In this task, noun-compounds are annotated to a pre-defined set of relations, and the model has to predict the correct
relation between the constituents of an unobserved compound. For example, olive oil may belong to the SOURCE relation
while morning meeting belongs to the TIME relation.
We train 3 models which are tuned on the validation sets to select the number of predicted paraphrases k classifier type (logistic regression and SVM), and the regularization types and values. The models differ by:
- Distributional (baseline) - in which
[w1]and[w2]are represented by a concatenation of their 300d GloVe embeddings. - Paraphrase-based - in which we use our paraphrasing model to get a vector representing the relation between
[w1]and[w2]. - Integrated - a concatenation of both feature vectors detailed above.
usage: classifier.py [-h] [--use_w1_w2_embeddings] [--use_paraphrase_vectors]
paraphrase_model_dir word_embeddings_for_model
word_embeddings_for_dist dataset_prefix model_dir
positional arguments:
paraphrase_model_dir the path to the trained paraphrasing model
word_embeddings_for_model
word embeddings to be used for the language model
word_embeddings_for_dist
word embeddings to be used for w1 and w2 embeddings
dataset_prefix path to the train/test/val/rel data
model_dir where to store the result
optional arguments:
-h, --help show this help message and exit
--use_w1_w2_embeddings
use w1 and w2 word embeddings as features
--use_paraphrase_vectors
use the paraphrase vectors as features