Extracting Complex Relations from Banking Documents

Proceedings of the Second Workshop on Economics and Natural Language Processing, pages 1–9 Hong Kong, November 4. c⃝2019 Association for Computational Linguistics 1 Extracting Complex Relations from Banking Documents Berke Oral1,2, Erdem Emekligil1, Sec¸il Arslan1, and G¨uls¸en Eryi˘git2 1R&D and Special Projects, Yapı Kredi Technology 2Department of Computer Engineering, Istanbul Technical University {erdem.emekligil, secil.arslan}@ykteknoloji.com.tr {oralbe, gulsen.cebiroglu}@itu.edu.tr Abstract In order to automate banking processes (e.g. payments, money transfers, foreign trade), we need to extract banking transactions from dif- ferent types of mediums such as faxes, e-mails, and scanners. Banking orders may be consid- ered as complex documents since they con- tain quite complex relations compared to tra- ditional datasets used in relation extraction re- search. In this paper, we present our method to extract intersentential, nested and complex relations from banking orders, and introduce a relation extraction method based on maximal clique factorization technique. We demon- strate 11% error reduction over previous meth- ods. 1 Introduction Despite recent efforts for digitalization in banking domain, formal letters (such as orders, petitions, demands or complaints) still remain as one of the main communication media in corporate banking. A mid-to-large scale bank receives millions of or- ders in a year, most of which are money transfer requests. Reading and manually processing those documents require significant amount of human labour. Moreover, since these documents require specialized knowledge to be interpreted, employ- ment and education of trustable personnel is often difficult. This situation makes the automatization of the process crucial. Automatic processing of banking documents is an Information Extraction (IE) task, where one seeks to extract structured information from semi- structured or unstructured texts. This is usually conducted with pipelined processes. The one used in this study depicted in Figure 1. The first step is to extract entities of interest from raw text, in our case is extracted via Optical Character Recogni- tion (OCR) system. This can be done with Named Entity Recognition (NER) algorithms or pattern OCR Noisy Text Tagged with Named Entities NER Relation Extraction Decoding Noisy Text Relation Graph Transactions Figure 1: Information extraction pipeline. based methods. Secondly, semantic relations be- tween extracted entities are predicted with Rela- tion Extraction (RE) algorithms. Finally, semantic structures are constructed on top of the relations as higher order structures such as templates or events. NER is an important step for many Natural Lan- guage Processing (NLP) tasks. It is a widely re- searched area, and current state-of-the-art algo- rithms (Lample et al., 2016) give satisfactory re- sults in most situations. RE is considered as a linguistically higher order task compared to NER, and proposed algorithms are often tailored to prob- lems in hand. Majority of research in this area is focused on extracting binary relations. Although there are some research dealing with complex re- lations (McDonald et al., 2005; Peng et al., 2017), work on nested complex relations as in the case of banking transactions is scarce. In this paper, we focus on money transfer re- quests as the major process type in banking or- ders. …

Content truncated to 500 words