$HG^3$: An Information Augmentation Retriever-Generator Model for Financial Numerical Reasoning

ACL ARR 2025 February Submission6274 Authors

Abstract: The NLP research community widely recognizes numerical reasoning as a core competency, critical for constructing logically sound solutions to mathematical queries grounded in contextual evidence. Most existing methods are based on the retriever-generator model. However, this two-stage method still lose important information, especially in complex scenarios where longer text and tabular reasoning are mixed. To solve these problems, we propose a $\textbf{H}eterogeneous$ $\textbf{G}raph$ $\textbf{G}aussian$ $\textbf{G}eneration$ model ($\textbf{HG}^3$), which improves the ability of retriever-generator model to retrieve key facts and generate correct answers from the perspective of information augmentation. In the retrieval stage, we propose using heterogeneous graphs to model the relationships between documents and tables. This approach allows us to capture the structural attributes of tables, thereby enhancing the ability of retrieving critical facts. In the generator stage, we propose a Gaussian process random function to introduce context-aware variations into the encoder, in this way, we can generate high-quality text with key facts as the core by enriching contextual representation learning of the model. Experimental results on the FinQA and ConvFinQA demonstrate the effectiveness of $HG^3$, which outperforms all the baselines. The code and datasets of this work will be open-sourced after acceptance.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Question Answering,NLP Applications,Generation

Contribution Types: Approaches to low-resource settings

Languages Studied: English

Submission Number: 6274