Comparing Dictionaries and Word Embeddings

Student Team: Anuroop John Abraham, Archit Bose, Kartik, and Utkarsh Garg

Project Mentor: Ashvini Jindal, LinkedIn, Dr. Ananth Sankar, LinkedIn

An extensive amount of NLP research focuses on learning word representations that accurately reflect their semantic properties. We investigate whether these word embeddings capture the same semantics of the word as its dictionary definition/gloss. To accomplish this, we leverage a reverse dictionary word retrieval task (i.e., given a definition, we retrieve the corresponding word by learning the word’s embedding space). Extending the idea to multilingual representation learning, we show the possibility of retrieving a word in any target language given a definition in any source language when trained on a single high-resource language. Through comprehensive experiments in both monolingual and multilingual settings, we show:

1. How different model architectures and word embeddings (static vs. contextual) affect the gloss to word preformance.
2. How adapter networks compare to fine-tuning for learning word representations, especially in a multilingual context.
3. Which layers are best at mapping phrasal semantics (gloss) to lexical semantics (word embeddings).
4. How increased model parameters and dataset sizes monotonically improve the model’s representation learning ability.

To evaluate our system in this many-to-many setting, we release the first gold-standard multilingual parallel test of 1200 sense-aware word-gloss pairs in 6 languages.