NLP中的实体任务

• NER (Recognition): detecting the phrase that is the name of an entity 实体发现
• NEC (Classification): assigning an entity type to the phrase 实体类型分类
• NEL (Linking): establishing the identity of the entity in a given reference
  database (Wikipedia, DBpedia, YAGO) 实体链接
• Coreference: any phrase that makes reference to an entity instance, including
  pronouns, noun phrases, abbreviations, acronyms, etc… 实体指代

Named Entity Linking （NEL）

Problem statement:
Potentially ambiguous entity mention (“Paris”) needs to be linked to a canonical identifier/instance (http://dbpedia.org/resource/Paris) that fits the intended referent in the context of the text

entity linking (text -> KG)

Named Entity Recognition and Disambiguation (NERD)

假设: the mentions are already recognized in text

Combine recognition and disambiguation/linking -> NERD
NERC + NED = NERD

Knowledge bases

A catalog of things, usually entities. Each one has:
one or more names;
other attributes;
Connections to other entities
Textual description;

Knowledge bases are also connected to each other

Example:
Structured knowledge bases: DBpedia & Wikidata
Unstructured knowledge bases: Wikipedia

实体链接的好处

Benefits of connecting text and knowledge bases:
Automatic knowledge base construction (AKBC) / Knowledge base completion (KBC)帮助KB的完善

实体链接的挑战

1. name ambiguity: Entities with the same name
2. name variation: Different names for the same entity
3. Missing (NIL) entities

实体链接框架

Mention Detection (NER)

在文档中识别出潜在的实体

Candidate generation/selection

Balance between generating too many candidates (too much ‘noise’) and generating too little candidates (missing the correct one)
Trade-off between precision and recall -> an art by itself!
In practice, something like 30 candidates per mention is usually enough

如何选择top30?
commonness: for a given mention, how relatively often it refers to some instance in Wikipedia.
Also, observe dominance within a form and topical bias

Disambiguation

Goal: decide which of the candidates (or none) is the correct referent.

方法:

1. Word-based methods: DBpedia Spotlight
   Compute cosine similarity between the text paragraph with an entity mention and Wikipedia descriptions of each candidate

2. Graph-based methods: AIDA and AGDISTIS
   Construct a subgraph that contains all entity candidates with some facts from a KB, then find the best connected candidates per mention.

实体链接的评估

1. Assign a true positive (TP), false positive (FP), and/or false negative (FN) per mention occurrence
2. Count the TPs, FPs, and FNs across all mentions
3. Compute precision, recall, and F1-scores once on top of these