Abstract
在本論文中，我們提出了一個結合解釋導向學習法(EBL) 和相似導向學習 法(SBL) 的語言習得模式。此外，我們提出了兩個習得問題來測試這個模 式。首先是詞典規則的習得，此處以習得動詞的子類(subcat)做為例子。 接著，設法習得介詞片語的附著傾向。本系統由領域理論(domain theory)，EBL 剖析器和SBL 模組所構成。領域理論包括一般化片語結構 文法(GPSG) 所提出的通用語言原則，本文無關的文法規則及一個含語法 特徵的字典。SBL 模組由一個簡單的語形處理單元及一個以 EBL 剖析器 輸出值為統計資料的統計單元。若EBL 剖析器之剖析失敗起因於某些字的 知識不全，系統將呼叫 SBL 模組，根據相關句子的剖析結果，去統計並 預測不全的部分。最後結果還需 EBL 剖析器的剖析驗證。若系統遭遇介 詞片語的附著問題， EBL 模組亦可統計 EBL 剖析器的剖析結果去預測介 詞片語的附著傾向。最後，上述的習得問題將透過實驗結果驗證之。 In this paper, we propose a language acquisition model which combines the Explanation-Based Learning(EBL) approach and the Similarity-Based Learning(SBL) approach. We illustrate with two kinds of acquisition problems to test this model. First, acquire and refine lexicon rules of corresponding lexical heads and generate suggestions if there are missing general grammar rules. The acquisition of unknown subcats of verbal lexicon entries will serve as an example. Second, acquire semantic preferences for prepositional phrase(PP) attachment. The system is composed of domain theory, EBL paser and SBL module. The domain theory in EBL consists of the universal linguistic principles proposed in the Generalized Phrase Structure Grammar( GPSG) formalism, the Context-Free grammar rules as well as a dictionary full of lexicon entries with syntactic/thematic features. The SBL module consists of a simple morphological processing unit and a statistical unit which analyze the partial parsing results from EBL parser. In parsing, if the EBL parser fails to parse a given sentence due to the incompleteness of relevant lexicon entries, the SBL module will be invoked to acquire the lexicon knowledge(lexicon rules with lexicon features). If the acquired lexicon knowledge can correctly complete the parsing, the parsing is terminated. Otherwise, no acquisition action is activated. Since the domain theory is assumed to contain no semantic information at the beginning of acquisition stage, the SBL module will be invoked to collect the partial results produced by the EBL parser to predict the implicitly implied semantic information if ambiguities such as PP attachment, are encountered by the EBL parser during parsing. The effects of the acquisition are verified by some experiments.