Rewrite quality checks allow grammatically malformed generalizations

Open
  • Is something wrong?
#276 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
48/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
python
Domain
ai, testing-qa

Research direction

Start by running docs/notebook_source/04_rewriting_biographies.py with the default rewrite configuration and inspect result.dataframe["biography_rewritten"]. Review the malformed examples and the quality metrics showing needs_human_review=False, then determine whether evaluation or repair should catch them. Done means malformed rewrites are corrected or marked for human review, with coverage beyond the reported biography cases.

Written by the indexing model from the issue text.

Description

bug
Priority Level

Medium (Annoying but has workaround)

Describe the bug

Summary

The rewrite pipeline can produce grammatically malformed sentences when generalizing an entity embedded in a noun phrase. These outputs can still pass the automated quality gate with relatively high utility scores and needs_human_review=False.

Observed behavior

While executing docs/notebook_source/04_rewriting_biographies.py, the pipeline produced:

Luca Hawthorne is a in his late 30s astronomer...

Robert Harper is a in his late 60s paramedic...

The age values were generalized, but the surrounding sentences were not restructured. Grammatically correct versions would be:

Luca Hawthorne is an astronomer in his late 30s...

Robert Harper is a paramedic in his late 60s...

The malformed outputs received the following metrics:

Output Utility Leakage Needs review
Luca Hawthorne… 0.908333 0.0 False
Robert Harper… 0.87 0.0 False

No repair was triggered.

Expected behavior

Generalized values should be integrated into grammatically correct prose. Clearly malformed rewrites should either be repaired automatically or marked for human review.

Likely failure mode

The model appears to substitute the generalized age phrase without restructuring the enclosing noun phrase:

  • a 37-year-old astronomer
  • a in his late 30s astronomer

Although the rewrite prompt requires natural prose, the current quality evaluation does not reliably detect this grammatical failure.

Impact

  • Generated text can be visibly ungrammatical despite passing evaluation.
  • Utility scores may overstate output quality.
  • needs_human_review does not identify malformed prose because it is score and threshold based.
  • Users may consume or publish defective rewrites without an obvious failure signal.

Suggested next steps

  1. Add prompt examples showing how attributive phrases must be restructured during generalization.
  2. Add explicit grammaticality or fluency criteria to the quality evaluator.
  3. Trigger repair when malformed syntax is detected.
  4. Study if this is the only case where it is happening, perhaps it's slipping through in other benchmark datasets too.
Steps/Code to reproduce bug

The malformed outputs were generated and committed in:

They were identified during review of PR #274:

To reproduce:

  1. Run docs/notebook_source/04_rewriting_biographies.py using the default rewrite configuration.
  2. Inspect result.dataframe["biography_rewritten"].
  3. Because generation is nondeterministic, repeated execution may be required.
  4. Observe that malformed output can receive a passing utility score and needs_human_review=False.
Expected behavior

Rewrite should generate correct grammar

Additional context

No response

Dominant language
Python
Stars
136
Forks
22
Avg merge
1d 22h
Merged PRs (30d)
47

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from NVIDIA-NeMo/Anonymizer

All issues in NVIDIA-NeMo/Anonymizer

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.