Models based on the transformer architecture, such as BERT, have marked a crucial step forward in the field of Natural Language Processing. Importantly, they allow the creation of word embeddings that capture important semantic information about words in context. However, as single entities, these embeddings are difficult to interpret and the models used to create them have been described as opaque. Binder and colleagues proposed an intuitive embedding space where each dimension is based on one of 65 core semantic features. Unfortunately, the space only exists for a small data-set of 535 words, limiting its uses. Previous work (Utsumi, 2018, 2020; Turton et al., 2020) has shown that Binder features can be derived from static embeddings and successfully extrapolated to a large new vocabulary. Taking the next step, this paper demonstrates that Binder features can be derived from the BERT embedding space. This provides two things; (1) semantic feature values derived from contextualised word embeddings and (2) insights into how semantic features are represented across the different layers of the BERT model.

Word embeddings such as Word2Vec not only uniquely identify words but also encode important semantic information about them. However, as single entities they are difficult to interpret and their individual dimensions do not have obvious meanings. A more intuitive and interpretable feature space based on neural representations of words was presented by Binder and colleagues (2016) but is only available for a very limited vocabulary. Previous research (Utsumi, 2018) indicates that Binder features can be predicted for words from their embedding vectors (such as Word2Vec), but only looked at the original Binder vocabulary. This paper aimed to demonstrate that Binder features can effectively be predicted for a large number of new words and that the predicted values are sensible. The results supported this, showing that correlations between predicted feature values were consistent with those in the original Binder dataset. Additionally, vectors of predicted values performed comparatively to established embedding models in tests of word-pair semantic similarity. Being able to predict Binder feature space vectors for any number of new words opens up many uses not possible with the original vocabulary size.