KB is a collection of records in database. Records are often stored as triples in the form subject predicate object.
Collections of records / triples from difference sources with different naming schemes are counted as Combined Data, which is discussed in Section 2.1.3.
Freebase里的数据格式:
一些常用的KB数据库:
Data Formats:
KB数据通常保存为RDF的格式。序列化后,可以用下面的形式保存:
N-triples
N-quads
TSV
Turtle
2.1.3 Combined Data
可以将Text和KB进行结合,也可以将多个KB进行结合。
combined data的两个原则:
link: link a text to a knowledge base by recognizing mentions of entities from the knowledge base in the text and linking to them。这里涉及到NER和disambiguation的任务。
mult: combine multiple knowledge bases with different naming schemes (such that the same entity or relation may exist with different names)
Commonly Used Datasets:
Text Linked to a Knowledge Base:
将link边骂道text中的格式是XML。下面是Wikipedia LOD的一个例子:
Semantic Web:
The data from the Semantic Web is often also called linked open data (LOD), because contents can be contributed and interlinked by anyone, just like web pages (but in a different format, see below). 我们将这种数据叫作semantic web data,它符合combining data的两个原则:
link: provided by semantic markup for ordinary web pages.
mult: anyone can contribute + absence of a global schema.
Summary:
这是一篇关于语义检索的survey,详细介绍了关于语义检索的各种范式,设计到的NLP任务,以及基于Text和基于KB的不同方法。
FTIR: Foundations and Trends in Information Retrieval
Resource:
Paper information:
Notes:
2 Classification by Data Type and Search Paradigm
2.1 Data Types and Common Datasets
2.1.2 Structured Data / Knowledge Bases
KB is a collection of records in database. Records are often stored as triples in the form subject predicate object.
Collections of records / triples from difference sources with different naming schemes are counted as Combined Data, which is discussed in Section 2.1.3.
Freebase里的数据格式:
一些常用的KB数据库:
Data Formats:
KB数据通常保存为RDF的格式。序列化后,可以用下面的形式保存:
2.1.3 Combined Data
可以将Text和KB进行结合,也可以将多个KB进行结合。
combined data的两个原则:
Commonly Used Datasets:
Text Linked to a Knowledge Base:
将link边骂道text中的格式是XML。下面是Wikipedia LOD的一个例子:
Semantic Web:
The data from the Semantic Web is often also called linked open data (LOD), because contents can be contributed and interlinked by anyone, just like web pages (but in a different format, see below). 我们将这种数据叫作semantic web data,它符合combining data的两个原则:
其中“mult”原则可以被RDF实现,可以把任意的文本链接到RDF的数据上。比如下面关于法国城市Emburn的例子。主要前缀 rdf: 和 gn: 可以让内容更紧凑,方便我们阅读:
“link”原则可以通过semantic markup来实现。比如下面是一个HTML页面,使用了Microdata markup:
下面是4个最常用的semantic markup。前3个直接使用HTML tags。而JSON-LD比其他3个的优势在于,可以更清晰地分割开ordinary content和semantic content:
2.2 Search Paradigms
2.2.1 Keyword Search
2.2.2 Structured Search
2.2.3 Natural Language Search
3 Basic NLP Tasks in Semantic Search
TODO
4 Approaches and Systems for Semantic Search
4.2 Structured Search in Knowledge Bases
4.2.1 Basic Techniques
两种保存KB的方法:保存到standard relational database management system (RDBMS),或者triple store。后者在4.2.2部分介绍。
如果是保存到RDBMS,query是SPARQL,那么query可以被转换为SQL queries.
Performance:
Dedicated triples stores 比 RDBMS 有优势。比如可以使用针对triple数据特化的index data structure。但是,RDBMS在处理complex query方面也有优势。
4.2.2 Systems
到2016年为止,3个最常用的系统是:Virtuoso,Jena, Sesame。
4.3 Structured Data Extraction from Text