concreteness

Models to estimate the concreteness of a word on a scale from 1 to 5 based on its fasttext embedding. Trained on the cc.nl.300.bin Dutch fasttext embeddings. Only use on nouns, verbs and adjectives; also works on unknown words.

Repository contains two models: svr and lgbm. The svr is slightly more accurate but relatively slow, lgbm is less accurate but faster.

Dependencies:

• scikit-learn
• LightGBM

Evaluation

	R2	MAE	r (test)	r (train)
SVR	0.74	0.39	0.86	0.95
LGBM	0.69	0.44	0.83	0.89
LR (Thompson & Lupyan, 2018)				0.8

Example usage

from concreteness import WordConcreteness

wc = WordConcreteness(model='svr')

wc.score('boek')
## 4.83339

wc.score('coronavirus')
## 3.08055

wc.score('ideologie')
## 1.4905

Or, if you don't want bother with this tiny wrapper, just use the models directly:

import pickle
import fasttext

svr = pickle.load(open('models/svr.p', 'rb'))
embeddings =  fasttext.load_model('cc.nl.300.bin')

word = 'boek'
word_embedding = embeddings.get_word_vector(word)

prediction = svr.predict([word_embedding])[0]
prediction
## 4.83339