Team jelarson at SemEval 2024 Task 8: Predicting Boundary Line Between Human and Machine Generated Text

Joseph Larson, Francis TyersNAACL 2024

TL;DR

This paper tackles the challenge of precisely identifying where a Large Language Model (LLM) begins writing within a document that starts with human text and is then completed by an AI. The authors developed a system that combines heuristic rules with linear regression on text length to predict this "boundary line," achieving a Mean Absolute Error (MAE) of 48.139 on the test set. This work highlights the importance of data quality and suggests that subtle linguistic anomalies can be key indicators of AI-generated content.

What Problem Does This Paper Address?

The proliferation of LLMs has led to an explosion of machine-generated text across various domains, raising concerns about its potential misuse (e.g., fake reviews, spam, misinformation). While distinguishing between human and machine-generated text is a known challenge, this paper focuses on a more granular task: identifying the exact point where an LLM takes over within a document that has a mixed human-AI origin. This is crucial for understanding the transition dynamics and potentially for developing more sophisticated detection methods that go beyond simple binary classification. The authors note that humans struggle to detect machine-generated text at chance level, underscoring the need for automated solutions.

Methods

The team employed a multi-pronged approach:

  • Data Examination and Heuristics: They meticulously analyzed the training data, identifying linguistic anomalies that frequently appeared in either human or machine-generated text, or specifically around the transition point. Key anomalies included:
  • Patterns like "Word..Word" and ".Word" (where '.' represents a punctuation mark or space).
  • The presence of single line breaks exclusively in human-written text.
  • Differences in formality, emotionality, punctuation usage, and dependency relation lengths between human and LLM text.
  • Based on these observations, they developed heuristic rules. For instance, one rule identified the transition point as occurring after a "Word..Word" pattern. Another rule leveraged the observation that single line breaks were a strong indicator of human text.
  • Linear Regression: They explored the relationship between document length and the transition point. A linear regression model was trained to predict the transition index based on text length. This proved to be a robust baseline, especially when combined with the heuristic rules.
  • Model Combination: The final system combined the most effective heuristic rules with the linear regression model. When a heuristic rule didn't apply, the linear regression model was used.

Datasets

The paper utilizes a dataset previously introduced by Wang et al. (2024) for the SemEval-2024 Task 8. This dataset consists of documents that begin with human-written text and are then completed by an LLM. The authors created scripts to manually examine transition words and anomalies within this dataset for training and evaluation.

Key Insights and Ideas

  • Linguistic "Fingerprints": Even sophisticated LLMs leave subtle linguistic traces. The paper highlights that patterns like "Word..Word" and the exclusive use of single line breaks by humans are surprisingly strong indicators of the text's origin.
  • Length Matters (and Outliers are Tricky): Document length is a significant predictor of the transition point, but the presence of outliers (documents with extreme lengths) can negatively impact the performance of simple linear models.
  • Data Quality is Paramount: The authors strongly emphasize that textual anomalies in the training and test data can significantly hinder the performance of even advanced models. Cleaning this data is crucial for accurate detection.
  • Heuristics Can Be Powerful: Despite the rise of complex neural models, well-crafted heuristic rules based on careful data analysis can provide significant improvements, especially when combined with other methods.

Key Results

  • The best heuristic rules (e.g., detecting "Word..Word" patterns) significantly reduced the Mean Absolute Error (MAE) compared to random guessing.
  • A baseline linear regression model achieved an MAE of 20.464 on the training data.
  • The final submitted system, a combination of heuristic rules and a linear regression model, achieved an MAE of 48.139 on the test set. This score was noted as being higher than the best models in the shared task (around 16.0 MAE), which the authors attribute to time constraints and the complexity of the task.
  • The paper found that the test data was less linearly distributed than the training data, leading to poorer performance for the linear model on unseen data.

Limitations and Future Work

  • Data Anomalies: The authors acknowledge that textual anomalies in the dataset significantly impacted their results. They hypothesize that removing these anomalies could lead to better performance for neural models.
  • LLM Sophistication: As LLMs become more advanced, current detection methods might become less effective. The paper suggests that future work could explore more sophisticated statistical models that analyze sampling frequencies.
  • Outlier Handling: The linear regression model struggled with outliers in text length. Future work could involve developing more robust methods for handling these cases, perhaps through multi-modal models or adaptive binning strategies.
  • LLM Agnosticism: The paper does not specify the exact LLM used to generate the test data, which could limit the generalizability of findings. Future research could investigate how detection methods perform across different LLMs.

Summary updated

BibTeX
@inproceedings{larson2024team,
  author = {Joseph Larson and Francis Tyers},
  title = {{Team jelarson at SemEval 2024 Task 8: Predicting Boundary Line Between Human and Machine Generated Text}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure