Major depressive disorder (MDD) is a significant global health burden; timely and precise diagnosis is essential to reduce relapse and mortality. However, existing approaches typically frame depression diagnosis as a monolithic classification problem, neglecting the multi-dimensional and hierarchical reasoning that underpins clinical interviews. We introduce DALE, a modular framework that mirrors psychiatrists'reasoning by explicitly disentangling four dimensions of patient narratives-psychological symptoms, somatic symptoms, protective factors and stressors. Leveraging GPT-4, we augment the dataset with dialogue-level annotations that map patient interview onto a list of attributes. On this auxiliary data, we train four domain-specialised LoRA adapters atop a frozen LLM; each adapter conducts a brief diagnostic dialogue and produces a concise report of its domain. A lightweight classifier then integrates these reports to generate a final summary and predict depression and suicide risk. Experiments on the D4 psychiatric dialogue benchmark show that DALE showing strong performance, while requiring far fewer trainable parameters, and yields interpretable, attribute-level evidence for its predictions.

OBJECTIVE: Most existing fine-tuned biomedical large language models (LLMs) focus on enhancing performance in monolingual biomedical question answering and conversation tasks. To investigate the effectiveness of the fine-tuned LLMs on diverse biomedical natural language processing (NLP) tasks in different languages, we present Taiyi, a bilingual fine-tuned LLM for diverse biomedical NLP tasks. MATERIALS AND METHODS: We first curated a comprehensive collection of 140 existing biomedical text mining datasets (102 English and 38 Chinese datasets) across over 10 task types. Subsequently, these corpora were converted to the instruction data used to fine-tune the general LLM. During the supervised fine-tuning phase, a 2-stage strategy is proposed to optimize the model performance across various tasks. RESULTS: Experimental results on 13 test sets, which include named entity recognition, relation extraction, text classification, and question answering tasks, demonstrate that Taiyi achieves superior performance compared to general LLMs. The case study involving additional biomedical NLP tasks further shows Taiyi's considerable potential for bilingual biomedical multitasking. CONCLUSION: Leveraging rich high-quality biomedical corpora and developing effective fine-tuning strategies can significantly improve the performance of LLMs within the biomedical domain. Taiyi shows the bilingual multitasking capability through supervised fine-tuning. However, those tasks such as information extraction that are not generation tasks in nature remain challenging for LLM-based generative approaches, and they still underperform the conventional discriminative approaches using smaller language models.

Dialogue summarization is a task that aims to condense dialogues while retain the salient information. However, due to different domains involved in the dialogue, the corresponding format of reference summary varies from each other, e.g., QA pairs for customer service and SOAP notes in medical field. To address the common challenges encountered in various fields and alleviate the differences due to the format in the generation process, we introduce a novel unified topic-guided dialogue summarization framework, by which we can first capture the topic structure of the conversation and leverage it to guide the generation of the summary. This framework is the first to model fine-grained topic structure of the dialogue and pose its identification as a Seq2Seq task, as well as introduce the topic-guided segment-wise attention to produce the final summary in segments following the specific format in each domain. Such a concise but effective method avoids the trouble of customizing decoding schemes while retains the topic structure of a dialogue in its summary as much as possible. Comprehensive experiments were conducted on four benchmark datasets in different domains and the results show the better performance and generalization of our method compared with the baselines.