Domain Adaption of Named Entity Recognition to Support Credit Risk Assessment

Domain Adaption of Named Entity Recognition to Support Credit Risk Assessment Julio Cesar Salinas Alvarado Karin Verspoor Timothy Baldwin Department of Computing and Information Systems The University of Melbourne Australia jsalinas@student.unimelb.edu.au karin.verspoor@unimelb.edu.au tb@ldwin.net Abstract Risk assessment is a crucial activity for financial institutions because it helps them to determine the amount of capi- tal they should hold to assure their sta- bility. Flawed risk assessment models could return erroneous results that trigger a misuse of capital by banks and in the worst case, their collapse. Robust models need large amounts of data to return ac- curate predictions, the source of which is text-based financial documents. Currently, bank staff extract the relevant data by hand, but the task is expensive and time- consuming. This paper explores a ma- chine learning approach for information extraction of credit risk attributes from fi- nancial documents, modelling the task as a named-entity recognition problem. Gener- ally, statistical approaches require labelled data for learn the models, however the an- notation task is expensive and tedious. We propose a solution for domain adaption for NER based on out-of-domain data, cou- pled with a small amount of in-domain data. We also developed a financial NER dataset from publicly-available financial documents. 1 Introduction In the years 2007–2008, the GFC (Global Fi- nancial Crisis) affected a vast number of coun- tries around the world, causing losses of around USD$33 trillion and the collapse of big-name banks (Clarke, 2010). Experts identified that one of the main causes of the GFC was the use of poor financial models in risk assessment (Clarke, 2010; news.com.au, 2010; Debelle, 2009). Risk assessment helps banks to estimate the amount of capital they should keep at hand to pro- mote their stability and at the same time to protect their clients. Poor risk assessment models tend to overestimate the capital required, leading banks to make inefficient use of their capital, or underes- timate the capital required, which could lead to banks collapsing in a financial crisis. Financial documents such as contracts and loan agreements provide the information required to perform the risk assessment. These texts hold rel- evant details that feed into the assessment process, including: the purpose of the agreement, amount of loan, and value of collateral. Figure 1 provides a publicly available example of a loan agreement, as would be used in risk assessment. Currently, bank staff manually extract the in- formation from such financial documents, but the task is expensive and time-consuming for three main reasons: (1) all documents are in unstruc- tured, textual form; (2) the volume of “live” doc- uments is large, numbering in the millions of doc- uments for a large bank; and (3) banks are con- tinuously adding new information to the risk mod- els, meaning that they potentially need to extract new fields from old documents they have previ- ously analyzed. Natural language processing (NLP) potentially offers the means to semi-automatically extract in- formation required for risk assessment, in the form of named entity recognition (NER) over fields of …

Content truncated to 500 words