This page hosts resources related to our work on commonsense knowledge representation.

For details on these resources, please see Li et al. (2016).

Downloads:

• ConceptNet knowledge base completion task data:
  • training set of 100,000 tuples: train100k.txt.gz
  • training set of 300,000 tuples: train300k.txt.gz
  • training set of 600,000 tuples (conjugated and unconjugated forms): train600k.txt.gz
  • dev1 development set: dev1.txt.gz
  • dev2 development set: dev2.txt.gz
  • test set: test.txt.gz


• Demo for scoring arbitrary tuples (see examples below): ckbc-demo.tar.gz


• Automatically-generated ConceptNet/Wikipedia tuples scored with our model (Bilinear AVG, trained on 600k training set):
  • tuples created by automatically modifying ConceptNet training tuples: tuples.cn.txt.gz
  • tuples automatically extracted from Wikipedia sentences: tuples.wiki.tar.gz


• ConceptNet-trained word embeddings:
  • actual word embeddings: embeddings.txt.gz
  • data for training word embeddings: embedding_training.txt.gz
  
  
• Code for training tuple models is here on github

Examples:

The demo mentioned above can be used to compute the score of any arbitrary tuple.
For example:

python demo_bilinear.py drive_fast accident topfive
causes score: 0.992529380836
hassubevent score: 0.986317631007
hasprerequisite score: 0.248943002166
usedfor score: 0.160948880494
hasfirstsubevent score: 0.160748517657

python demo_bilinear.py gift birthday_party topfive
atlocation score: 0.940286085436
usedfor score: 0.802385884092
isa score: 0.439336382139
createdby score: 0.358644232972
motivatedbygoal score: 0.250502613254

python demo_bilinear.py hot summer sum
total score is: 11.0485060314
python demo_bilinear.py hot winter sum
total score is: 5.44850904225

python demo_bilinear.py eat_food feel_full max
causes score: 0.836753527212

References:

Xiang Li, Aynaz Taheri, Lifu Tu, and Kevin Gimpel. Commonsense Knowledge Base Completion. Proc. of ACL, 2016. (bib)