce;and spam filtering. As a result;with the growing depth of related research;short text classification has emerged as a critical challenge that urgently needs to be addressed in information retrieval. However;short text classification faces several challenges that complicate the task. First;short texts tend to be brief;with sparse semantic features and weak contextual connections;often requiring additional feature information to assist in classification (Wu;2023). For example;Hua et al. (2024) integrated text;entities;and words into a heterogeneous graph convolutional network;representing features through a word graph and enhancing word features with BiLSTM;thereby enabling document category prediction. However;due to length limitations;GNNs typically do not incorporate additional information when processing short texts;treating each short text as a single node in the graph. This leads to insufficient feature information;which in turn affects performance. Therefore;enhancing feature information and semantic relevance becomes crucial. Second;in rapidly changing text data environments (such as social media and news);expert knowledge required to organize and label training documents properly is often limited;expensive;and sometimes even unavailable. In some cases;training text classifiers using well-organized and accurately labeled documents is prohibitively costly;making unsupervised learning a crucial solution. Unsupervised tasks allow for fast classification without labeled data;avoiding the time-consuming and error-prone process of manual annotation. For example;Cameron-Steinke (2019) employed a novel text expansion approach in an unsupervised manner;whereby new words are generated directly for each input sentence. The newly generated keywords are formed within the hidden states of a pre-trained language model;which are then used to create expanded pseudo-headlines. However;it has limitations in terms of the accuracy and redundancy of the expanded content. In addition;Mishra (2024) compared and analyzed unsupervised text classification techniques;especially Latent Dirichlet Allocation (LDA) and BERTopic;and applied these topic modeling techniques to very short documents. However;since their evaluation datasets are based on specific scenarios;the generalization of these methods is weak. Therefore;how to effectively classify large amounts of unlabeled data using unsupervised methods is a research-worthy problem. To address the aforementioned issues;we propose a novel approach called UEDE: The Unsupervised Short Text Classification Method based on GCN Encoder-Decoder and Local Enhancement. This method uses unsupervised learning to cleverly integrate the GCN encoder–decoder and local enhancement technology under the autoencoder framework to achieve accurate short-text classification. Specifically;we first design the MRFasGCN modeling approach within the autoencoder framework;learning the hierarchical representations of node attributes and the network’s deep structure. We model and infer the relationships between nodes by integrating these deep representations with network structural information;thereby capturing text features and semantic information more effectively. Then;we introduce a dual decoder suited for unsupervised learning;which alleviates the issue of short text feature sparsity from multiple perspectives by reconstructing the topology and node attributes. Furthermore;we further enhance the latent semantic module locally by constructing an information fusion graph;ensuring that each pair of connected nodes shares a similar module distribution and complementary features. This improves the model’s ability to understand the local context of the text while addressing the overfitting issues caused by short text feature sparsity;significantly boosting the model’s performance and classification accuracy. The main contributions of this article can be summarized as follows: • We propose an unsupervised learning-based short text classification method;utilizing multiple convolutional layers as encoders to learn short text attribute networks’ deep representations and structural information. This approach facilitates module partitioning and inference;more effectively capturing text features and semantic information. • We design a dual decoder structure that;through unsupervised learning;separately reconstructs the topology and node attributes while accounting for the heterogeneity of node degrees. This approach compensates for insufficient short text feature information from multiple perspectives. • We introduce a local enhancement method that enables nodes to share more common neighbors and similar node attributes;strengthening the connections between related nodes. This improves the model’s understanding of the text’s local context while alleviating the overfitting issues caused by short text feature sparsity. • We perform comprehensive experiments on real-world datasets encompassing news articles;concise comments;and search snippets to assess the efficacy of our model in comparison to fourteen baseline approaches. The experimental findings unequivocally establish that our model surpasses the current state-of-the-art baseline methods on the benchmark datasets. We propose an unsupervised learning-based short text classification method;utilizing multiple convolutional layers as encoders to learn short text attribute networks’ deep representations and structural information. This approach facilitates module partitioning and inference;more effectively capturing text features and semantic information. We design a dual decoder structure that;through unsupervised learning;separately reconstructs the topology and node attributes while accounting for the heterogeneity of node degrees. This approach compensates for insufficient short text feature information from multiple perspectives. We introduce a local enhancement method that enables nodes to share more common neighbors and similar node attributes;strengthening the connections between related nodes. This improves the model’s understanding of the text’s local context while alleviating the overfitting issues caused by short text feature sparsity. We perform comprehensive experiments on real-world datasets encompassing news articles;concise comments;and search snippets to assess the efficacy of our model in comparison to fourteen baseline approaches. The experimental findings unequivocally establish that our model surpasses the current state-of-the-art baseline methods on the benchmark datasets. The remainder of the paper is structured as follows: Section 2 reviews previous work. Section 3 describes our proposed method and model;including the unsupervised short text classification model and local enhancement method. In Section 4;we perform comprehensive experiments on the datasets and analyze the outcomes. Lastly;Section 5 concludes the paper;offering insights into the future research directions.

计算机与信息工程学院