In the current wave of rapid technological advancement, many cutting-edge large language models (LLMs) have emerged, such as GPT-4 and Llama. However, the ability of these LLMs to generate smooth and coherent text has led to concerns about potential misuse. Therefore, in practical applications, the ability to distinguish them from texts created by human hands becomes especially crucial. A model for detecting machine-generated text is proposed through the PAN Task 4 - Voight-Kampf Generative AI Authorship Verification task. The model emphasizes the extraction of additional semantic information from the text. Various pre-trained language models (PLMs) were employed in the experiments, and the incorporation of multiple text features, such as word frequency and perplexity, was explored to enhance the outcomes. Ultimately, four runs were submitted, with the highest-performing approach attaining a mean score of 0.884 across all test datasets.
With the advent of Language Models, detecting machine-generated text has become a popular research direction[9, 10]. Existing detection methods can generally be classified into three categories: probabilitybased detection methods, machine learning-based detection methods, and deep learning-based detection methods. Probability-based methods typically do not require training samples. Gehrmann et al.[11] designed a statistical tool called GLTR, highlighting the distribution diferences between generated text and human-written text using diferent colors. GLTR detects the number of high, medium, and low probability words and presents the results in a visualized format. Mitchell et al.[12] proposed DetectGPT, which calculates log probabilities by perturbing text and comparing the changes between the original and perturbed text. It assumes that machine-generated text tends to lie within the negative log probability curve, while human-written text exhibits higher or lower probabilities compared to perturbed text. Machine learning-based methods, on the other hand, typically employ classical machine learning models. These methods often have fewer parameters and can be easily deployed. For instance, Solaiman et al.[13] applied the TF-IDF method to design a regression classifier that utilizes the Top-K sampling strategy to identify machine-generated text. Fröhling et al.[14] proposed a method that employs advanced features to simulate text’s experience, syntax, and semantics. They utilized simple classification models for detection, achieving performance comparable to mainstream deep learning methods.
Deep learning-based detection methods have gained widespread application in recent research and often outperform the previous two categories[15]. The emergence of the Transformer architecture has provided many advanced methods for deep learning. For example, Chen et al.[16] employed RoBERTa and T5 architectures to design a feature extraction and discrimination process between human-written text and text generated by ChatGPT. They trained two text classification models for text detection. Gambini et al.[17] demonstrated that fine-tuning XLNET on tweets written by humans, GPT-2, and earlier generation techniques, such as Markov chains and RNNs, can efectively identify GPT-3-generated tweets with an accuracy of 82.1%. It shows that fine-tuning XLNET on a mixed dataset can be a successful method for detecting machine-generated text.
It has been shown through practical applications that incorporating text features into PLMs is an efective method[ 18, 19, 20]. Some researchers have observed that incorporating features such as word embeddings[21], predictability[19], linguistic inquiry and word count[22], and other linguistic features can yield text encodings that contain high-quality semantic information. This feature integration enhances the models’ representation power and improves their performance in text detection tasks. The present study derives inspiration from the methods above and employs a similar approach by incorporating multiple features with text encoding to improve the model’s performance.
This section describes the model employed for automatic diferentiation between human and machinegenerated text. It includes an overview of the explored fusion features and the final ensemble architecture utilized. The architecture of the model is depicted in Figure 1. A PLM encodes the text and derives a pooled output for text representation. To capture features, an LSTM is utilized for language modeling [23, 24], generating text-based feature vectors. The pooled vectors and feature vectors are concatenated and inputted into a linear layer to obtain fusion logits. Subsequently, the softmax function is applied to derive the output probabilities. Alternatively, the pooled vector can be directly passed through a linear layer to yield PLM logits, which are then softmaxed to obtain the output probabilities.
Input text
PLM Hybrid model
Feature extraction module
Word frequency extraction
Perplexity extraction
Pooling layer
LSTM
Pooling output direct output Fusion features incorporate feature
Linear Feature vector
PLM logits
Output probabilities
(w/o feature) Fusion logits Output probabilities
(feature)
The inclusion of word frequency features is considered in the text encoding process. The word frequency corpus used in this work is sourced from the English dataset of the Google Book Corpus Ngrams1. This dataset encompasses textual data from millions of books published between 1500 and 2008, providing a comprehensive resource for word frequency information. The data is presented as "n-grams," consecutive sequences of n words. It can be utilized to study the distribution of token frequencies in machinegenerated text compared to human-authored text. Specifically, a word frequency dictionary is created using the Google Book Corpus, where each word group is mapped to its corresponding frequency of occurrence. Then, the tokenizer will be used to segment the input text. Look up the dictionary for each word group to get its counts (). If it does not exist in the dictionary, take 1. Finally, take the log of (). The process is expressed as Equation 1.
Freq = log (max (1, ()))
When encountering a lengthy word group comprising multiple words, the complete count is assigned to each word within the group. Ultimately, the word frequency feature representation of the text is obtained through the LSTM.
Perplexity is used to evaluate the predictive ability of a set of sample data. It measures the uncertainty or confusion level of the given data. Research by Tang et al. [10] has shown that language models tend to focus on common patterns in the text they are trained on, resulting in lower perplexity scores for text generated by LLMs. In contrast, human authors can express themselves in multiple styles, making it more challenging for language models to make accurate predictions and resulting in higher perplexity values for text created by humans.
GPT2 is considered the underlying generative model for perplexity features since most LLMs are based on the same Transformer architecture. GPT2 tends to assign low perplexity to common text and higher perplexity to text with varied styles. Perplexity can be measured by the entropy of the probability distribution. For a given sequence of n tokens , the probability distribution entropy () can be represented as Equation 2, where p represents the probability of generating the token corresponding to the vocabulary, and eps is the deviation term.
() = −
∑︁ () log2 (() + eps)
=1
1https://storage.googleapis.com/books/ngrams/books/datasetsv3.html
(
In addition, for a given , +1 represents the next token immediately following , ′ represents the
GPT2-predicted token for (the token with the highest probability in the vocabulary). Therefore, when
considering the token probabilities across the entire vocabulary in GPT2, the logarithmic probabilities
of the occurrence of the succeeding token +1 and the predicted token in the vocabulary are also
taken into account. As shown in Equation 3 and Equation 4, these measures assess the probability of
the context token occurring and the model’s confidence in predicting the token, respectively.
() = (+1)
() = (︀ ′
︀)
(
The number of context probability and prediction confidence values is n-1. Missing values are filled with 0 to align with the number of entropy values for the probability distribution. Finally, these features are concatenated together, and the LSTM network is used to obtain the perplexity feature representation of the text.
pooling layer. includes LSTM and PLM: Due to the excellent performance of Transformer-based PLMs in downstream classification tasks [ 15], three diferent variants of PLMs are explored: BERT, BERT-Large, and Roberta-Large. The [CLS] token representation with a length of either 768 or 1024 is utilized, and the output is extracted from the The primary focus is on incorporating text features. As shown in Figure 1, the hybrid model designed previous two sections and concatenated.
are used as the feature vector. • Text features are extracted using the word frequency and perplexity extraction introduced in the • The concatenated features are fed into the LSTM for encoding, and the hidden layer’s state values • The feature vector is concatenated with the PLM pooled output to obtain fusion features, which are then passed through a linear classification layer.
During the training phase, the hybrid model’s logits are directly outputted as the binary classification scores for each text after applying the softmax function. Since the provided samples are text pairs during the prediction phase, potential prediction biases caused by individual texts are minimized by predicting the probabilities for text1 and text2 separately.
An ensemble of two models in Figure 2 is considered. For instance, two hybrid models with PLM
using BERT-Large and Roberta-Large, respectively, are fine-tuned. The corresponding logits for text1
and text2 are obtained from the models and weighted. The final logits for each text are calculated as the
weighted sum, as depicted in Equation 5, where represents the weight coeficient.
logitsfinal = · logitsBERT_Large + (1 − ) · logitsRoberta_Large
(
Finally, the final logits of text1 and text2 are averaged after applying the softmax function. Specifically, the logits of text1 and (1 - logits) of text2 are added together and averaged. This results in the output probabilities during the prediction phase.
The oficial train dataset is derived from a collection of real and fake news articles from multiple news headlines in the United States in 2021. On the other hand, the test dataset includes various types of texts, such as news articles, Wikipedia summaries, or fan fiction. The training set consists of one portion Text1 Text2
Hybrid model (BERT-Large) Hybrid model (Roberta-Large) Logits of text2 weighted sum of human-authored data and 13 portions of data generated by diferent LLMs, each corresponding to a human sample topic. Each text data portion contains 1,087 samples and is provided in a dictionary format separated by a set of newline characters, such as {"id": "...", "text": "..."}. The test set will be provided in a diferent format, where each line contains a pair of texts, such as {"id": "...", "text1": "...", "text2": "...}. One of the texts is human-authored, while the other is machine-generated.
Figure 3 displays the text lengths and their corresponding probability density distributions plotted as KDE graphs for each data portion. Most text lengths are distributed between 300 and 500, implying that when utilizing PLMs, the information loss resulting from truncating excessively long texts does not need to be excessively concerned with.
Considering the imbalance between human-authored and machine-generated texts (1:13), data augmentation techniques are employed to expand the original training set, ensuring the quality of the model’s learning. Expressly, a benchmark dataset2 proposed by Sarvazyan et al.[25], similar in the domain to the provided human dataset and consisting of Wikipedia and news articles, is referred to. Since the LLMs used for machine-generated data difer from the training set, it was chosen not to utilize 2https://huggingface.co/datasets/symanto/autextification2023 the machine-generated text portion of this benchmark dataset. Instead, only human-authored data was extracted to augment and align the dataset with the training set.
A pre-trained sentence-transformers model3, capable of tasks such as similarity judgment or semantic search, was employed. Similarity comparisons were conducted between each text in the training set and the reference dataset. By encoding the texts using the sentence-transformers model, cosine similarity was calculated, and the top 12 texts with the highest similarity scores were selected. In the end, 13,044 similar texts were extracted for expansion.
The training set was re-divided into train and test sets using a 7:3 ratio. BERT, BERT-Large, and Roberta-Large were utilized as the base PLMs. The model was developed based on the PaddlePaddle framework for the framework and parameter selection. The batch size was 64, the max length was 512, and the learning rate was 2e-5. The model was trained for 10 epochs using the AdamW optimizer on an environment with an A800 GPU.
Additionally, the method introduces two parameters, , and _ , which specify the ensemble models’ weight proportion and decide whether the hybrid model outputs fusion logits or PLM logits. It provides us with the flexibility to generate diferent approaches.
The approach’s efectiveness was validated by running four diferent approaches on the final test set of the TIRA platform [26]. Their introductions are as follows: • rapid-pole: As a baseline for comparison, using only the prediction results from BERT, without including any word frequency or perplexity feature, _ set to False. • savory-plate: Similar to rapid-pole, but employs the Bert-Large model without any word frequency or perplexity feature, _ set to False. • lazy-iteration: Ensembles Bert-Large and Roberta-Large hybrid models, with the weighting coeficient in Equation 5 set to 0.9, _ set to True. • gritty-producer: Similar to lazy-iteration, but with the weighting coeficient in Equation 5 set to 0.1, _ set to True.
This section discusses the oficial metrics and baselines used in the task. The metrics employed encompass six diferent dimensions: • ROC-AUC: The area under the ROC (Receiver Operating Characteristic) curve. • Brier: The complement of the Brier score (mean squared loss). • C@1: A modified accuracy score that assigns non-answers (score = 0.5) the average accuracy of the remaining cases. • F1: The harmonic mean of precision and recall. • F0.5u: A modified F0.5 measure (precision-weighted F measure) that treats non-answers (score = 0.5) as false negatives.
• The arithmetic mean of all the metrics above.
The LLM detection baseline includes seven implementations: PPMd Compression-based Cosine (PPMd CBC), Authorship Unmasking, Binoculars, DetectLLM LRR and NPR, DetectGPT, Fast-DetectGPT, and Text length. PPMd CBC and Authorship Unmasking utilize a bag-of-words model, while Binoculars, DetectLLM, and DetectGPT employ LLMs to measure text perplexity. Text length serves as a randomness indicator for data integrity checks. The results of the partial baseline implementations can be found in the section 4.4. 3https://huggingface.co/annakotarba/sentence-similarity
The external results of the model can be seen in Table 1 and 2. Table 1 shows the results that the oficial baselines provided by the PAN organizers and summary statistics of all submissions to the task (i.e., the maximum, median, minimum, and 95-th, 75-th, and 25-th percentiles over all submissions to the task). Table 2 shows the summarized results averaged (arithmetic mean) over 10 variables of the test dataset.
The best-performing approach, "gritty-producer," achieved an average score of 0.966, surpassing all baseline methods. In the quantile results, the scores of participating teams are arranged in ascending order. This approach falls within the range of 75% to 95%.
As expected, the performance of approaches that incorporate features (lazy-iteration and grittyproducer) outperforms the individual PLM models (rapid-pole and savory-plate) on both datasets.
The results show that the scheme focusing on using Roberta_Large probability with language features (gritty-producer) is better than the scheme focusing on using Bert_Large probability with language features (lazy-iteration), which shows that the choice of PLM is also important. The model with more advanced pre-training skills is usually better for this use case.
This task uses PLMs to maintain a lead over almost all advanced baselines, including DetectLLM and DetectGPT using LLM, so introducing LLM to achieve the task was not explored. 95-th quantile 75-th quantile Median 25-th quantile Min
This paper proposes a text-based multi-feature fusion hybrid model for addressing the VoightKampf Generative AI Authorship Verification 2024 task. Experiments were conducted using various Transformer-based PLMs, and detailed insights into diferent feature extraction methods are provided. These methods efectively enhance the model’s performance, resulting in the task’s mean score of 0.884. For future work, improving the selection of hyperparameters, such as the output weights for each model, can be focused on, where grid search techniques can help identify better values. Additionally, further exploration can be done by incorporating additional text features to enhance the final output results. Acknowledgments
F1 0.884 0.905 0.875 0.884
F0.5 This work is supported by the Social Science Foundation of Guangdong Province, China (No.GD24CZY02) the Fifteenth International Conference of the CLEF Association (CLEF 2024), Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2024. [6] J. Bevendorf, M. Wiegmann, J. Karlgren, L. Dürlich, E. Gogoulou, A. Talman, E. Stamatatos, M. Potthast, B. Stein, Overview of the “Voight-Kampf” Generative AI Authorship Verification Task at PAN and ELOQUENT 2024, in: G. Faggioli, N. Ferro, P. Galuščáková, A. G. S. de Herrera (Eds.), Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum, CEUR Workshop Proceedings, CEUR-WS.org, 2024. [7] M. Davies, Making google books n-grams useful for a wide range of research on language change,
International Journal of Corpus Linguistics 19 (2014) 401–416. [8] P. Przybyła, Detecting bot accounts on twitter by measuring message predictability, 2019. [9] E. Crothers, N. Japkowicz, H. L. Viktor, Machine-generated text: A comprehensive survey of threat models and detection methods, IEEE Access (2023). [10] R. Tang, Y.-N. Chuang, X. Hu, The science of detecting llm-generated text, Communications of the ACM 67 (2024) 50–59. [11] S. Gehrmann, H. Strobelt, A. M. Rush, Gltr: Statistical detection and visualization of generated text, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, 2019, pp. 111–116. [12] E. Mitchell, Y. Lee, A. Khazatsky, C. D. Manning, C. Finn, Detectgpt: Zero-shot machine-generated text detection using probability curvature, in: International Conference on Machine Learning, PMLR, 2023, pp. 24950–24962. [13] I. Solaiman, M. Brundage, J. Clark, A. Askell, A. Herbert-Voss, J. Wu, A. Radford, J. Wang, Release strategies and the social impacts of language models (2019). [14] L. Fröhling, A. Zubiaga, Feature-based detection of automated language models: tackling gpt-2, gpt-3 and grover, PeerJ Computer Science 7 (2021) e443. [15] D. Macko, R. Moro, A. Uchendu, J. Lucas, M. Yamashita, M. Pikuliak, I. Srba, T. Le, D. Lee, J. Simko, et al., Multitude: Large-scale multilingual machine-generated text detection benchmark, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 9960–9987. [16] Y. Chen, H. Kang, V. Zhai, L. Li, R. Singh, B. Raj, Gpt-sentinel: Distinguishing human and chatgpt generated content, arXiv preprint arXiv:2305.07969 (2023). [17] M. Gambini, T. Fagni, F. Falchi, M. Tesconi, On pushing deepfake tweet detection capabilities to the limits, in: Proceedings of the 14th ACM Web Science Conference 2022, 2022, pp. 154–163. [18] A. Palmer, N. Schneider, N. Schluter, G. Emerson, A. Herbelot, X. Zhu, Proceedings of the 15th international workshop on semantic evaluation (semeval-2021), in: Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), 2021. [19] P. Przybyła, N. Duran-Silva, S. Egea-Gómez, I’ve seen things you machines wouldn’t believe: Measuring content predictability to identify automatically-generated text, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2023). CEUR Workshop Proceedings, CEUR-WS, Jaén, Spain, 2023. [20] P. Fivez, W. Daelemans, T. Van de Cruys, Y. Kashnitsky, S. Chamezopoulos, H. Mohammadi, A. Giachanou, A. Bagheri, W. Poelman, J. Vladika, et al., The clin33 shared task on the detection of text generated by large language models, Computational Linguistics in the Netherlands Journal 13 (2024) 233–259. [21] E. Ferracane, S. Wang, R. Mooney, Leveraging discourse information efectively for authorship attribution, in: Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2017, pp. 584–593. [22] A. Uchendu, T. Le, K. Shu, D. Lee, Authorship attribution for neural text generation, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 8384–8395. [23] M. Lippi, M. A. Montemurro, M. Degli Esposti, G. Cristadoro, Natural language statistical features of lstm-generated texts, IEEE Transactions on Neural Networks and Learning Systems 30 (2019) 3326–3337. [24] M. Sundermeyer, R. Schlüter, H. Ney, Lstm neural networks for language modeling., in: Interspeech, volume 2012, 2012, pp. 194–197. [25] A. Mikael Sarvazyan, J. Ángel González, M. Franco-Salvador, F. Rangel, B. Chulvi, P. Rosso, Overview of autextiflcation at iberlef 2023: Detection and attribution of machine-generated text in multiple domains., Procesamiento del Lenguaje Natural 71 (2023). [26] M. Fröbe, M. Wiegmann, N. Kolyada, B. Grahm, T. Elstner, F. Loebe, M. Hagen, B. Stein, M. Potthast, Continuous Integration for Reproducible Shared Tasks with TIRA.io, in: J. Kamps, L. Goeuriot, F. Crestani, M. Maistro, H. Joho, B. Davis, C. Gurrin, U. Kruschwitz, A. Caputo (Eds.), Advances in Information Retrieval. 45th European Conference on IR Research (ECIR 2023), Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2023, pp. 236–241. doi:10.1007/ 978-3-031-28241-6_20.