Ner's past, present and future Overview - Future

background

Named entity recognition （NER, Named Entity Recognition）, It refers to the recognition of entities with specific meaning in the text , Mainly including people's names 、 Place names 、 Organization name 、 Proper nouns, etc .

About NER Past 、 Now you can refer to NER Past 、 Present and future overview - Now? NER Past 、 Present and future overview - In the past

This article is about NER The future is at a point in time 2021 Some development points in the future

• Few-show & zero shot. How to pass a few samples , Even zero samples can get a good model , For example, how to introduce regular expression templates 、prompt Methods such as .
  • Template-Based Named Entity Recognition Using BART
• Integrating knowledge . Before that Present article Some have been mentioned , future , As the pre training model gets larger , If you can strip knowledge from it , Use smaller language models to speed up training . Then through other ways to integrate knowledge , For example, how to search , image DeepMind Of RETRO and OpenAI Of WebGPT
• The migration study . This may be a little big , How to use the language model to learn . Why can people recognize the entities in them , With the migration of past experience , To draw inferences from one instance ; Grammatical information （ Sentence pattern, etc ）; Specific sentence patterns ; Imitation learning, etc .
• Decoding method . Personally feel span、 classification 、 Sequence tagging doesn't seem to be perfect ,span The method does not consider the dependencies between the overall sequence labels ; For classification, the length of the entity should also be considered , In practice, the length of an entity can be any length （ It makes sense that extreme cases exist ）; Sequence annotation can not solve the nesting problem well . At present, there are some methods of combination , for example Span+ Fragment arrangement 、BIO+ classification , But there is still room for optimization .

Personal view , For reference only .

Reference

• Template-Based Named Entity Recognition Using BART
• Improving Language Models by Retrieving from Trillions of Tokens
• WebGPT: Browser-assisted question-answering with human feedback
• Boundary Enhanced Neural Span Classification for Nested Named Entity Recognition
• https://aclanthology.org/D19-1034/