Improving Low Compute Language Modeling with In-Domain Embedding Initialisation

Abstract

Many NLP applications, such as biomedical data and technical support, have 10-100 million tokens of in-domain data and limited computational resources for learning from it. How should we train a language model in this scenario? Most language modeling research considers either a small dataset with a closed vocabulary (like the standard 1 million token Penn Treebank), or the whole web with bytepair encoding. We show that for our target setting in English, initialising and freezing input embeddings using in-domain data can improve language model performance by providing a useful representation of rare words, and this pattern holds across several different domains. In the process, we show that the standard convention of tying input and output embeddings does not improve perplexity when initializing with embeddings trained on in-domain data.

North Korean Defectors and Human Resource Development in South Korea

“…For the input embeddings of the newly added tokens, we adopt the approach of using the average embeddings of the subword tokens that make up these new tokens as in (Hewitt, 2021;Welch et al, 2020). This method utilizes the semantic richness of the model's existing subword embeddings to offer a meaningful starting point for the new tokens' representations.…”

Compositional Demographic Word Embeddings

“…We explored various hyperparameter configurations on our validation set and found the best results using dropout with the same mask for generic and demographic-specific embeddings, untied weights, and fixed input embeddings. Untying and fixing input embeddings is supported by concurrent work (Welch et al, 2020b). Each model is trained for 50 epochs.…”