TransMORE: 基於翻譯的多語言開放式關係抽取

Abstract

開放式關係抽取 (Open Relation Extraction Systems, ORE) 系統從句 子中找出關係對。與傳統關係抽取不同，「開放式」意指不事先制定 (pre-define) 欲抽取的關係種類。 現存的所有英文 ORE 系統的設計都與英文文法或詞性 (part-of- speech) 高度相關，因此不容易直接應用於別種語言。而對於目前非英 文 ORE 系統，大部分基於英文 ORE 系統提出的概念，適度的修改以 合乎其他語言，此方式往往需要重新訓練資料或甚至重新設計系統。 有鑑於此，本論文提出一套基於翻譯器 (translator) 的多語言 (multilingual) ORE 系統，TransMORE，能應用於任何能翻譯至英文的自 然語言。由於 TransMORE 使用了第三方的翻譯器與英文 ORE 系統， 隨著翻譯器或英文 ORE 系統的進步，TransMORE 能同時與之進步， 毋需重新設計或再訓練任何資料。 在本研究中，除了與先前作品比較外，我們也實驗了使用不同英 文 ORE 系統對產出的影響，同時展現 TransMORE 易於抽換第三方 ORE 系統的特性。

Open domain relation extraction (ORE) systems identify relation and arguments phrases in a sentence, without any pre-defined underlying schema. Most English ORE systems, including current state-of-the-art system, can only extract relations from English because their methods highly rely on lingustic grammar and/or part-of-speech tagging. For non-English ORE systems, most of them use ideas from English ORE systems, and redesign the language-dependent rules to fit their target language. This thesis presents a new multilingual ORE system, TransMORE, which uses a novel method for extracting relations from multilingual corpus. TransMORE bases on a third-party translator and English ORE system. With better performance translator and/or English ORE system, TransMORE can have better result as well. In this work, besides compare with previous work, we also experiment on TransMORE within different English ORE systems.