Large language models (LLMs) have led to a surge in machine-generated content across various platforms, posing a challenge for humans to distinguish between machine-generated and human-written text. To address this, there is an urgent need for automated systems that can identify machine-generated content and mitigate related risks. In response, the PAN@CLEF, in collaboration with the Voight-Kampf Task (ELOQUENT Lab) [ 1][2][3], proposed the Generative AI Authorship Verification task. This study presents a novel model architecture integrating Graph Neural Networks (GNNs), pre-trained Language Models (LLMs), and stylometric features to classify text documents as human-generated or machine-generated. Our approach employs a two-path structure: the first path processes text documents through data augmentation, transforming them into co-occurrence graphs for GNN processing, while the second path extracts and fine-tunes embeddings using a BERT-BASE model along with stylometric features. These embeddings are combined to enhance classification accuracy and robustness. We also detail our data stratification strategy, which involves augmenting human-generated texts to balance the dataset. The efectiveness of our model is demonstrated through extensive evaluation metrics, achieving superior performance over baseline methods.
Large language models (LLMs) have become widely available and easily accessible, increasing machinegenerated content across various platforms, including Q&A forums, social media, educational resources, and academic settings. Recent advancements in LLM technology, like ChatGPT and GPT-4, enable these models to produce coherent responses to most user inquiries, making them increasingly attractive for replacing human labor in various applications. However, this accessibility has raised concerns about potential misuse, such as generating fake news, impacting the financial services industry, afecting the legal domain, and causing disruptions in educational settings. Given humans’ dificulty distinguishing between machine-generated and human-written text, there is an urgent need to develop automated systems capable of identifying machine-generated content to mitigate the associated risks.
Motivated by these challenges, the PAN@CLEF [
In recent years, shared tasks focused on automatically detecting AI-generated text have risen, including
the prominent Autextification challenge [
Another exciting challenge task, SemEval-2024 Task-8 [
This section describes the system overview of our approach: Model Architecture, Data Stratification, and Graph Representation. Model Architecture lays out the structure and detailed workings that drive our method. Data Stratification shows the partition and data augmentation process for the data. Graph Representation explains the text-to-graph representation process.
The dataset consists of a Training Set and a Validation Set for model training and evaluation (see Table 1). The Training Set includes 23,433 text documents, with 12,129 human and 11,304 machine-generated texts. The Validation Set contains 5,654 text documents, evenly split between 2,827 human and 2,827 machine texts. Initially, there was a significant imbalance, with 14 machine-generated texts for every human-generated text. To balance the dataset, text augmentation was applied to the human texts using the back-translation method with a Large Language Model 1, along with techniques such as word insertions, synonyms, substitutions, and deletions 2. This augmentation ensured a balanced dataset, facilitating more efective training of models to distinguish between human and machine-generated content. 1HuggingFace Models used: Helsinki-NLP/opus-mt-en-ROMANCE to translate from English to another language (Spanish, French, etc.) and Helsinki-NLP/opus-mt-ROMANCE-en to back translate to enlish 2We used a python library called TextAttack: https://pypi.org/project/textattack/0.0.3.1/
Figure 1 illustrates a dual-path architecture that classifies text documents as human-generated or machine-generated. This architecture efectively integrates Graph Neural Networks (GNN) with pretrained Language Models (LLMs) and stylometric features to enhance classification accuracy and robustness.
The top path of the architecture employs a Graph Neural Network (GNN) approach [
The bottom path of the architecture leverages LLMs and stylometric features to complement the GNN approach. Like the top path, text documents undergo data augmentation, with specific techniques applied to human texts to ensure a balanced dataset. In addition, stylometric features are extracted, composed of linguistic and stylistic elements unique to each text. The stylo feature extracted includes mainly text-document stats such as: mean word length, mean sentence length, and the standard deviation of sentence length; furthermore, includes the counter of diferent punctuation marks such as commas, semicolons, quotes, exclamations, dashes, etc; finally, includes some connector word such as: and, buts, however, mores, this, etc. As an output, it generates a normalized stylometric feature vector that contains all these metrics for each text document from the corpus.
Concurrently, a pre-trained BERT-BASE model is fine-tuned on the text data, and CLS tokens are extracted to obtain document-level embeddings. These embeddings provide a comprehensive representation of the text at the document level. The architecture combines these CLS document embeddings with the previously extracted stylometric features. These combined features are concatenated with the GNN embeddings to form a unified representation that captures both the high-level semantic information and the detailed stylistic nuances of the text. This concatenated embedding is fed into another dense network, which performs the final classification into human-generated or machine-generated categories.
Finally, for each test case (pair of text: text1 and text2), it is required to output the ID of the input text pair and a confidence score between 0.0 and 1.0. A score < 0.5 means that text1 is believed to be human-authored. A score > 0.5 means that text2 is believed to be human-authored. A score of exactly 0.5 means the case is undecidable. To solve this within our architecture, as an output of our classification layer, a soft score probability is obtained for the documents to belong to each class (between 0.0 and 1.0); if text1 is human-authored, apply the rule: 1 - prob_text1; if text2 is believed to be human-authored, apply the rule: prob_text2; lastly, if the prediction is exactly 0.5 (or around an epsilon of 0.01) leave the score as 0.5 which means the problem is indecidable.
For the graph representation, We used the Co-Occurrence graph, where the words are represented as a node, and the Co-Occurrence of two words within the text document is defined as an edge between the words/nodes. As attributes/weights, edges have the frequency of co-occurrences between words in the text document, and the point-wise mutual information (PMI) measure between each word-to-word relation 4. As output, we will have one graph representation for each text document in the corpus.
For instance, let’s take the following sentence (from one of the documents in the corpus): millions in Texas lose power as the winter storm falls to -22c , represented in Figure 2. Each node in the graph corresponds to a unique word from the text, such as "power," "lose," "texas," etc. These nodes are connected by edges, which signify that the words co-occur within the same context or proximity in the text (window size of 2 in this example). The frequency weight indicates how often the connected words appeared together in the document. For example, an edge labeled "freq: 2" means that the two words it connects appeared together twice. The PMI weight quantifies the strength of association between two words, indicating how often the words co-occur more than would be expected by chance. A higher PMI value denotes a stronger association. For instance, PMI values such as 6.08 or 7.40 suggest a significant contextual relationship between the word pairs, even if they do not appear together frequently. This metric helps highlight meaningful word associations that might not be immediately obvious from frequency alone.
Finally, we use the text graph representations for each document and apply an LLM-based node feature initialization process. Subsequently, a GNN processes these graphs to obtain representations that encapsulate semantic meanings, syntactic structures, and contextual information.
4PMI measures the strength of association between two words by comparing their joint probability to the product of their probabilities
The six experimental approaches showcase strong performance overall. The final-run4gnnllm_llmft_stylofeat-partitionB stands out with a consistent score of 0.992 across all metrics, indicating highly reliable results. Similarly, final-run7-gnnllm_llmft_stylofeat-fullpartitionA exhibits slightly higher performance in ROC-AUC (0.994) but a marginally lower Brier Score (0.987), with an overall mean score of 0.99. final-run10-gnnllm_llmft_stylofeat-fullpartitionB also demonstrates robust performance, achieving a ROC-AUC of 0.989 and a Brier Score of 0.984, resulting in an overall mean of 0.987. Meanwhile, final-run8-gnnllm_stylofeat and final-run6-gnnllm_llmft_stylofeat show slightly lower performances with mean scores of 0.974 and 0.97, respectively.
Regarding the baselines, Binoculars perform comparably well, achieving an overall mean of 0.965 with a strong ROC-AUC (0.972) and F1 score (0.966). But, all of our approaches outperformed the baseline score reported.
On the other hand, Table 3 shows the summarized results averaged (arithmetic mean) over 10 variants of the test dataset. Each dataset variant applies one potential technique to measure the robustness of authorship verification approaches. In this evaluation, our best run final-run4-gnnllm_llmft_stylofeatpartitionB achieved an score of 0.997
Finally, Our submission scores achieved 14th out of 30 on the leaderboard with a ranking score of 0.727 overall test datasets. On the other hand, for part of the test datasets, this score had to be estimated since the system failed to run on short texts (35 words or less).
This study explores the eficacy of advanced model architectures in diferentiating between humangenerated and machine-generated text, addressing the growing concern of automated content misidentification in various domains. We implemented an architecture that combines Graph Neural Networks (GNNs) and pre-trained Language Models (LLMs) with stylometric features, showcasing superior performance compared to the baselines proposed.
Our experimental runs, particularly final-run4-gnnllm_llmft_stylofeat-partitionB and final-run7gnnllm_llmft_stylofeat-fullpartitionA, consistently achieved good scores across multiple evaluation metrics, significantly outperforming all baseline methods. These results afirm the robustness and reliability of our approach in accurately identifying machine-generated text.
Overall, our results highlight the promise of utilizing advanced GNN and LLM architectures and extensive feature sets to tackle the issues arising from the surge in machine-generated content. Future studies could enhance these approaches further and investigate their applicability across various languages and settings. Also, diferent text graph representations and graph neural network architecture should be tried with the LLM combination.
This paper has been supported by PAPIIT projects IN104424, TA101722, and CONAHCYT CF-2023-G-64. Also, the authors thank CONAHCYT for the computing resources provided through the Deep Learning Platform for Language Technologies of the INAOE Supercomputing Laboratory.