共指消解

共指消解——Coreference Resolution

目录

• 共指消解——Coreference Resolution
  • Coreference Resolution
  • Mention Detection
  • Deal with bad mentions
  • Avoid traditional pipeline system
  • Coreference Models

Coreference Resolution

Coreference Resolution: find all the words that refer to the same entity in the world.

antecedent / anaphor: 先行词与回指。a term (anaphor) refers to another term(antecedent). For example, Pronouns! But not all noun phrases have reference (like everyone)

Cataphor: 后行词与后指。antecedent出现在后面

Mention Detection

• Pronouns

  Part-of-speech(POS) tagger

• Named entities

  named entity recognition(NER) system

• Noun phrases

  use a parser

Deal with bad mentions

• train classifier to filter out spurious mentions
• Keep all mentions as candidate mentions and in final discard all singleton mentions (not have coreference with anything else)

Avoid traditional pipeline system

• train classifier specifically for mention detection
• end-to-end model (Mention detection and co-reference resolution simultaneously)

Coreference Models

• rule-based (Hobbs Algorithm, make a tree)

• mention pair

• mention ranking

• clustering

• Neural Coreference Model

  1 D Convolution feature.

  Character-level convolutional feature and word2vec , sliding spans as candidate inputs

  Learn Character-level representations for POS tagging, BERT