Advacheck at GenAI Detection Task 1: AI Detection Powered by Domain-Aware Multi-Tasking

TL;DR

A system designed by Advacheck team to recognise machine-generated and human-written texts in the monolingual subtask of GenAI Detection Task 1 competition is described, finding that multi-task learning outperforms single-task mode and simultaneous tasks form a cluster structure in embeddings space.

Abstract

The paper describes a system designed by Advacheck team to recognise machine-generated and human-written texts in the monolingual subtask of GenAI Detection Task 1 competition. Our developed system is a multi-task architecture with shared Transformer Encoder between several classification heads. One head is responsible for binary classification between human-written and machine-generated texts, while the other heads are auxiliary multiclass classifiers for texts of different domains from particular datasets. As multiclass heads were trained to distinguish the domains presented in the data, they provide a better understanding of the samples. This approach led us to achieve the first place in the official ranking with 83.07% macro F1-score on the test set and bypass the baseline by 10%. We further study obtained system through ablation, error and representation analyses, finding that multi-task learning outperforms single-task mode and simultaneous tasks form a cluster structure in embeddings space.

Figures (7)

• Figure 1: Overview of the proposed multi-task architecture. Modules marked only with are trainable at all stages. The weights of Transformer Encoder are frozen at the first stage of training and trainable at the second one. The Custom Classification Head (CCH) described in Appendix \ref{['sec:appendix_a']} is used for predictions.
• Figure 2: Two principal component decomposition of PCA for texts from the development subsample. In (a) the vector space structure for the deberta-v3-base fine-tuned in single-task mode is shown, while figure (b) shows the same model but fine-tuned in MTL mode with two additional custom classification heads.
• Figure 3: Macro $F_1$-score on the test set of different configuration of the systems depending on the threshold.
• Figure 4: Proportion of predictions for different generators from test set. Labels in bold are generators texts from which are present in the train set.
• Figure 5: The architecture of the custom classification head used in the described approach. Value of dropout is equal to 0.5.