Abstract:
LLM-based representation learning is widely used to build effective information retrieval systems, including scientific domains. For making science more open and affordable, it is important that these systems support multilingual (and cross-lingual) search and do not require significant computational power. To address this we propose SciRus-tiny, light multilingual encoder trained from scratch on 44M abstracts (15B tokens) of research papers and then tuned in a contrastive manner using citation data. SciRus-tiny outperforms SciNCL, English-only SOTA-model for scientific texts, on 13/24 tasks, achieving SOTA on 7, from SciRepEval benchmark. Furthermore, SciRus-tiny is much more effective than SciNCL: it is almost 5x smaller (23M parameters vs 110M), having approximately 2x smaller embeddings (312 vs 768) and 2x bigger context length (1024 vs 512). In addition to the tiny model, we also propose the SciRus-small (61M parameters and 768 embeddings size), which is more powerful and can be used for complicated downstream tasks. We further study different ways of contrastive pre-training and demonstrate that almost SOTA results can be achieved without citation information, operating with only title-abstract pairs.
Keywords:information retrieval, interpretability and analysis of NLP models, large language models, representation learning.
