Repository for the course project of 263-5000-00L Computational Semantics for Natural Language Processing FS2022 at ETH.
Code and data for experiments in section 5 can be found under extrapolation.
Code and data for experiments in section 4 and 6 can be found under embedding_distribution.
Results for ablation models are under ablation.
Other branches in the repo are our attempts to improve SimKGC. Most of these results are not ideal and thus are not covered in the report. A brief overview of these branches:
similarity-metrics: Replacing cosine similarity with dot product or euclidean distance.negsamples: Attempt to design a commonsense-aware negative sampler similar to CAKE.cake-lossandloss-func: Attempt to design a commonsense-aware or a relation type-aware loss function.ensemble: Reverse the hr + t structure of SimKGC to investigate the backward performance.adapter: Incorporating concept prediction task through AdapterFusion.