Classification-Aware Neural Topic Model Combined with Interpretable Analysis - for Conflict Classification

Proceedings of Recent Advances in Natural Language Processing, pages 666–672 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_072 666 Classification-Aware Neural Topic Model Combined With Interpretable Analysis - For Conflict Classification Tianyu Liang1, Yida Mu 1, Soonho Kim2, Darline Larissa Kengne Kuate2, Julie Lang2, Rob Vos2, Xingyi Song1 1Department of Computer Science, University of Sheffield, Sheffield, United Kingdom 2International Food Policy Research Institute, Washington, DC, USA {tliang9, y.mu, x.song}@sheffield.ac.uk Abstract A large number of conflict events are affect- ing the world all the time. In order to anal- yse such conflict events effectively, this paper presents a Classification-Aware Neural Topic Model (CANTM-IA) for Conflict Information Classification and Topic Discovery. The model provides a reliable interpretation of classifica- tion results and discovered topics by introduc- ing interpretability analysis. At the same time, interpretation is introduced into the model ar- chitecture to improve the classification perfor- mance of the model and to allow interpretation to focus further on the details of the data. Fi- nally, the model architecture is optimised to reduce the complexity of the model. 1 Introduction Hundreds of conflicts break out every day around the world, many of which have a major impact on the world’s political and economic situation. A recent example is Ukraine Crisis, which has caused energy scarcity in Europe, a reduction in world food production and many other repercussions. For governments and institutions such as the IFPRI, the impact of conflict events can be greatly reduced if they are classified, analysed and responded to in the shortest possible time. Our goal is to develop a deep learning model suitable for the classification of conflict informa- tion. This model should be able to classify con- flict categories and discover category-related top- ics. Most importantly, the model must have high reliability as the consequences of conflicting infor- mation are often very serious. We therefore want to combine text classification, topic modelling and interpretable analysis to solve the problem. Text classification assigns category labels to dif- ferent texts for the purpose of distinguishing textual information. Recurrent neural networks (RNNs), convolutional neural networks (CNNs) and graph neural networks (GNNs) have all been applied to text classification tasks (Tai et al., 2015; Zhu et al., 2015; Cheng et al., 2016; Kalchbrenner et al., 2014; Kim, 2014; Johnson and Zhang, 2015; Peng et al., 2018). More recently, Sun et al. (2019) provides a fine-tuned BERT-based pre-training model (De- vlin et al., 2019) for text classification tasks generic solution with new state-of-the-art results on eight extensively studied text classification datasets. The topic model is designed to automatically find a range of topics and topic words from a col- lection of documents. One of the most classic topic models is latent dirichlet allocation (LDA) (Blei et al., 2003), which is an unsupervised, non- hierarchical model. Many subsequent research has been based on LDA, such as the Hierarchical Latent Dirichlet Allocation (HLDA) proposed by Griffiths et al. (2003). In 2016, Miao et al. (2016) proposed a generative neural variational document model (NVDM), which models the likelihood of …

Content truncated to 500 words