Excel Data Annotation and Validation Tool

What Problem Does It Solve?

In Excel data processing, manual verification of annotation results—such as validating classification, sentiment analysis, or information extraction accuracy—is often required. Traditional methods rely on human inspection, which is time-consuming and prone to subjective errors. Automated scripts may overlook semantic complexities and data noise, leading to misjudgments. The Excel数据检查 skill addresses this by providing a structured reasoning workflow to validate result fields directly in original cells using colors and comments, without adding new columns, thus enabling efficient and reliable verification.

How Does It Work?

The core of this skill is the precision reasoning validation pipeline, ensuring independent judgment per row. The workflow includes:

• Data Characteristic Analysis: First, use the analyze command from scripts/validate_data.py to summarize result field value distributions, identify data types (e.g., speech transcription text), and infer annotation task types, forming validation rule hypotheses.
• Row-by-Row Reasoning Chain: For each row, execute a four-step process:
• Step 1 — Original Text Understanding: Comprehensively read data field content, noting noise (e.g., typos, colloquialisms) and key information in long texts.
• Step 2 — Independent Judgment: Infer expected results independently based on understanding, without viewing current annotation values.
• Step 3 — Comparison Analysis: Compare independent judgment with actual result field values; pause if inconsistent.
• Step 4 — Re-review Confirmation: Re-examine original text, considering semantic equivalence, reasonableness, and label granularity, ultimately determining "accurate," "error," or "to be confirmed." Standards emphasize preferring "to be confirmed" over erroneous "error" judgments.
• Writing Validation Results: Use the write-batch command to write results to a new file. Error cells are marked with red background and comments, to-be-confirmed cells with yellow background and comments, and accurate cells with no marking. Output files are named 原文件名-标注版.xlsx.

Boundaries and Considerations

This skill is designed for Excel file verification, but note:

• For large files (over 1000 rows), process in batches and inform the user of progress to avoid batch fatigue.
• If the result field is empty, directly mark as "to be confirmed" with a comment stating "result field is empty, cannot validate."
• When data fields contain sensitive information (e.g., ID numbers), protect privacy by avoiding full references in comments.
• For speech transcription text, restore the speaker's true intent to prevent misjudgment due to noise.
• Validation logic follows references/validation_rules.md, suitable for tasks like classification and information extraction in data analysis, but requires semantic reasoning for unstructured data (e.g., free text).