FiNER: Financial Numeric Entity Recognition for XBRL Tagging

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4419 - 4431 May 22-27, 2022 c⃝2022 Association for Computational Linguistics FiNER: Financial Numeric Entity Recognition for XBRL Tagging Lefteris Loukas1,2, Manos Fergadiotis1,2, Ilias Chalkidis3, Eirini Spyropoulou1, Prodromos Malakasiotis1,2, Ion Androutsopoulos1,2, George Paliouras1 1Institute of Informatics and Telecommunications, NCSR “Demokritos” 2Department of Informatics, Athens University of Economics and Business 3Department of Computer Science, University of Copenhagen Abstract Publicly traded companies are required to sub- mit periodic reports with eXtensive Business Reporting Language (xbrl) word-level tags. Manually tagging the reports is tedious and costly. We, therefore, introduce xbrl tagging as a new entity extraction task for the finan- cial domain and release finer-139, a dataset of 1.1M sentences with gold xbrl tags. Un- like typical entity extraction datasets, finer- 139 uses a much larger label set of 139 en- tity types. Most annotated tokens are nu- meric, with the correct tag per token depend- ing mostly on context, rather than the token itself. We show that subword fragmentation of numeric expressions harms bert’s perfor- mance, allowing word-level bilstms to per- form better. To improve bert’s performance, we propose two simple and effective solutions that replace numeric expressions with pseudo- tokens reflecting original token shapes and nu- meric magnitudes. We also experiment with fin-bert, an existing bert model for the fi- nancial domain, and release our own bert (sec-bert), pre-trained on financial filings, which performs best. Through data and er- ror analysis, we finally identify possible limita- tions to inspire future work on xbrl tagging. 1 Introduction Natural language processing (nlp) for finance is an emerging research area (Hahn et al., 2019; Chen et al., 2020; El-Haj et al., 2020). Financial data are mostly reported in tables,but substantial information can also be found in textual form, e.g., in company filings, analyst reports, and economic news. Such information is useful in numerous fi- nancial intelligence tasks, like stock market predic- tion (Chen et al., 2019; Yang et al., 2019), financial sentiment analysis (Malo et al., 2014; Wang et al., Source code: https://github.com/nlpaueb/finer Correspondence: eleftheriosloukas@aueb.gr Figure 1: Sentences from finer-139, with xbrl tags on numeric and non-numeric tokens. xbrl tags are ac- tually xml-based and most tagged tokens are numeric. 2013; Akhtar et al., 2017), economic event detec- tion (Ein-Dor et al., 2019; Jacobs et al., 2018; Zhai and Zhang, 2019), and causality analysis (Tabari et al., 2018; Izumi and Sakaji, 2019). In this work, we study how financial reports can be automatically enriched with word-level tags from the eXtensive Business Reporting Language (xbrl), a tedious and costly task not considered so far.1 To promote transparency among shareholders and potential investors, publicly traded companies are required to file periodic financial reports. These comprise multiple sections, including financial ta- bles and text paragraphs, called text notes. In ad- dition, legislation in the us, the uk, the eu and elsewhere requires the reports to be annotated with tags of xbrl, an xml-based language, to facili- tate the processing of …

Content truncated to 500 words