ROBUST NATURAL LANGUAGE PROCESSING FOR URBAN TRIP PLANNING论文学习

研究内容

• In this article we present a Natural Language interface for trip planning in complex, multimodal, urban transportation networks.（在这篇文章中，我们提出了一个用于复杂、多模式的城市交通网络中的行程规划的自然语言接口。）
• Our objective is to provide robust understanding of complex requests while giving the user flexibility in their language.（我们的目标是在为用户提供语言灵活性的同时，为复杂的请求提供健壮的理解。）

技术方案

• We designed TRANQUYL, a transportation query language for trip planning.（我们设计了出行规划的交通查询语言TRANQUYL）
  • Our query structure builds on the standard “SELECT , FROM, WHERE” structure of SQL.（我们的查询结构建立在 SQL 的标准“SELECT、FROM、WHERE”结构之上。）
  • We retain the same base syntax and structure but extend it in two important ways.（我们保留了相同的基本语法和结构，但以两种重要的方式对其进行了扩展。）
    • to query trips we introduce an operator ALL_TRIPS（为了查询行程，我们引入了一个运算符 ALL_TRIPS）
    • we introduce four new clauses that allow further specification of the parameters of the trip.（我们引入了四个新的子句，允许进一步说明行程的参数）
• We developed a user-centric ontology, which defines the concepts covered by the interface and allows for a broad vocabulary. We utilize the ontology to infer the meaning of personal references and preferences so that they do not need to be explicitly stated for each request.（我们开发了一个以用户为中心的本体，它定义了接口所涵盖的概念，并允许使用广泛的词汇表。我们利用以用户为中心的本体来推断个人引用和偏好的含义，这样就不需要为每个请求显式地声明它们。）
• NL2TRANQUYL, the software system built on these foundations, translates English requests into formal TRANQUYL queries.（NL2TRANQUYL是建立在这些基础上的软件系统，它将英语请求转换成正式的TRANQUYL查询。）

系统架构

The translation from natural language to TRANQUYL occurs in several distinct steps.The four stages correspond to parsing, concept identification, concept attachment, and query generation.（从自然语言到TRANQUYL的转换需要几个不同的步骤，这四个阶段分别对应解析、概念识别、概念连接和查询生成。）

• It begins by parsing the input with the Stanford Parser in order to obtain both constituency and dependency parses.（它首先使用Stanford Parser解析输入，以获得选区解析和依赖解析。）
• In order to determine which concept in the ontology a specific relevant node n∈N corresponds to, two approaches are taken: (1) comparing nodes to concepts in the ontology (getOntologyConcepts), and (2) identifying specific concepts via regular expressions (getRegExConcepts).（为了确定特定相关节点n∈N对应本体中的哪个概念，采取两种方法：（1）将节点与本体中的概念进行比较（getOntologyConcepts），以及（2）通过正则表达式识别特定概念（getRegExConcepts ）。）
• We use the dependency parse and three sets of strategies, guided by the ontology, to generate a knowledge map. In general, there are three primary tasks in building the map: identifying personal references, aligning modifiers, and determining which data are missing.（我们使用依赖解析和三组策略，在本体的指导下，生成知识图谱。一般来说，构建地图有三个主要任务:识别个人引用、调整修饰符和确定丢失了哪些数据。）
• The final step is generating the TRANQUYL query using the knowledge map.（最后一步是使用知识图生成 TRANQUYL 查询。）

如何做实验

• formal evaluation: we present an evaluation of system performance on three sets of requests: well-formatted and grammatical requests as collected from external informants, grammatical paraphrases of the requests, telegraphic or fragmented requests.（我们给出了系统在三组请求上的性能评估:从外部信息者收集的格式良好和语法规范的请求(集合A)，请求的语法释义(集合B)，电报或碎片化请求(集合C)）
• informal evaluation: we solicited requests from laypersons with no experience with our research. This was done by posting a “note” on Facebook and asking random contacts to read the note and respond with requests.（我们向没有研究经验的外行人征求意见。这是通过在Facebook上发布一个“通知”，并让随机联系人阅读该通知并回应请求来实现的。）