Method: Implicit Tensor Factorization
Learning Embeddings for Transitive Verb Disambiguationby Implicit Tensor Factorization
Qualitative Evaluation: Nearest Neighbor Phrases & Multiple Meanings in Verb Matrices
Conclusions Adjuncts are very helpful in learning the meaning of transitive
verb phrases Tensor-based joint learning method is effective in capturing
the changes of the meaning of transitive verb phrases
Future Work Incorporating other informative contextual information, such as adjective-
noun relationships Applying our SVO embeddings to real-world NLP applications, such as
semantic search engines
Overview: Transitive Verb Sense Disambiguation using Embeddings and Matrices
CVSC 2015
Quantitative Evaluation: Results on Transitive Verb Sense Disambiguation Task
Kazuma Hashimoto and Yoshimasa Tsuruoka (University of Tokyo){hassy, tsuruoka} @logos.t.u-tokyo.ac.jp
A plausibility score for each tuple (p, 𝑎1, 𝑎2)An observed tuple Collapsed tuples
Jointly learning the composition function
Noun embeddings
Verb matrix
Cost function
Model parameters(randomly initialized)
Backpropagation
English Wikipedia corpus SVO: 23.6 million instances SVOPN: 17.3 million instances
Co-occurrence statistics of phrases!
Using Predicate-argument structures (Hashimoto et al., 2014)(Relationships between predicates and their arguments (phrases))
Extracting Subject-Verb-Object (SVO) tuples SVO-Preposition-Noun (SVOPN) tuples
Why prepositional adjuncts?Prepositional adjuncts complement the meaning of verb phrases!
Enju parser(Miyao et al., 2008)
Nearest neighbor verb-object phrases
make moneymake cash, make dollar, make profit,
earn baht, earn pound, earn billion
make paymentmake loan, make repayment, pay fine,
pay amount, pay surcharge, pay reimbursement
make use (of)use number, use concept, use approach,
use method, use model, use one
Biomedical domain
Computer science
Computer science
Management
Measuring semantic similarities between pairs of transitive verbs taking the same subjects and objects (Grefenstette et al., 2011)
Verb pair
with the same argumentsHuman rating
student write name
student spell name7
child show sign
child express sign6
system meet criterion
system visit criterion1
Evaluation:Spearman’s rank correlation between human ratings and the cosine similarity scores produced by the SVO embeddings
MethodSpearman’s rank
correlation score
This work (only SVO data) 0.480
This work (SVO and SVOPN data) 0.614
Tensor-based method (Milajevs et al., 2014) 0.456
Joint learning method (Hashimoto et al., 2014) 0.422
Adjuncts improve the score!
Finally, 50-dimensional embeddings are used