Improving Low Compute Language Modeling with In-Domain Embedding Initialisation
This addresses the challenge of efficient language modeling for specific domains like biomedical data and technical support, though it is incremental as it builds on existing embedding methods.
The paper tackles the problem of training language models with limited computational resources and 10-100 million tokens of in-domain data, showing that initializing and freezing input embeddings using in-domain data improves performance by better representing rare words, with gains observed across multiple domains.
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 byte-pair 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.