A Latent Variable Model for Learning Distributional Relation Vectors

Authors: Jose Camacho-Collados, Luis Espinosa-Anke, Shoaib Jameel, Steven Schockaert

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that our relation vectors outperform those from existing methods.
Researcher Affiliation Academia 1School of Computer Science and Informatics, Cardiff University, United Kingdom 2School of Computing, University of Kent, United Kingdom
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes These pre-trained relation embeddings, along with the code to generate them, are available at https://github.com/pedrada88/relative.
Open Datasets Yes To learn the relation vectors we use the English Wikipedia dump of January 2018, as in Joshi et al. [2019]. We evaluate our model on graded lexical entailment using the Hyper Lex dataset [Vuli c et al., 2017]. As test sets we used Diff Vec [Vylomova et al., 2016] and BLESS [Baroni and Lenci, 2011]. We specifically used the following standard datasets11: (1) 20news [Lang, 1995]; (2) reuters [Lewis et al., 2004]; (3) bbc12 [Greene and Cunningham, 2006]; and (4) ohsumed.
Dataset Splits Yes For both protocols, training and test partitions are available. (Hyper Lex) BLESS includes noun-noun relations such as hypernymy, meronymy, and co-hyponymy, including 13,258 and 6,629 instances for training and testing, respectively. (BLESS) using 10-fold cross-validation in the case of Diff Vec. For bbc, which does not include train-test splits, we performed 10-fold cross-validation.
Hardware Specification Yes learning relation vectors for all pairs in the vocabulary took around a day on a standard desktop computer on CPU. The Titan Xp used for this research was donated by the NVIDIA Corporation.
Software Dependencies No The paper mentions '300-dimensional Fast Text word embeddings [Bojanowski et al., 2017]' and 'Model developed in Keras: https://github.com/fchollet/keras.' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes The number of iterations of the EM algorithm (see Section 3.2) is set to three, which we empirically found sufficient for the model to converge.