研究内容

Given an RDF graph G and a natural language question qNL, our goal is to interpret qNL as a SPARQL query qS, by mapping the semantic items — relations(i.e. properties), entities and classes expressed by qNL to the corresponding triple patterns inqS.（给定一个 RDF 图 G 和一个自然语言问题 qNL，我们的目标是将 qNL 解释为一个 SPARQL 查询 qS，通过将 qNL 表示的语义项目——关系（即属性）、实体和类映射到相应的三重模式 inqS。）

系统架构

the system framework is separated into offline and online two parts.

• Offline Processing: To enable semantic relation extraction in online processing, a dictionary of the RDF relations and their natural language paraphrases is automatically built in advance.（离线处理：为了能够在在线处理中提取语义关系，预先自动构建 RDF 关系及其自然语言释义的字典。）
• During the online stage, the input question qNL is fed into the following four-step pipeline:（在线阶段，输入问题 qNL 被送入以下四步管道：） 
  1. qNL is parsed into a dependency tree tNL.（qNL 被解析为依赖树 tNL）
  2. the phrases in qNL that mention any semantic relation are recognized in tNL with the help of the paraphrase dictionary. （在释义词典的帮助下，在 tNL 中可以识别 qNL 中提到任何语义关系的短语。）
  3. these phrases are mapped to the RDF fragments to find their matches of semantic items in the RDF graph G.（将这些短语映射到 RDF 片段以在 RDF 图 G 中找到它们的语义项匹配。）
  4. the RDF fragments are joined to compute a reasonable subgraph of graph G, by checking their compatibility based on the dependency tree tNL.（通过基于依赖树 tNL 检查它们的兼容性，连接 RDF 片段以计算图 G 的合理子图。）
  5. The results are ranked according to the score of semantic similarity and coherence, leading to the target SPARQL.（结果根据语义相似性和连贯性的分数进行排名，从而得出目标 SPARQL。）

如何做实验

The system is implemented on DBpedia dataset.（该系统是在 DBpedia 数据集上实现的。）