漢語構式語意自動分類研究

Abstract

由多義引起的語義歧義常阻礙機器對於人類語言的理解。因此，詞義消歧（WSD）一直是自然語言處理中重要的任務。除了字詞，作為形式-意義配對的構式同樣具有多義特徵。本論文旨在將詞義消歧任務從單詞層次擴展到構式層次，並利用語料庫與計算語言學方法研究構式的多義性。為了了解空槽和構式之間的相互作用，我們運用搭配分析（collostruction analysis）將兩者的吸引程度進行量化。隨後，本研究鎖定十二個多義構式，並標記構式中兩個空槽間的語義關係。透過標記資料，我們分析了空槽間語義關係與構式語義的配對。最後，本研究使用 3×2 因素之實驗設計，以探討各模型預測空槽間語義關係之能力。研究結果顯示，模型能夠區分空槽間的語義關係，代表其具有分類構式語義的能力。在本論文中，帶有弱監督線索的 BERT 模型可以達到 80% 的預測準確度，說明弱監督線索和上下文脈絡有助於提高預測構式語義的準確度。

Semantic ambiguity arising from polysemy has hindered the machine from thoroughly understanding human language. Thus, Word Sense Disambiguation (WSD) has always been a significant task in natural language processing. Similar to words, constructions, which are conventionalized form-meaning pairings, also possess the polysemous trait. Therefore, this thesis aims to extend the disambiguation task from word-level to construction-level. This thesis investigated the polysemy of constructions with corpus-based and computational approaches. To understand the interaction between slots and constructions, a collostruction analysis was conducted. Subsequently, examples of 12 potentially polysemous construction forms were annotated with the semantic relation between two open slots X and Y. Afterwards, the mapping between X/Y relations and construction senses was examined directly from the annotated data. Finally, a 3×2 experiment was employed to investigate the predictive abilities of X/Y relations of different models. The results showed that models could distinguish between X/Y relations, which implied that models could classify the construction senses. The best performance was achieved by BERT with weak supervision signals, reaching an accuracy of 0.8. It can be concluded that weak supervision signals and contextual information can help enhance prediction accuracy.