基於統計方法之中文搭配詞自動擷取

本研究採取以下四個步驟擷取出雙連詞、三連詞、四連詞之詞彙或詞性組合之搭配詞。首先採用Smadja's Xtract的平均數及變異數的方法，擷取具有變動距離模式所共同出現的詞彙或詞性的組合，接著使用搭配詞顯著性的衡量方法：相互資訊值及T檢定值。通過以上檢驗的候選搭配詞，經由對照中央研究院詞義標示語料庫之目標詞的結果，在同樣的跨距下，若同為一個詞義者，則我們以此搭配詞作為詞義標示知識。並且，本研究將產出之搭配資訊應用於詞義自動標示處理，達到20.07%的應用率及90.83%的正確率。

We take the four following steps to extract collocations made of combinations of 2, 3, 4 words and/or part of speech, espectively. First, we use 'Smadja's Xtract' to extract theco-occurrence combinations of words and/or part of speech of varying distance by computing means and variances. Second, we evaluate the significances of collocation candidates by 2 metrics: mutual information and t-test value. At last, we compare the head words of tagged word sense corpus made by Academic Sinica with the collocation candidates. If in the same distance, the head words of collocation candidates match the ones made by Academic Sinica, we say they are collocations. In addition, we apply the collocation information produced from this research to word sense disambiguation. It reaches application rate of 20.07% and precision rate of 90.83%.