AASC is a corpus of natural language text extracted from scientific papers. It contains 2,339,195 sentences from PDF-format papers from the ACL Anthology [1], a comprehensive scientific paper repository on computational linguistics and natural language processing.
For PDF document analysis, we use PDFNLT 1.0 [2], a PDF paper analysis tool specifically trained for ACL Anthology. After excluding papers with non-standard structures (eg. no abstract, or no references), the rest 13,923 papers were further processed by (1) sentence splitting, and (2) section type labeling.
The ACL_2018_v2.tar.gz file contains the extracted natural language sentences for each <paper_ID>, where the <paper_ID> is the unique identifier of the paper on the ACL Anthology. The corresponding PDF version can be found using the URL:
http://aclweb.org/anthology/<paper_ID>.
Each sentence file is named as <paper_ID>.ss within which each line represents tab-separated values of a sentence:
| Column | Example (A00-1001.ss) |
|---|---|
| Sentence ID | s-1-1-0-0 |
| Section type | abstract |
| Sentence text: | The paper describes a natural language based expert system route advisor for the public bus transport in Trondheim, Norway. |
A simple dictionary-based classifier was used for the section type labeling.
For details, see also our Overview of AASC
- Following the copyright policy of the original ACL Anthology, AASC materials are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 International License.
- This work was supported by National Institute of Informatics and JST Crest JPMJCR1513.
