| 摘要: | 命名實體鏈結 (NEL, Named Entity Linking) 是自然語言處理 (NLP,
Natural Language Processing) 的一項研究，在 NLP 中的研究中和應用
有著重要的作用，是不可或缺的一環，若能有效地提升 NEL 的準確性
的話就能更好的為開發高性能的 NLP 系統奠定基礎。
NEL 的主要挑戰是缺少帶標註的文本，在漢籍文本上尤為困難，
原因是因為古代人名時常出現重複的人名，使得註釋者除了必須會閱
讀漢籍文本之外也必須將每個候選人名的個人資料與文本的上下文做
比較，而使得研究人物的關係和社會網路更為困難，而本研究為了解
決此問題本篇提出了一套架構，除了上述問題之外也解決標註資料過
少的問題，該系統利用中國歷代人物傳記資料庫與中研院的人名權威
資料庫裡人名的履歷、時間、關係人等欄位自行產生訓練資料後再使
用 BERT 模型達成古人名的實體消歧與鏈結。
本研究以《明實錄》做為實驗文本，《明實錄》是中國明代官修的編
年體史書，該書中記錄了從明太祖朱元璋到明熹宗朱由校共十五代皇
帝，約兩百五十年的大量歷史文本，其中包含十三部，三千零五十五
卷，共計一千七百多萬字，而其中文本包含朝廷各院所呈繳之章奏、
批件等，並以各省官員收集的先朝紀錄作補充，逐年紀錄各個皇帝詔
赦、律令等，並含括了政治、經濟、文化、祭祀等大事而成。目前本
研究總共成功標註 8,787 個人名、257,302 個標籤，準確率 92.08%。;NEL plays an important role both in the study and application of NLP. If
the accuracy of NEL is effectively improved, the foundation of high-performance
NLP development can be laid.
The main challenge of NEL is the lack of annotated texts, especially in
studying Classical Chinese, because ancient names often appear repeatedly,
which makes it difficult to study the historical figures relationships and their
social networks. Our system used the China Biographical Database Project
(CBDB) and Ming Qing Biographical Database to generate training data and
then uses BERT model to eliminate the physical disambiguation of the names.
This study took the Ming Shilu as the experiment text. The Ming Shilu is
an official chronological history book of the Ming Dynasty in China, chroni-
cling 15 generations emperors, from Zhu Yuan-Zhang to Zhu You-Jiao, cov-
ering about 250 years. There is over 17 million characters including 30,055
volumes and 13 parts in the Ming Shilu. The text records the imperial pardons
and laws of each emperor as well as political, economic, cultural, and ritual
events year by year, including the imperial decrees and approvals submitted
by the imperial ministries, and the records of previous dynasties collected by
the provincial officials.
8,787 names and 257,302 tags were successfully tagged in this study, with
92.08% accuracy. |
