Ophthalmic Disease Zero-Shot Question Answering Through Knowledge Triple Augmented Language Model Prompting

Ophthalmic disease zero-shot question answering aims to answer questions about ophthalmic disease in natural language without any model training. Although the pretrained large language models (LLMs) have stored quantities of general knowledge of the real-world and made zero-shot question answering come true, they ususally have insufficient ophthalmic knowledge, and fine-tuning LLMs to update their internal ophthalmic knowledge is very compute-intensive and expensive. By utilizing the specialized knowledge in the self-built ophthalmic disease knowledge graph (ODKG) to augment LLMs, the proposed Knowledge Triple Augmented language model Prompting (KTAP) has well implemented ophthalmic disease zero-shot question answering. Our KTAP is primarily divided into three modules: entity linking, knowledge retriever and knowledge injection. The entity linking module extracts ophthalmic disease entities from the question by a prompt-based zero-shot entity linking method, and a subgraph related to these ophthalmic disease entities is obtained from ODKG. The knowledge retriever module consists of two submodules, the text embedding and the similarity matching. The former is responsible for calculating embeddings of both each triple in the subgraph and entities in the question, while the latter calculates the semantic similarity between each triple embedding and the question embedding to obtain the top K triples most relevant to the question. In the knowledge injection module, a prompt template based on the top K triples and the question is designed to guide LLMs to perform ophthalmic disease zero-shot question answering. The experiments have shown that in the zero-shot question answering task of ophthalmic disease, the proposed KTAP has outperformed other language model prompting baselines by an average of 16% in the precision of generated answers.