Human Perception in Natural Language Generation

Lorenzo De Mattei, Huiyuan Lai, Felice Dell’Orletta, Malvina NissimACL 2021

TL;DR

This paper explores how to make AI-generated text more human-like by fine-tuning a GPT-2 model on human perception data. They found that fine-tuning improved human perception of generated text and developed an automatic evaluation method that correlates well with human judgments.

What Problem Does This Paper Address?

Large language models (LLMs) like GPT-2 can generate remarkably human-like text. However, existing evaluations show that while many generated sentences are perceived as human-written, a significant portion are not. This raises the question: can we explicitly train LLMs to produce text that humans perceive as more human-like, even if the underlying generation process is still artificial? The authors also highlight that current evaluation methods for generated text might not fully capture human perception, and they aim to bridge this gap.

Methods

The core of this research involves two main methodological contributions:

  • Fine-tuning for Perception: Instead of just training on "gold" (human-written) data, the researchers collected human judgments on whether texts were human or machine-generated. This perception-labeled data was then used to fine-tune a pre-trained GPT-2 model (specifically, an Italian version called GePpeTto). They explored two fine-tuning strategies:
  • GePpeTto-F (Fine-tuned): Standard fine-tuning using the perception labels.
  • GePpeTto-R (Rewarded): An additional step where the model is rewarded for generating text that a separate "perception classifier" deems human-like. This is inspired by reinforcement learning techniques used in style transfer.
  • Automatic Evaluation: To overcome the cost and scalability issues of human evaluation, the authors developed an automatic evaluation strategy. They trained a regressor model on the perception-labeled data to predict human judgments. This regressor was then used to evaluate the generated texts from different models.

Datasets

  • Training Data: The authors created a custom dataset by:
  • Selecting 1400 human-written sentences from the original GePpeTto training corpus.
  • Generating 1400 sentences using the base GePpeTto model.
  • Having native Italian speakers rate these sentences on a 1-5 Likert scale for human-likeness.
  • This data was then binarized (human < 3, machine >= 3) for classifier training.
  • Test Data: A separate set of 1400 sentences was used for evaluation:
  • 350 human-written sentences.
  • 1050 generated sentences from three models (GePpeTto, GePpeTto-F, GePpeTto-R).
  • These sentences were also rated by human judges.

Key Insights and Ideas

  • Perception is Key: The study demonstrates that explicitly training a model to be perceived as human-like, rather than just mimicking human text, can lead to better results. This shifts the focus from pure linguistic fidelity to subjective human judgment.
  • Fine-tuning Works: The fine-tuned model (GePpeTto-F) was consistently perceived as more human-like than the original GePpeTto model by human evaluators.
  • Automatic Evaluation is Reliable: The developed automatic evaluation regressor showed a strong correlation with human judgments, suggesting it can be a viable alternative for large-scale evaluations. This is a significant contribution for the field.
  • Simplicity Can Be Deceptive: Interestingly, the linguistic analysis revealed that texts perceived as human-like by machines (i.e., those the regressor rated highly) often exhibited simplicity (e.g., shorter sentences, fewer subordinate clauses). This contrasts with human judgments, where complexity is often associated with human-likeness. This suggests that current automatic evaluation metrics might be susceptible to simpler, less complex outputs that are easier for models to generate.

Key Results

  • Human Evaluation:
    • Gold (human-written) texts were perceived as most human-like by human judges (average score of 2.41).
  • The fine-tuned model (GePpeTto-F) performed best among the generated texts, outperforming the base GePpeTto and the rewarded model (GePpeTto-R).
  • GePpeTto-F achieved a human perception score of 2.74, while GePpeTto-R scored 2.84, and the base GePpeTto scored 2.85. (Lower scores indicate more human-like perception).
  • Automatic Evaluation:
    • The automatic regressor also ranked GePpeTto-F as the best performing model, aligning with human judgments.
  • The regressor showed good correlation with human scores (Pearson=0.54).
  • Linguistic Analysis:
    • For human texts, complexity (longer sentences, deeper syntactic links) correlated with human perception.
  • For machine-perceived texts (those rated highly by the regressor), simplicity was a key feature, possibly because it reduces the chance of exposing machine-like errors.

Limitations and Future Work

  • Model Limitations: The rewarded model (GePpeTto-R) did not perform as well as the simply fine-tuned model (GePpeTto-F) in human perception, suggesting that the reward mechanism might not have been optimally tuned or that the perception classifier itself has limitations.
  • Data Limitations: The study used a relatively small dataset for fine-tuning and evaluation. The authors suggest expanding the perception-labeled data to assess the impact of training size and to perform finer-grained analysis across different text genres and subject demographics.
  • Automatic Evaluator Bias: The authors note that the automatic regressor might overestimate the human-likeness of generated texts, especially when the training data for the regressor itself was derived from machine-generated text. This is an area for further refinement.
  • Future Work: The authors propose exploring the impact of training data size, analyzing different text genres, and incorporating demographic information into the evaluation. They also acknowledge the potential for malicious use of text generation models and emphasize the importance of raising awareness about these risks.

Summary updated

BibTeX
@inproceedings{mattei2021human,
  author = {Lorenzo De Mattei and Huiyuan Lai and Felice Dell’Orletta and Malvina Nissim},
  title = {{Human Perception in Natural Language Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure