idopinto/co-specter-biomed-recon

Model Card for Model ID

SPECTER-CoCite model included in a paper for modeling fine-grained similarity between documents in the biomedical domain.

Title: "Multi-Vector Models with Textual Guidance for Fine-Grained Scientific Document Similarity"

Authors: Sheshera Mysore, Arman Cohan, Tom Hope

Paper: https://arxiv.org/abs/2111.08366

Github: https://github.com/allenai/aspire

Note: In the context of the paper, this model is referred to as Specter-CoCite_Spec and represents a baseline bi-encoder for scientific document similarity. This model is similar in architecture to the allenai/specter model but is trained on co-citation data instead of citation data.

Refer to https://huggingface.co/allenai/aspire-biencoder-biomed-spec for more details and usage.

Evaluation Results

Differences might be in co-citations training data which constructed from scratch from different release of S2ORC (originaly, 2019-09-28, which I didn't have access to.)

Model	TRECCOVID-MAP	TRECCOVID-NDCG%20	RELISH-MAP	RELISH-NDCG%20
specter	28.24	59.28	60.62	77.20
aspire-biencoder-biomed-spec	26.07	54.89	61.47	78.34
aspire-biencoder-biomed-spec-full	28.87	60.47	61.69	78.22
aspire-biencoder-biomed-spec-recon	27.73	59.18	60.36	77.09
aspire-biencoder-biomed-spec2	29.38	61.45	60.26	77.25