FiNER: Financial Numeric Entity Recognition for XBRL Tagging

TL;DR

FiNER-139 introduces a real-world XBRL tagging task and releases a large corpus of 1.1M sentences annotated with 139 numeric-focused XBRL tags. The study analyzes baseline architectures and reveals that heavy numeric subword fragmentation hurts BERT, motivating simple pseudo-token replacements for numbers ([num], [shape]) and domain-specific pretraining. Domain-adapted models, particularly SEC-BERT with numeric tokens, achieve the best performance, confirming the value of targeted pretraining for financial text. The work provides practical guidance for automated XBRL tagging and highlights future directions, including hierarchical classification and annotation quality controls.

Abstract

Publicly traded companies are required to submit 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 financial domain and release FiNER-139, a dataset of 1.1M sentences with gold XBRL tags. Unlike typical entity extraction datasets, FiNER-139 uses a much larger label set of 139 entity types. Most annotated tokens are numeric, with the correct tag per token depending mostly on context, rather than the token itself. We show that subword fragmentation of numeric expressions harms BERT's performance, allowing word-level BILSTMs to perform 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 numeric magnitudes. We also experiment with FIN-BERT, an existing BERT model for the financial domain, and release our own BERT (SEC-BERT), pre-trained on financial filings, which performs best. Through data and error analysis, we finally identify possible limitations to inspire future work on XBRL tagging.

Figures (5)

• Figure 1: Sentences from 1.25fi1.25ner-139, with 1.25xbrl tags on numeric and non-numeric tokens. 1.25xbrl tags are actually 1.25xml-based and most tagged tokens are numeric.
• Figure 2: Frequency distribution of the 139 1.25xbrl tags used in this work over the entire 1.25fi1.25ner-139 dataset. Label indices shown instead of tag names to save space.
• Figure 3: 1.25xbrl tag predictions of 1.25bert (top), 1.25bert + 1.25 [num ] (middle), 1.25bert + 1.25 [shape ] (bottom) for the same sentence. 1.25bert tags incorrectly the amounts in red. 1.25bert + 1.25 [num ] and 1.25bert + 1.25 [shape ] tag them more successfully (green indicates correct tags).
• Figure 4: $\mathrm{Hits@}k$ results (%, avg. of 3 runs with different random seeds) on test data, for different $k$ values. Standard deviations were very small and are omitted.
• Figure 5: A manually inspected sentence from 1.25fi1.25ner-139 showing some inconsistencies in the gold 1.25xbrl tags of the auditors. The green '1' is correctly annotated with the 1.25xbrl tag Lessee Operating Lease Term Of Contract. The red '16' should have also been annotated with the same tag, but is not, possibly because the annotator thought the (same) tag was obvious. The orange numbers '0.1' and '6' lack 1.25xbrl annotations; they should have both been annotated as Lessee Operating Lease Renewal Term. We can only speculate that the auditor might not have been aware that there is an 1.25xbrl tag for lease renewal terms, in which case the recommendation engine of Section \ref{['sec:business_study']} might have helped.