In this paper we present our approach for the detection of text written by Language Models (task 1). We further train models to classify the specific Language Model that creates the texts (task 2). Our final model ranks 49/76 for task 1 and 30/38 for task 2. Even though results are not among the best ones, we provide a more in-depth analysis of the employed features, which show the importance of character-level features and the low importance of TF-IDF weighted word-level features.
we can see that the text resembles John Milton’s words. Without attempting to compare this tiny language model to a Large Language Model as GPT, we tried to obtain some similar text using the popular chatbot ChatGPT. We provided the following prompt and we got the following answer on the 22nd of May, 2023:
Prompt: In John Milton’s style, how would you follow the sentence "And Adam"? Response: And Adam, with steadfast gaze and resolute countenance, beheld the vast expanse before him, a tapestry of untamed wilderness, teeming with uncharted wonders and boundless potential.
We further continued the interaction with ChatGPT and asked the following: "can you do it in iambic pentameter?". The response was:
And Adam, with his steadfast gaze and mien, Surveyed the vast expanse before his eyes, A tapestry of wilds yet unexplored, Where wonders dwelled, untamed and unrevealed,
A realm of boundless promise, yet unknown.
It is no doubt that recent language models sound notably more natural and fluent, especially when we compare them to older models. This brings many opportunities, but it also poses many dilemmas and dangers.
On the bright side, many repetitive tasks that involve reading the same data and writing very similar reports, could be simplified by the use of these models, such as, table report writing. These models seem to work also very well in style change, for instance writing a paragraph in a very diferent register, for instance writing an academic text for teenagers.
What seems to be a nice help, though, it can also be seen as a dangerous step. For instance, until recently, spam messages or scams were rather easy to spot (fool language, typical repeated scenarios, and so on). But with the advent of new LLMs, we, e-mail recipients, can now be tricked in an easier way. For the prompt "Can you write a catchy SMS message to buy an umbrella?" we get the following answer on the 23rd of May, 2023 (the answer even included emojis that could not be incorporated because of encoding issues):
Don’t let the rain spoil your day! Stay dry and stylish with our fabulous umbrellas! Whether it’s a sudden shower or a drizzle, our umbrellas have got you covered. Get yours now and dance in the rain with confidence! Shop today and embrace the rainy season with a touch of elegance. Hurry, limited stock available! Don’t miss out on this weather-defying deal. #StayDryStayHappy
Then, even though there are many benefits with these models, we believe it is relevant to be able to detect whether a text was created by a human or a computer.
In this paper, we present our approach to detect whether a text was written by a Language
Model or by a human, as part of the AuTexTification shared task [
With the improvement of language models, it is no surprise that recognizing whether a text was written by a human or a computational model is becoming more challenging. We review some works that attempted to solve this problem in the following lines.
Some researchers have worked on deepfake detection in tweets. In their work [
Many general classification works rely on the availability of large language models. In spite
of that, Frohling and colleagues [
We use the data [
This task is tackled following a general methodology that we expected to work for many tasks,
including bot vs. human discrimination, subjectivity detection, sexism detection, and many
others. Our original goal in the framework is to faithfully compare feature importance across
diferent tasks, but in this article we will emphasize in the characteristics of bot vs. human
and bot discrimination problems. Our feature set is mainly inspired by work on authorship
attribution and profiling [
In this study, we employ seven distinct classifiers for the purpose of training. These classifiers consist of K-Nearest Neighbors (with = 5), Logistic Regression, Linear Support Vector Machine, Multilayer Perceptron, Decision Tree, Random Forest, and Gradient Boosting.
To ensure model evaluation, we adopt a train/test validation procedure. Specifically, we allocate two thirds of the available data for training purposes, while the remaining one third is reserved for testing. This approach allows for reliable assessment of the classifiers’ performance and generalization capabilities.
The results for task 1 and task 2, obtained from various classifiers, are provided in Tables 2 and 3, respectively. The evaluation metrics considered include the weighted F1-score, accuracy, and training time. We believe that training time holds significant importance in practical applications of such models, particularly in real-world scenarios involving private companies or public institutions.
Notably, the inference time is not included in the reported results as it demonstrates relatively similar performance across all classifiers, except for the K-Nearest Neighbors classifier. However, due to its comparatively low performance, the inference time is considered less relevant in this context.
For task 1, the classifiers achieved a ranking of 49 out of 76, with a macro F1-score of 57.35. In the case of task 2, the classifiers attained a ranking of 30 out of 38, yielding a macro F1-score of 45.72. These metrics provide an overview of the classifiers’ performance in the respective tasks.
Rather unexpectedly, we can observe that our test results from task 1 are rather low, compared to the validation results from table 2. It could have been a case of overfitting to the training data or also that the domain shift of the test data afected more than what we expected, making our model domain dependent. In task 2 we observed that the GB method showed slightly better results but a significantly larger training time (as it can be seen in table 3). Consequently, we ultimately opted to utilize the Multilayer Perceptron as our chosen classifier for the test set, as it demonstrated similar results to GB.
Based on the feature importance of the Logistic Regression models trained, it seems like BERT related features can be good for finding out whether a text is written by a human or a bot. But they do not seem to hold the same power for establishing the system used (task 2).
The analysis of feature performance in our study reveals certain trends. Firstly, the utilization of Character Bag-of-Words features exhibits favorable predictive capability for both tasks under investigation. On the other hand, the incorporation of morphological features extracted through the Stanza library does not appear to provide predictive value for either of the tasks.
Notably, POS uni- or ngrams demonstrate relatively strong predictive potential as feature representations. These features, based on part-of-speech information, exhibit promising predictive power for the tasks at hand.
However, our findings indicate that TFIDF normalization does not significantly contribute to improved performance. The TFIDF normalized BOW features do not appear to possess substantial relevance in the context of the tasks examined, especially when compared to the regular BOW features.
In this paper, we presented our approach for tasks 1 and 2 of the AuTexTification Shared
Task [
Considering that our results are still quite low, there is still room for improvement. We
would like to include more complex features, such as the ones proposed in [