This paper describes a proposed system design for Style Change Detection (SCD) tasks for PAN at CLEF 2022. We propose a unified architecture of ensemble neural networks to solve three SCD-2022 edition tasks. We fine-tune the BERT, RoBERTa and ALBERT transformers and their connecting classifiers to measure the similarity of two given paragraphs or sentences for authorship analysis. Each transformer model is regarded as a standalone method to detect differences in the writing styles of each testing pair. The final output prediction is then combined using the majority voting ensemble mechanism. For SCD-2022 Task 1, which requires finding the only one position of a single style at the paragraph level, our approach achieves a macro F1-score of 0.7540. For SCD-2022 Task 2 to detect the actual authors of each written paragraph, our method achieves a macro F1-score of 0.5097, a Diarization error rate of 0.1941 and a Jaccard error rate of 0.3095. For SCD-2022 Task 3 to find located writing style changes at the sentence level, our model achieves a macro F1-score of 0.7156. In summary, our method is the winning approach in the list of all intrinsic approaches.
|頁（從 - 到）||2565-2573|
|期刊||CEUR Workshop Proceedings|
|出版狀態||已出版 - 2022|
|事件||2022 Conference and Labs of the Evaluation Forum, CLEF 2022 - Bologna, Italy|
持續時間: 5 9月 2022 → 8 9月 2022