TL;DR
This paper presents a hybrid approach for distinguishing human-written text from machine-generated text, combining linguistic features with token-level probabilistic features processed by a BiLSTM. Despite a promising methodology, the system underperformed significantly compared to baselines and other participants in the SemEval-2024 Task 8, highlighting the challenges in this domain.
What Problem Does This Paper Address?
The rapid advancement and widespread adoption of generative language models (like ChatGPT) have created a pressing need to reliably distinguish between human-written and machine-generated text. This is crucial for various applications, from academic integrity to combating misinformation. However, humans, even experts, struggle to consistently identify AI-generated content, and current commercial solutions are not yet robust enough. The SemEval-2024 Task 8 aimed to address this challenge by focusing on multidomain, multimodal, and multilingual machine-generated text detection. This paper specifically tackles the problem of authorship attribution, aiming to determine if a text is human-written or machine-generated, and if machine-generated, which model produced it.
Methods
The Groningen Team A's approach is a hybrid system that combines two main types of features:
- Document-level Features: These are extracted directly from the text and include:
- Perplexity: Measured using the XLM-ROBERTa language model, it quantifies how surprised a model is by a given text, with lower perplexity often indicating machine generation.
- TF-IDF: Standard term frequency-inverse document frequency features based on word-level unigrams.
- Stylometric Features: A set of features like average sentence length, punctuation count, and the distribution of capitalized words and Part-of-Speech (POS) tags.
- Information Redundancy: Measures lexical or topical repetition, hypothesized to be higher in machine-generated text.
- Entity-based Coherence: Analyzes how entities are referenced across sentences, based on the idea that human and machine texts might differ in their entity transition patterns. This was only used in English-only tracks due to limitations in co-reference resolution.
- Token-level Features (Predictability Features): These features, inspired by Przybyła et al. (2023), assess the likelihood of token sequences using generative language models. Key components include the log-probability of observed tokens, the log-probability of the most likely token, and the entropy of the token probability distribution. These features were processed by a Bidirectional Long Short-Term Memory (BiLSTM) network to capture sequential patterns.
The document-level and token-level features were concatenated and fed into a Multi-Layer Perceptron (MLP) for final classification. They also explored a separate BERT-LSTM model as a baseline, using BERT for sentence embeddings followed by an LSTM layer.
Datasets
The datasets used were provided by the SemEval-2024 Task 8 organizers. These datasets are an extension of the M4 dataset from Wang et al. (2024b) and include texts from various domains (e.g., Reddit, Wikipedia, arXiv) and multiple languages (English, Russian, Chinese, Arabic, Urdu, Indonesian, Bulgarian). Subtask B specifically contained machine-generated texts from multiple models.
Key Insights and Ideas
- Hybrid Approach Limitations: The core finding is that their combination of linguistically motivated features and token-level probabilistic features, while theoretically sound, did not yield significant performance gains. This suggests that the chosen features, despite showing differences between human and machine texts, did not translate into effective learning for their hybrid classifier.
- Feature Differences Exist but Don't Guarantee Performance: The analysis of raw features revealed notable differences in values for human vs. machine texts (e.g., higher frequency of pronouns and auxiliary verbs in human texts, higher frequency of nouns and adjectives in machine texts). However, these differences did not translate into improved model performance, underscoring the complexity of the detection task.
- Bias Towards "Machine" Label: The models, particularly for the monolingual Subtask A and Subtask B, exhibited a strong bias, predicting all examples as machine-generated or specifically "ChatGPT." This indicates a failure to learn nuanced distinctions.
- Multilingual Subtask A Showed Some Promise: The multilingual Subtask A model was the only one that could distinguish between examples to some extent, though it also showed an extreme bias towards the "machine" label.
Key Results
- Overall Underperformance: The feature-based system performed poorly across all subtasks.
- Monolingual Subtask A: Accuracy of 0.525 (Baseline: 0.884), ranking 128/137.
- Multilingual Subtask A: Accuracy of 0.512 (Baseline: 0.808), ranking 61/68.
- Subtask B: Accuracy of 0.166 (Baseline: 0.746), ranking 74/77.
- BERT-LSTM Model: This model also underperformed, achieving only 16.67% accuracy on Task B and was not experimented with for Task A due to time constraints.
- Confusion Matrix (Multilingual A): The confusion matrix for multilingual Subtask A showed a significant imbalance, with 17147 predicted "Machine" texts versus only 406 predicted "Human" texts, highlighting the bias.
Limitations and Future Work
- Feature Set Limitations: The authors explicitly state that their examined subset of linguistically motivated features, alongside probabilistic features, was not able to contribute almost any performance. This suggests a need for further exploration and refinement of feature engineering strategies.
- Computational Overhead: The BERT-LSTM model introduced significant computational overhead, limiting experimentation.
- Time Constraints: Due to time constraints, the BERT-LSTM model could not be fully experimented with for all tasks.
- Future Directions: The paper emphasizes the need for further exploration and refinement of feature engineering strategies to effectively address this evolving domain. The nuanced challenges in developing robust detection mechanisms for machine-generated text require continued research.
Summary updated