Paper: arXiv 2504.12319

Authors: Duc Tuyen TA, Wajdi Ben Saad, Ji Young Oh

Abstract

With the introduction of the PSD2 regulation in the EU which established the Open Banking framework, a new window of opportunities has opened for banks and fintechs to explore and enrich Bank transaction descriptions with the aim of building a better understanding of customer behavior, while using this understanding to prevent fraud, reduce risks and offer more competitive and tailored services. And although the usage of natural language processing models and techniques has seen an incredible progress in various applications and domains over the past few years, custom applications based on domain-specific text corpus remain unaddressed especially in the banking sector. In this paper, we introduce a language-based Open Banking transaction classification system with a focus on the french market and french language text. The system encompasses data collection, labeling, preprocessing, modeling, and evaluation stages. Unlike previous studies that focus on general classification approaches, this system is specifically tailored to address the challenges posed by training a language model with a specialized text corpus (Banking data in the French context). By incorporating language-specific techniques and domain knowledge, the proposed system demonstrates enhanced performance and efficiency compared to generic approaches.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s mathematics is primarily applied machine learning (text preprocessing, vectorization, classifiers) without advanced theoretical derivations, keeping complexity moderate. It demonstrates strong empirical rigor through a concrete system design with data processing steps, specific dataset description, and performance evaluation on real banking data, making it backtest-ready.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Specialized French Open Banking<br>Transaction Classification"] --> B["Data Collection<br>French Banking Transactions"]
  B --> C["Preprocessing & Labeling<br>Domain-specific cleaning &<br>Manual annotation"]
  C --> D["Modeling<br>Custom NLP vs. Generic Models"]
  D --> E["Evaluation<br>Performance Metrics<br>(Accuracy, F1-Score)"]
  E --> F["Key Findings<br>Custom system outperforms<br>generic approaches in<br>French banking context"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a language‑based Open Banking transaction classification system tailored to French banking text.

Source sentence
In this paper, we introduce a language-based Open Banking transaction classification system with a focus on the french market and french language text.

Assumes 5

  • The transaction descriptions are in French and can be processed with French‑specific language techniques. data
    Source sentence
    focus on the french market and french language text
  • A multi‑class classifier can replace pairwise binary classifiers and scale to many categories. model
    Source sentence
    using a multi-class classifier is a more practical alternative
  • Weighted precision, recall and F1 are appropriate evaluation metrics for the heavily imbalanced dataset. statistical
    Source sentence
    we employed precision, recall, and F1 metrics using a weighted-average approach
  • TF‑IDF vectorisation and cosine similarity are sufficient to detect similar transaction descriptions at scale. model
    Source sentence
    transforming the transaction descriptions into vector form using Term Frequency-Inverse Document Frequency (TF-IDF) as terms to identify similar strings
Show 1 more
  • The training and test sets (94,356 and 109,509 transactions) are representative of French Open Banking data. data
    Source sentence
    The data consists of a training set with 94, 356 transactions and a test set with 109, 509 transactions, both from customer accounts of major French banks

Data

  • Open Banking transaction dataset (French customers) text-news proprietary
    December 2021 to January 2023 · major French banks customers
    Source sentence
    The data consists of a training set with 94, 356 transactions and a test set with 109, 509 transactions, both from customer accounts of major French banks, written mostly in French, from December 2021 to January 2023.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Linear SVM achieves weighted precision, recall and F1 of 90% with 50% of the training data and 94% with 80% of the training data using only transaction descriptions. performance
    Source sentence
    we observe with only transaction description, the system with Linear SVM model can get the (weighted) precision, recall, and F1 of 90% with only 50% of the training set and 94% with 80% of the training set.
  • Random Forest with Word2Vec embeddings reaches weighted precision, recall and F1 of 95% on the test set. performance
    Source sentence
    the training model with Random Forest can get the (weighted) precision, recall, and F1 of 95% in the tests set.
  • Category‑specific performance is high overall, with the worst results for categories having few training examples (e.g., ADVANCE SALARY F1 83%). performance
    Source sentence
    The worst performance corresponds to the categories with fewer entries in the training set. In the case of ”GROCERIES”, the classification highly relied on the merchant name which explains the relatively low performance.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (17)

  1. Deep learning enhancing banking services: a hybrid transaction classification and cash flow prediction approach (2022)
  2. Text document categorization by term association (2002)
  3. Keyword extraction from short texts with a text-to-text transfer transformer (2022)
  4. Theory-driven analysis of large corpora: Semisupervised topic classification of the UN speeches (2022)
  5. Simplified Graph Learning for Inductive Short Text Classification (2022)
  6. Transformers are Short Text Classifiers: A Study of Inductive Short Text Classifiers on Benchmarks and Real-world Datasets (2022) — link
  7. HGAT: Heterogeneous graph attention networks for semi-supervised short text classification (2021)
  8. Machine learning for financial transaction classification across companies using character-level word embeddings of text fields (2021)
and 9 more
  1. Document and word representations generated by graph convolutional network and Bert for short text classification (2020)
  2. Identifying banking transaction descriptions via support vector machine short-text classification based on a specialized labelled corpus (2020)
  3. Using topic modeling methods for short-text data: A comparative analysis (2020)
  4. Short text topic modeling techniques, applications, and performance: a survey (2020)
  5. Deeptrax: Embedding graphs of financial transactions (2019)
  6. Short text classification using graph convolutional network (2019)
  7. Continuous embedding spaces for bank transaction data (2017)
  8. Automatic classification of bank transactions (2017)
  9. Word2Vec (2017)

0 of 17 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.