Interpretable Operational Risk Classification with Semi-Supervised Variational Autoencoder

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 846–852 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 846 Interpretable Operational Risk Classification with Semi-Supervised Variational Autoencoder Fan Zhou1, Shengming Zhang1, Yi Yang2 1University of Electronic Science and Technology of China. 2Hong Kong University of Science and Technology. fan.zhou@uestc.edu.cn, shmizhang@gmail.com, imyiyang@ust.hk Abstract Operational risk management is one of the biggest challenges nowadays faced by finan- cial institutions. There are several major challenges of building a text classification system for automatic operational risk predic- tion, including imbalanced labeled/unlabeled data and lacking interpretability. To tackle these challenges, we present a semi-supervised text classification framework that integrates multi-head attention mechanism with Semi- supervised variational inference for Opera- tional Risk Classification (SemiORC). We em- pirically evaluate the framework on a real- world dataset. The results demonstrate that our method can better utilize unlabeled data and learn visually interpretable document repre- sentations. SemiORC also outperforms other baseline methods on operational risk classifi- cation. 1 Introduction In the decade since the global financial crisis, banks and regulators have become increasingly alert to operational risks (OR). However, the banks still struggle to deal with operational risks effectively (Hoffman, 2002). It is reported that major banks global wide have suffered nearly $210 billion in op- erational risk losses since 2011 1. Operational risks refer to the risks of loss due to errors, breaches, in- terruptions or damages caused by people, internal processes, systems or external events (Coleman, 2010). One of the daily jobs of risk officers is screening potential operational risks from a mas- sive amount of online news outlets. Therefore, there is an urgent need for financial organizations to use artificial intelligence methods for OR predic- tion. While this task can be easily formulated as a classic document classification problem, there are 1https://www.bain.com/insights/ how-banks-can-manage-operational-risk/ at least two challenges in designing such an in- telligent OR prediction system. First, acquiring labels from risk officers is time-costly, and there is no standard labeled dataset for this task. Second, providing explanations is critical for OR predic- tion as risk officers cannot solely rely on prediction outcomes for subsequent decision making. There- fore, these practical issues call for an interpretable semi-supervised text classification framework for OR prediction. However, little prior literature has specifically studied these issues in one framework. To tackle the above-mentioned practical chal- lenges, we propose a semi-supervised text clas- sification model based on the semi-supervised variational autoencoder (SemiVAE) (Kingma et al., 2014) and multi-head attention mechanism (Vaswani et al., 2017) for OR prediction task. Semi- VAE allows effective learning of latent represen- tation from both labeled and unlabeled data, and multi-head attention mechanism produces the di- rect visualization of informative words associated with multi-label predictive outcomes. Learning the model parameters is effective and scalable under the variational inference method. This paper contributes to the burgeoning body of research on using NLP techniques in key finan- cial applications. For example, the prior study leverages the textual features in firm annual re- ports to …

Content truncated to 500 words