FactCheck: Financial Event Fact-Checking

• FactCheck is a web-based system that aggregates multilingual financial news and clusters related events using unsupervised techniques.
• It employs a transformer-based fact-checking model, specifically WWM-BERT with adversarial training, to predict merger and acquisition outcomes.
• The interface presents cross-source event clusters with credibility indicators, helping analysts navigate information overload in finance.

FactCheck is a web-based financial news analysis system that combines multilingual news aggregation, unsupervised event clustering, and a transformer-based fact-checking model specialized for financial events. It was proposed in response to the explosion in the sheer magnitude and complexity of financial news data, which makes it increasingly challenging for investment analysts to extract valuable insights and perform analysis. In its reported form, the system aggregates English and Chinese financial news, groups articles about the same event by unsupervised clustering, and exposes a web interface that lets analysts inspect cross-source coverage together with credibility-oriented predictions for merger and acquisition (M&A) outcomes (Yang et al., 2021).

1. Scope and motivating problem

FactCheck in finance is designed for an environment in which hundreds of thousands of financial news articles are published across sources and languages. The paper identifies three operational difficulties for analysts: information overload and redundancy, geographically limited coverage if only a few outlets are followed, and difficulty in obtaining a holistic view of a single event such as an M&A deal or an accounting scandal across multiple news organizations and countries (Yang et al., 2021).

The system also responds to the limits of traditional manual fact-checking in finance. The paper notes that fact-checking in finance is largely manual, as in sites like factcheck.org and Snopes, and that this does not scale with the volume and speed of financial news and is constrained by the expertise and bandwidth of human fact-checkers. FactCheck therefore frames automated event-centric aggregation and credibility-aware analysis as a practical infrastructure for analysts rather than as a replacement for expert judgment (Yang et al., 2021).

A common misconception is that the system performs fine-grained claim verification in the sense of checking individual numeric or textual assertions inside an article. In fact, the reported case study operationalizes fact-checking as a financial event prediction task: predicting whether an M&A deal is completed or not, and using the model’s output probability as a proxy for assessing the credibility and likely outcome of the reported event (Yang et al., 2021).

2. Architecture and multilingual news processing

Architecturally, FactCheck has two main components: a News Aggregator and a Fact Checker. The News Aggregator ingests news from Reuters, Fox, and Sina; translates non-English content into English; and uses an unsupervised event-clustering model to assign event hashtags to each article. The Fact Checker is a transformer-based classifier, specifically WWM-BERT with adversarial training, that takes as input news articles related to a financial event and outputs a prediction of the event outcome (Yang et al., 2021).

The reported data pipeline covers the period from January 1, 2020 – April 22, 2020 and yields a corpus of 183,991 articles, consisting of 116,390 English and 67,601 Chinese items. Reuters and Fox provide English articles, while Sina provides Chinese articles. All non-English articles are translated into English using Baidu’s translation API (https://fanyi-api.baidu.com/), producing a bilingual setting in which each article has an English representation (Yang et al., 2021).

The paper does not detail tokenization or normalization steps explicitly, but it presents the corpus abstraction

$\mathcal{D} = \{d_1, d_2, \dots, d_N\},$

where each $d_i$ is a news article, original or translated into English. This unified English representation allows event clustering to be performed in a single language space while still linking original English reports and translated Chinese reports to the same underlying event (Yang et al., 2021).

The website provides both English and Simplified Chinese interfaces. This bilingual presentation is significant because the aggregation step is not merely multilingual collection; it is intended to produce a cross-lingual, cross-source view of financial events that analysts can inspect directly in the interface (Yang et al., 2021).

3. Event construction and event-centric presentation

FactCheck does not use a supervised event extraction pipeline. Instead, it uses unsupervised event clustering to build events from news articles. The paper adopts the unsupervised news clustering method of Ribeiro et al. (2017), which assigns hashtags corresponding to news events and relies on textual similarity between articles (Yang et al., 2021).

The conceptual pipeline is: represent each article by a feature vector, compute similarity between articles, cluster similar articles into event clusters, and assign hashtags to each cluster. The paper presents this representation abstractly as

$x_i = \phi(d_i),$

with article similarity typically measured by cosine similarity,

$$\text{sim}(x_i, x_j) = \frac{x_i \cdot x_j}{\|x_i\|\, \|x_j\|}.$$

It references the underlying clustering method but does not re-specify the exact clustering algorithm in the FactCheck paper itself (Yang et al., 2021).

Once clustering is performed, each cluster corresponds to a financial event or a tight set of related sub-events, and hashtags serve as event identifiers. In the interface, the event-centric representation is especially important: rows correspond to events, and columns correspond to news sources such as Reuters, Fox, and Sina. A query can be matched using both article content and event hashtags, and the resulting table shows side-by-side coverage of the same event across sources and languages (Yang et al., 2021).

This arrangement gives analysts a cross-lingual, cross-source view of each financial event. Chinese articles translated into English can appear next to English originals, making it easier to compare reporting on a deal or scandal across outlets. The paper’s emphasis is therefore not only on classification accuracy, but also on aligning multilingual reporting around event identity (Yang et al., 2021).

4. Transformer-based fact-checking as financial event prediction

The Fact Checker is implemented as WWM-BERT with adversarial training with Projected Gradient Descent (PGD) applied to the input embeddings. In the case study, the task is binary classification of M&A outcomes, such as completed versus failed deals. The paper explicitly characterizes this as proxy fact-checking: if the classifier predicts that a deal is unlikely to complete, the associated reporting can be flagged as potentially less credible, whereas a high probability of completion supports the credibility of reports asserting that the deal will go through (Yang et al., 2021).

Formally, if $x$ denotes the input text and $(t_1,\dots,t_n)$ its token sequence, the embeddings $(e_1,\dots,e_n)$ are perturbed adversarially: $$\tilde{e}_i = e_i + \delta_i,\quad \|\delta\| \le \epsilon.$$ The transformer encoding is written as

$h = f_\theta(\tilde{e}_1,\dots,\tilde{e}_n),$

and the classifier predicts

$p(y \mid x) = \text{softmax}(W h + b).$

Training uses cross-entropy loss, with adversarial training minimizing the worst-case loss under small input perturbations (Yang et al., 2021).

The model’s specialization is domain-specific rather than knowledge-base-driven. The paper states that the model learns patterns predictive of deal success or failure from a large labeled M&A dataset, including patterns related to regulatory issues, financing problems, and hostile takeovers, and that no additional structured financial knowledge bases or hand-engineered features are described. Specialization is therefore achieved through fine-tuning on financial news, not through explicit symbolic financial knowledge (Yang et al., 2021).

The paper also interprets the predicted probability of success as a credibility or likelihood score: $d_i$0 This score is not introduced as a general theory of credibility. Rather, it is a practical interface between event prediction and analyst-facing credibility assessment in the M&A case study (Yang et al., 2021).

5. M&A dataset, baselines, and empirical results

The quantitative evaluation centers on M&A outcome prediction. The data source is Zephyr, with a time range of January 1, 2007 – August 12, 2019 and 14,539 original M&A deals. After preprocessing, the final classification dataset contains 4,098 instances, each corresponding to a specific M&A deal and its related news, labeled with the deal’s binary outcome (Yang et al., 2021).

The baselines are Random Guessing, CNN-Text, BiGRU, HAN, and BERT-WWM without adversarial training. Evaluation uses MCC, Accuracy (Acc), and F1 score. The paper’s main results are as follows (Yang et al., 2021):

Model	MCC	Acc	F1
Random Guess	0.013	0.510	0.462
CNN-Text	0.729	0.848	0.847
BiGRU	0.734	0.836	0.849
HAN	0.742	0.848	0.853
BERT-WWM	0.751	0.874	0.879
BERT-WWM + adv.	0.788	0.894	0.894

These results show that all deep models substantially outperform random guessing, that BERT-WWM already outperforms CNN-Text, BiGRU, and HAN, and that BERT-WWM + adversarial training (PGD) achieves the best performance. The improvement over non-adversarial BERT-WWM is reported as statistically significant by an independent t-test with p-value = 0.014 (Yang et al., 2021).

The paper does not present a broader ablation study on multilingual aggregation or clustering. Its main quantitative ablation is the comparison between BERT-WWM and BERT-WWM + adv., which it interprets as evidence that adversarial training is a key component of the fact-checker (Yang et al., 2021).

6. Interface, interpretation, and relation to later fact-checking systems

The web front-end is built using Flask and AllenNLP as the backend framework. The typical workflow is: search or browse events using company names, tickers, or keywords; inspect an event grid where rows are clusters and columns are sources; and, for M&A events, view predicted outcomes and softmax confidence scores that act as credibility indicators. Analysts can then drill down into individual articles and compare how different sources describe the same event (Yang et al., 2021).

The system’s limitations are explicit. It includes only three sources and two languages; the fact-checker is evaluated only on M&A outcome prediction; it operates at the level of event outcome rather than fine-grained claim verification; and its translation-first strategy may introduce translation noise and bias. Future directions named in the paper include expanding to more sources and languages, extending to other financial event types such as earnings, fraud, bankruptcy, and macroeconomic events, and developing more sophisticated claim extraction and structured representations (Yang et al., 2021).

Within the broader fact-checking literature, FactCheck occupies a specific position: it is an event-centric, multilingual financial analysis system rather than a general open-domain verifier. Earlier end-to-end systems such as FAKTA integrated document retrieval, stance detection, evidence extraction, and source-type analysis for claim verification (Nadeem et al., 2019), while FaxPlainAC emphasized explainable fact-checking with human correction in the loop and life-long learning support (Zhang et al., 2021). Later work expands other axes of the design space: QACHECK exposes question-guided multi-hop reasoning with explicit question–answer chains (Pan et al., 2023); FactCheck Editor embeds multilingual claim detection, retrieval, NLI-based veracity prediction, and LLM-based correction directly into a text editor and supports over 90 languages (Setty, 2024); OpenFactCheck standardizes modular checking pipelines and benchmarked evaluation of both LLMs and fact-checkers (Iqbal et al., 2024); and LiveFC extends fact-checking to live audio streams with transcription, diarization, retrieval, and NLI-based verification in real time (V et al., 2024).

This suggests a broader trajectory in fact-checking research from source- and claim-centric verification toward systems that are multilingual, interactive, benchmarked, or real-time. FactCheck’s distinctive contribution within that trajectory is narrower but technically specific: it ties multilingual news aggregation, unsupervised event clustering, and transformer-based event-outcome prediction into a single financial workflow, and uses that workflow to give analysts a holistic view of multilingual financial events together with a proxy measure of event credibility (Yang et al., 2021).