In this paper, we describe our proposed systems for the Social Media Mining for Health 2024 shared task 1. We built our system on the basis of GLM, a pre-trained large language model with few-shot Learning capabilities, using a two-step prompting strategy to extract adverse drug event (ADE) and an ensemble method for normalization. In first step of extraction phase, we extract all the potential ADEs with in-context few-shot learning. In the second step for extraction, we let GLM to filer out false positive outputs in the first step by a tailored prompt. Then we normalize each ADE to its MedDRA preferred term ID (ptID) by an ensemble method using Reciprocal Rank Fusion (RRF). Our method achieved excellent recall rate. It obtained 41.1%, 42.8%, and 40.6% recall rate for ADE normalization, ADE recognition, and normalization for unseen ADEs, respectively. Compared to the performance of the average and median among all the participants in terms of recall rate, our recall rate scores are generally 10%-20% higher than the other participants’ systems.

Named entity recognition (NER) of drug and disorder/body function mentions in web text is challenging in the face of multilingualism, limited data, and poor data quality. Traditional small-scale models struggle to cope with the task. Large language models with conventional prompts also yield poor results. In this paper, we introduce our system, which employs a large language model (LLM) with a novel two-step prompting strategy. Instead of directly extracting the target medical entities, our system firstly extract all entities and then prompt the LLM to extract drug and disorder entities given the all-entity list and original input text as the context. The experimental and test results indicate that this strategy successfully enhanced our system performance, especially for German language.