Scholarlink

<p>In this paper, we introduce a baseline approach for solving the authorship verification task by incorporating a straightforward Automated Machine Learning (AutoML) method, including a defined neural network search space and an exhaustive search strategy. Our approach aims to automatically identify prominent neural networks that generate well-performing models simulating a general Neural Architecture Search (NAS) process in the authorship identification problem. To this end, we compare our approach against traditional Machine Learning (ML) methods for generating models across different text corpora in English and Spanish documents. Additionally, we demonstrate that a pre-trained model like a Bidirectional Encoder Representation from Transformers (BERT), which captures the context of words in a bidirectional manner, combined with our approach, can be a suitable option for capturing the author’s writing style. General NAS’s theoretical and practical implications are also discussed, highlighting the challenges of using this technique across more complex architectures on different corpora and its inherent limitations, resulting in several guidelines for integrating NAS to solve authorship verification. Finally, experimental results show that our proposed pipeline can be applied to a Natural Language Processing (NLP) problem without extensive computational resources, outperforming most common techniques.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Toward the Use of Neural Architecture Search: A Baseline Approach for Solving the Authorship Verification Problem

  • 期刊论文

  • 发表日期：

    2025年7月16日

查看全文
  • Luis Ignacio Ferro Salinas,
  • Esteban Castillo,
  • Víctor Adrián Sosa Hernández

摘要

In this paper, we introduce a baseline approach for solving the authorship verification task by incorporating a straightforward Automated Machine Learning (AutoML) method, including a defined neural network search space and an exhaustive search strategy. Our approach aims to automatically identify prominent neural networks that generate well-performing models simulating a general Neural Architecture Search (NAS) process in the authorship identification problem. To this end, we compare our approach against traditional Machine Learning (ML) methods for generating models across different text corpora in English and Spanish documents. Additionally, we demonstrate that a pre-trained model like a Bidirectional Encoder Representation from Transformers (BERT), which captures the context of words in a bidirectional manner, combined with our approach, can be a suitable option for capturing the author’s writing style. General NAS’s theoretical and practical implications are also discussed, highlighting the challenges of using this technique across more complex architectures on different corpora and its inherent limitations, resulting in several guidelines for integrating NAS to solve authorship verification. Finally, experimental results show that our proposed pipeline can be applied to a Natural Language Processing (NLP) problem without extensive computational resources, outperforming most common techniques.