TEASE is an algorithm for Web-based acquisition of entailment relations. Given a lexical-syntactic input template, a parse sub-tree with variable slots, the algorithm automatically learns other templates that are candidates for entailment relation with the input template. The direction of the entailment relation is not learned in this version, so the resulting relation can be either that the input entails the candidate, the candidate entails the input or both entail each other (paraphrases). The structure of the candidates is also learned as part of the acquisition.
The current knowledge collection available consists of 136 different templates that were given as input. Under each directory you can find two files: pivot.xml, containing the description of the input template, and learned_templates.xml, containing the description of all the learned templates for that input template. The current collection was created by Idan Szpektor at Bar Ilan University.
Acquiring the Resource
The current TEASE knowledge collection is available for research purposes: TEASE_collection.zip
Note that in the current knowledge collection, the first candidate in the list of learned templates may be at times the input template itself. This will be filtered out in future knowledgebase versions.
Please refer to the following publication when using this resource:
- Idan Szpektor, Hristo Tanev, Ido Dagan and Bonaventura Coppola. 2004. Scaling Web-based Acquisition of Entailment Relations. in Proceedings of EMNLP 2004.