Text2Structure - Structure Extraction Language
( (Spot Name: Info Span (Asso Name: Info Span) (Asso Name: Info Span) ) ) # Structure extrating language (SEL) for Universal IE
Spot Namerepresents there is a specific information piece with the type of spot name existing in the source text.
Asso Nameindicates there exists a specific information piece in the source text that is with the AssoName association to its upper-level Spotted information in the structure.
Info Spanrepresents the text span corresponding to the specific spotting or associating information piece in the source text.
Following is an example:
( (person: Steve (work for: Apple) ) (start-position: became (employee: Steve) (employer: Apple) (title: CEO) (time: 1997) ) (orgnization: Apple) (time: 1997) ) # The SEL representation for "Steve became CEO of Apple in 1997."
feature engineering -> neural network architechure engineering -> fine tuning -> prompt engineering
How to choose prompt
Different prompts have different zero-shot or few-shot capbilities. For example, the prompt for extrating person name could be:
- Which people are contained in the original text?
- Who are in the text?
- What are the names?
- Yaojie Lu, etc, from CAS, Baidu and BAAI. Unified Structure Generation for Universal Information Extraction., ACL 2022.
- PaddleNLP - UIE. 通用信息抽取 UIE(Universal Information Extraction)
- 通用信息抽取技术与产业应用实战, 2022.5.21
- 《UIE：基于统一结构生成的通用信息抽取》-韩先培, 2022.7.22