CLFEC: A New Task for Unified Linguistic and Factual Error Correction in paragraph-level Chinese Professional Writing

1. Problem Definition

Current Chinese text correction research primarily focuses on Grammatical Error Correction (GEC) and Chinese Spelling Correction (CSC), treating these as purely linguistic tasks. Conversely, Factual Error Correction (FEC) is typically studied in isolation, focusing on claim extraction and evidence retrieval.

However, in paragraph-level professional writing (e.g., finance, law, medicine, current affairs), linguistic errors (typos, grammar, punctuation) and factual errors (incorrect entities, outdated regulations, wrong numbers) frequently co-occur and interact. Existing datasets and models fail to address this complexity because:

• Separation of Concerns: Decoupled pipelines often miss deep inconsistencies when surface-level errors are present.
• Data Bias: Existing "gold" correction datasets often contain residual factual errors, biasing model training.
• Over-correction: Large Language Models (LLMs) tend to "over-correct" clean text, altering style or preference-based wording under the guise of error correction.

The paper introduces CLFEC (Chinese Linguistic & Factual Error Correction), a unified task designed to jointly correct linguistic and factual errors within a single context.

2. Methodology

A. Dataset Construction (CLFEC)

The authors constructed a new, mixed, multi-domain dataset from scratch to avoid contamination from existing corpora.

• Domains: Current Affairs, Finance, Law, and Medicine.
• Composition: 925 paragraphs with an average length of ~466 characters.
• Error Injection: Errors were injected using LLMs followed by rigorous manual verification.
  • Linguistic Errors (LEC): Word (spelling/omission/redundancy), Grammar (structure/collocation), and Punctuation.
  • Factual Errors (FEC): Entity substitution, temporal/regulatory mismatches, and numerical inconsistencies.
• Splits: The dataset is divided into four diagnostic subsets: MIX (both LEC and FEC), LEC-only, FEC-only, and Error-free (to test over-correction).
• Annotation: Follows the CHERRANT paradigm, representing corrections as edit operations (start/end indices, original span, corrected span).

B. Correction Paradigms Evaluated

The study systematically evaluates three LLM-based approaches:

1. Prompting (Baseline): Direct instruction to the LLM to correct errors.
2. Retrieval-Augmented Generation (RAG):
   • S-RAG (Sequential): A two-stage pipeline. Stage 1 corrects linguistic errors; Stage 2 uses RAG to correct factual errors.
   • U-RAG (Unified): A single-pass approach where the LLM handles both LEC and FEC simultaneously using retrieved evidence.
3. Agentic Framework: A ReAct-based workflow where an agent plans, executes, and verifies corrections.
   • Plan-and-Execute: The agent decomposes the document into a "to-do list" of localized tasks (e.g., verify a specific claim, fix a specific sentence).
   • Tools: Uses todo_write for state management, search_tool for evidence retrieval, and verify_tool to ensure edits are anchored to the source text to prevent hallucinations.

3. Key Contributions

1. New Task Definition: Formalized CLFEC as a unified task for paragraph-level professional writing, bridging the gap between linguistic and factual correction.
2. High-Quality Dataset: Released a multi-domain, mixed-error dataset with diagnostic splits (MIX, LEC, FEC, Error-free) specifically designed to test generalization and over-correction.
3. Empirical Analysis of Paradigms: Provided a comprehensive comparison of Prompting, RAG variants, and Agentic workflows, offering actionable guidance for industrial proofreading systems.

4. Key Results and Findings

A. Model Performance

• Specialized vs. General Models: Specialized error correction models (e.g., CEC3-4B) underperformed general foundation models (e.g., Qwen3-4B, GLM-4.7) on the LEC task, particularly in punctuation and grammar, suggesting that foundation models have better transferability to new domains.
• Unified Context Superiority: The U-RAG (Unified RAG) approach consistently outperformed the S-RAG (Sequential) pipeline.
  • Reason: Sequential processing leads to accumulative over-correction (the model edits style twice) and contextual misalignment (factual modules fail to map corrections to the already-modified text).
  • Note: While U-RAG improved factual correction significantly, it did not necessarily improve linguistic recall compared to prompt-only, as retrieved evidence is often optimized for facts, not linguistic nuance.
• Agentic Framework: Performance was highly dependent on the backbone model's reasoning capabilities.
  • DeepSeek-V3.2 showed significant gains with the Agent framework, effectively planning and verifying facts.
  • Qwen3-235B and GLM-4.7 showed degraded performance with the Agent loop, likely due to instruction drift or formatting errors in multi-turn interactions.

B. Specific Challenges Identified

• Mixed-Error Difficulty: The MIX setting yielded the lowest recall across all models. A "masking effect" was observed where models fixed surface linguistic errors (e.g., typos) but missed the underlying factual inconsistency (e.g., correcting "urethritis" spelling but failing to realize the diagnosis was actually "vaginitis" based on symptoms).
• Over-Correction: LLMs exhibit a strong tendency to over-correct "clean" text. Precision drops as error density decreases because models actively seek out stylistic changes to satisfy the "correction" prompt.
• Weakly Normative Errors: Punctuation and grammar errors were the hardest to recall. These are often "weakly normative" (subjective or subtle) and are frequently overlooked even by native speakers, leading to low recall rates.
• Evidence Grounding: Factual correction relies heavily on external evidence. Models relying solely on parametric knowledge achieved sub-optimal F1 scores (~26.87%), whereas RAG methods boosted this to ~49.73%.

5. Significance

• Industrial Application: The paper provides a blueprint for building reliable, fully automatic proofreading systems for professional sectors. It highlights that unified processing is superior to decoupled pipelines for mixed-error scenarios.
• Reliability: The introduction of the Agentic workflow with explicit verification steps offers a path to reduce hallucinations and over-correction, though it requires robust backbone models.
• Future Directions: The authors suggest that future work should focus on low-resource methods (e.g., knowledge distillation) to make these complex agentic workflows feasible for real-time industrial applications, and on multi-reference annotations to better evaluate open-ended stylistic corrections.

In conclusion, CLFEC demonstrates that unifying linguistic and factual correction is not only necessary for professional writing but also technically challenging, requiring a shift from simple sequence-to-sequence models to evidence-grounded, agentic, or unified retrieval-augmented architectures.