利用雙重註釋編碼器於中文健康照護實體連結;Leveraging Dual Gloss Encoders in Chinese Healthcare Entity Linking

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/90047

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/90047

Title:	利用雙重註釋編碼器於中文健康照護實體連結;Leveraging Dual Gloss Encoders in Chinese Healthcare Entity Linking
Authors:	洪滿珍;Hung, Man-Chen
Contributors:	電機工程學系
Keywords:	實體連結;詞義消歧;語言轉譯器;自然語言理解;健康資訊學;entity linking;word sense disambiguation;language transformers;natural language understanding;health informatics
Date:	2022-08-25
Issue Date:	2022-10-04 12:09:01 (UTC+8)
Publisher:	國立中央大學
Abstract:	詞義消歧是自然語言理解的一項重要且艱難的任務，尤其是對於醫療領域經常有多種語義含意的詞彙。我們提出一個雙重註釋編碼器 (Dual Gloss Encoders, DGE) 模型，以 BERT 轉譯器為基礎，將中文句子中健康照護領域的命名實體，連結到多國語言詞彙語義網路 BabelNet，以實現上下文感知語義理解。消歧目標詞的每個註釋都源自 BabelNet，我們將原句嵌入註釋得到語境化的目標詞嵌入向量，目標詞嵌入向量再和每個註釋嵌入配對，以計算語義消歧的分數，將分數最高者選為語句中消歧目標詞的註釋選項。由於在健康照護領域缺乏中文實體連結數據，我們收集了適當的領域單詞並在句子中手動標記它們的註釋。最後，我們總共有 10,218 個句子，包含 40 個不同的消歧目標詞和 94 個不同的語義註釋。我們將建構的數據劃分為訓練集 7,109 筆、發展集 979 筆與測試集 2,130 筆。實驗結果表明，我們提出的 DGE 模型的性能優於三個實體連結模型，即 BERTWSD、GlossBERT 與 BEM，獲得了 F1-Score 97.81%。;Word sense disambiguation is an important and difficult task for natural language understanding, especially for those lexical words with many semantic meanings in the healthcare domain. We propose a BERT transformer based Dual Gloss Encoder (DGE) model to link Chinese healthcare entities to the multi-lingual lexical network BabelNet for contextaware semantic understanding. The target word along with its context in original sentence is encoded to obtain embedding vector. Each gloss of the target word is originated from BabelNet to encode the gloss embedding. Target word embedding and each gloss embedding will be paired to calculate the scores for sense disambiguation. The gloss with the highest score is returned as predicted gloss for the target word in a given sentence. Due to a lack of Chinese entity linking data in the healthcare domain, we collected proper domain-specific words and manually annotated their glosses in the sentence. Finally, we have a total of 10,218 sentences containing 40 distinct target words with 94 various semantic glosses. Our constructed data was divided into three mutually exclusive datasets, including training set (7,109 sentences), development set (979 sentences), and test set (2,130 sentences). Experimental results indicate that our proposed DGE model performs better than three entity linking models, i.e., BERTWSD, GlossBERT and BEM, obtaining the best F1-score of 97.81%.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	56	View/Open

社群 sharing

Loading...