Evaluating LLM-generated text

April 3, 2024 · 3 minute read · NLP

Key Ideas

  • Human judgement can be learnt by LLMs to serve as replacements for text quality evaluation tasks
  • However this runs into problems of generalization and LLM specific bias

Notes

Fixed Scope Task Evaluation - Human Evaluation

Generally indicated by scores based on subjective measures on a 5 point scale

  • Adequacy: Is the meaning correct?
  • Fluent: Is it easy to read?

Cons

  • Subjective!
  • Difficult to calibrate
  • Expensive and time consuming

Fixed Scope Task Evaluation - Automatic Metrics

Precision, Recall and F-Scores

  • Use the precision (common words in y_cap and y_pred/y_pred length), the recall (common words in y_cap and y_pred/y_cap length) and F-Score (precision * recall/((precision + recall)/2))

Pros

  • Quick, easy and cheap

Cons

  • Does not handle synonyms
  • Does not handle order
  • We may not at always have a reference
  • Does not take into account meaning of the constructed sentence

Bilingual Evaluation Understudy (BLEU)

  • Based off n-gram overlap between y_pred and y_cap
  • Computes precision for n-grams of size 1 to 4.
  • Has a brevity penalty
  • `BLEU = min(1, output-length/ref-length) (PROD_{i, 1:4} precision_i)^1/4
  • Allows use of multiple references, and we can match against all refs, so recall will not be very useful
    • Closest reference length is used

Cons

  • All words/n-grams are treated equally
  • Human translations also score lower than machines
  • The score does give any indication, cannot be used comparatively with other test strings

ROUGE

  • Based off n-gram overlap between y_cap and y_pred
  • Computes recall for n-grams of size 1 to 4
  • Used for text summarization systems

Cons

  • All words/n-grams are treated equally
  • Human translations also score lower than machines
  • The score does not matter
  • Can game the score by just replicating the string n-times

Fixed Scope Task Evaluation - Learned Metrics

  • Finetune a model directly on scores from human evaluations to perform evaluation

BLEURT

  • Finetune a pretrained BERT model on synthetic tasks with perturbed data and automatic metrics. Then finetune again with human evaluation metrics.

COMET

  • SOTA LLM-as-a-judge for similarity scores

Open Ended Task Evaluations - Human Evaluations

Challenges when using humans

  • Subjective
  • Needs experts to evaluate
  • Annotators might not do a good job

Long Eval

  1. Split long form text into atomic claims
  2. Get each claim verified for support from the long form text
    • (optional) Send only subset of claims to each individual annotator to reduce workload
  3. Calc %age of facts being supported by the text

Open Ended Task Evaluations - LLM Evaluations

GPTEval

  • Create a prefix/context for LLM with instructions on how to evaluate and to give a score
  • Aggregate scores w/ probabilities to calculate evaluation

Win Rate

  • Ask LLM to select one of two outputs and use that as a score
  • One of the two outputs should ideally come from the same base model to ensure fair comparison between two models
  • Caveat: The selected annotator model might prefer a specific class of model (responses created by OpenAI models may be preferred by OpenAI models)

Decompose, Eval and Aggregate

  1. Use an LLM to break down a text into claims
  2. Verify each claim with an LLM + an evidence source
  3. Calculate the retrieval score

Resources

Twitter, Facebook