CrowNER at ROCLING 2023 MultiNER-Health Task: Enhancing NER Task with GPT Paraphrase Augmentation on Sparsely Labeled Data

CrowNER at ROCLING 2023 MultiNER-Health Task: Enhancing NER Task with GPT Paraphrase Augmentation on Sparsely Labeled Data Yin-Chieh Wang∗ Telexpress Co., Ltd. pony.wang@telexpress.com Feng-Yu Kuo∗ Telexpress Co., Ltd. bruce.kuo@telexpress.com Te-Yu Chi Department of CSIE National Taiwan University d09922009@ntu.edu.tw Sheh Chen Telexpress Co., Ltd. shepherd.chen@telexpress.com Wen-Hong Wu∗ Telexpress Co., Ltd. vincent.wu@telexpress.com Han-Chun Wu∗ Telexpress Co., Ltd. andy.wu@telexpress.com Te-Lun Yang Department of CSIE National Taiwan University d12944007@ntu.edu.tw Jyh-Shing Roger Jang Department of CSIE National Taiwan University jang@mirlab.org Abstract In this research, we utilized the training dataset from the ROCLING 2023 Chinese Multi-genre Named Entity Recognition in the Healthcare Domain, which comprises the Chi- nese HealthNER Corpus (Lee and Lu, 2021) and the ROCLING 2022 CHNER Dataset (Lee et al., 2022), along with the test set (Lee et al., 2023). The objective was to address the named entity recognition task within the Chinese healthcare domain. Our initial step in- volved preprocessing the training dataset. We identified instances in the training set where sentences with identical structural patterns ex- hibited ambiguities and errors in named en- tity definitions. Prioritizing data validation, we manually excluded erroneous entries. In specialized domains such as medicine, domain- specific terminologies and proprietary names are often defined within sentences as merged labels, rather than separate ones. Thus, we employed the ’Entity Relationship Construc- tion and Merging Strategies’ approach to con- solidate related named entities. Subsequently, ∗These authors contributed equally to this work. we computed the frequencies of sentence and entity occurrences. We extracted sparsely la- beled data and applied two techniques for data augmentation: GPT Paraphrase and entity re- placement while preserving sentence structure. These steps resulted in an augmented training set. Finally, we conducted fine-tuning experi- ments on various state-of-the-art BERT-based models to obtain a model suitable for the RO- CLING Shared Task. Keywords: GPT 3.5, Data augmentation, GPT paraphrase, Entity Relationship Con- struction and Merging Strategies 1 Introduction Named Entity Recognition (NER) aims to identify specific meaningful entities from text, such as person names, locations, organization names, dates, and times. In specific domains like the medical field, these named entities of- ten have unique naming conventions and char- acteristics. To accurately identify entities in these specialized domains, it’s common to use domain-specific training data to train NER models that cater to the named entity recog- 1 The 35th Conference on Computational Linguistics and Speech Processing (ROCLING 2023) Taipei City, Taiwan, October 20-21, 2023. The Association for Computational Linguistics and Chinese Language Processing nition requirements of that field. The main goal of this research is to develop and refine a Named Entity Recognition (NER) model fo- cused on the medical field, aiming to inves- tigate and improve its accuracy. The study involves various stages, including data prepro- cessing, model evaluation and selection, and experimentation with data augmentation tech- niques. In the data preprocessing phase, this in- volves data cleaning, entity relationship con- struction, and merging strategies. We discov- ered several issues in the data, such as non- compliance with BIO tagging standards and inconsistent entity labels within the same …

Content truncated to 500 words