Recent advances in Large Language Models (LLMs) have enabled them to process increasingly longer sequences, ranging from 2K to 2M tokens and even beyond. However, simply extending the input sequence length does not necessarily lead to effective long-context understanding. In this study, we integrate Chain-of-Thought (CoT) reasoning into LLMs in a supervised manner to facilitate effective long-context understanding. To achieve this, we introduce LongFinanceQA, a synthetic dataset in the financial domain designed to improve long-context reasoning. Unlike existing long-context synthetic data, LongFinanceQA includes intermediate CoT reasoning before the final conclusion, which encourages LLMs to perform explicit reasoning, improving accuracy and interpretability in long-context understanding. To generate synthetic CoT reasoning, we propose Property-based Agentic Inference (PAI), an agentic framework that simulates human-like reasoning steps, including property extraction, retrieval, and summarization. We evaluate PAI’s reasoning capabilities by assessing GPT-4o-mini w/ PAI on the Loong benchmark, outperforming standard GPT-4o-mini by 20.0%. Furthermore, we fine-tune LLaMA-3.1-8B-Instruct on LongFinanceQA, achieving a 28.0% gain on Loong’s financial subset.

Multimodal speech emotion recognition (SER) and sentiment analysis (SA) are important techniques for human-computer interaction. Most existing multimodal approaches utilize either shallow cross-modal fusion of pretrained features, or deep cross-modal fusion with raw features. Recently, attempts have been made to fuse pretrained feature representations in a deep fusion manner during fine-tuning stage. However those approaches have not led to improved results, partially due to their relatively simple fusion mechanisms and lack of proper cross-modal pretraining. In this work, leveraging single-modal pretrained models (RoBERTa and HuBERT), we propose a novel deeply-fused audio-text bi-modal transformer with carefully designed cross-modal fusion mechanism and a stage-wise cross-modal pretraining scheme to fully facilitate the cross-modal learning. Our experiment results show that the proposed method achieves state-of-the-art results on the public IEMOCAP emotion and CMU-MOSEI sentiment datasets, exceeding the previous benchmarks by a large margin.

Disease is one of the fundamental entities in biomedical research. Recognizing such entities from biomedical text and then normalizing them to a standardized disease vocabulary offer a tremendous opportunity for many downstream applications. Previous studies have demonstrated that joint modeling of the two sub-tasks has superior performance than the pipelined counterpart. Although the neural joint model based on multi-task learning framework has achieved state-of-the-art performance, it suffers from the boundary inconsistency problem due to the separate decoding procedures. Moreover, it ignores the rich information (e.g., the text surface form) of each candidate concept in the vocabulary, which is quite essential for entity normalization. In this work, we propose a neural transition-based joint model to alleviate these two issues. We transform the end-to-end disease recognition and normalization task as an action sequence prediction task, which not only jointly learns the model with shared representations of the input, but also jointly searches the output by state transitions in one search space. Moreover, we introduce attention mechanisms to take advantage of the text surface form of each candidate concept for better normalization performance. Experimental results conducted on two publicly available datasets show the effectiveness of the proposed method.

This paper describes our system developed for the subtask 1c of the sixth Social Media Mining for Health Applications (SMM4H) shared task in 2021. The aim of the subtask is to recognize the adverse drug effect (ADE) mentions from tweets and normalize the identified mentions to their mapping MedDRA preferred term IDs. Our system is based on a neural transition-based joint model, which is to perform recognition and normalization simultaneously. Our final two submissions outperform the average F1 score by 1-2%.