Abstract:

In order to find near-synonyms/synonyms phrases in large corpus, a near-synonym/synonym phrase finding model based on common surrounding context is proposed in this paper. It captures semantic similarity via [n]-gram distribu-
ted method, and implicitly preserves local syntactic structure without parsing, making the underlying method language independent. The specific implementation is divided into two phases：The first phase is context collection and filtering, that is, it uses the local contexts surrounding the query phrase as features to the conditional model to capture both semantic and syntactic information. The second phase is the collection and screening of candidate phrases, that is, it iterates over all the instances of each “left”, “right” and “pairing” in the data to collect a set of near-synonym/synonym candidate phrases. And the elements that make up the model and the scoring functions used to evaluate the performance of the model are also given. The experimental results based on different large corpus show that the proposed modeling method is superior to other common finding method models in terms of statistical scoring finding performance and overall scalability.

Key words: near-synonyms/synonyms, query phrases, semantic similarity, context, scoring function

摘要：

为了实现大型语料库中近义词/同义词短语的查找，提出了一种基于共同语境的近义词/同义词短语查找模型，它通过[n]-gram分布式方法捕获语义相似性，不需要解析就能隐式地保存局部句法结构，使底层方法语言独立；具体实现分为两个阶段：第一阶段是上下文收集和过滤，即用围绕查询短语的本地上下文作为条件模型的特征来捕获语义和语法信息。第二阶段是候选词短语收集和筛选，即对数据中的每个“左”“右”和“配对”的全部实例进行迭代，以收集一组近义词/同义词候选短语；还给出了构成模型的要素和用于评价模型性能的评分函数；基于不同大型语料库的实验结果表明，提出的建模方法在总的统计评分查找性能和整体可扩展性方面都优于常用的其他查找方法模型。

关键词: 近义词/同义词, 查询短语, 语义相似性, 上下文, 评分函数