Scholarlink

<p>Question answering (QA) is the most challenging and important task in natural language processing (NLP). The majority of QA researchers rely on open-domain or monolingual settings based on specific languages or domains. Recently, multilingual question answering (MLQA) has played an essential role in accessing natural language interfaces. This scheme provides efficient access to multilingual data and provides precise answers based on the language. The limited benchmark dataset for MLQA is the major barrier to the evolution of the MLQA system. Thus, integrating deep learning with the MLQA system helps in improving the performance of the system. To encounter this issue, an effective MLQA system is designed using the optimal BERT-based convolutional neural network (OptBerConvoNet). The designed model is comprised of two phases, namely training and testing. Here, the input to the Bidirectional Encoder Representations from Transformers (BERT) tokenization is the group of questions and passages, where the sentences are converted into tokens. Thereafter, the extraction of features is done using various feature extractors based on the question and passage tokens. Then, the extracted features are subjected to the OptBerConvoNet to train the network. The answer obtained from trained OptBerConvoNet is matched with the target answer. In the testing phase, the trained OptBerConvoNet is tested to attain the relevant answer by considering a passage and questions as input. The OptBerConvoNet achieved an exact match of 95.67%, a precision of 97.57%, a recall of 92.77%, an F-measure of 95.11%, a Recall-Oriented Understudy for Gisting Evaluation (ROUGE) of 95.79%, and a bilingual evaluation understudy (BLEU) of 96.57%.</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

OptBerConvoNet: a framework based on BERT and optimized deep convolutional neural network for multilingual question and answering

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • S. Kanimozhi,
  • Selvamani K,
  • Amol Dattatray Dhaygude,
  • Vamsidhar Talasila

摘要

Question answering (QA) is the most challenging and important task in natural language processing (NLP). The majority of QA researchers rely on open-domain or monolingual settings based on specific languages or domains. Recently, multilingual question answering (MLQA) has played an essential role in accessing natural language interfaces. This scheme provides efficient access to multilingual data and provides precise answers based on the language. The limited benchmark dataset for MLQA is the major barrier to the evolution of the MLQA system. Thus, integrating deep learning with the MLQA system helps in improving the performance of the system. To encounter this issue, an effective MLQA system is designed using the optimal BERT-based convolutional neural network (OptBerConvoNet). The designed model is comprised of two phases, namely training and testing. Here, the input to the Bidirectional Encoder Representations from Transformers (BERT) tokenization is the group of questions and passages, where the sentences are converted into tokens. Thereafter, the extraction of features is done using various feature extractors based on the question and passage tokens. Then, the extracted features are subjected to the OptBerConvoNet to train the network. The answer obtained from trained OptBerConvoNet is matched with the target answer. In the testing phase, the trained OptBerConvoNet is tested to attain the relevant answer by considering a passage and questions as input. The OptBerConvoNet achieved an exact match of 95.67%, a precision of 97.57%, a recall of 92.77%, an F-measure of 95.11%, a Recall-Oriented Understudy for Gisting Evaluation (ROUGE) of 95.79%, and a bilingual evaluation understudy (BLEU) of 96.57%.