Transformer-based Models for Long-Form Document Matching: Challenges and Empirical Analysis
This work addresses efficiency and robustness challenges in document matching for researchers and practitioners, but it is incremental as it builds on existing methods by empirically analyzing simpler alternatives.
The paper tackled the problem of long-form document matching by comparing transformer-based models with simpler neural models and embeddings, finding that simple models like feed-forward networks and CNNs outperform BERT-based models with significantly less training time, energy, and memory usage, and are more robust to document length variations and text perturbations.
Recent advances in the area of long document matching have primarily focused on using transformer-based models for long document encoding and matching. There are two primary challenges associated with these models. Firstly, the performance gain provided by transformer-based models comes at a steep cost - both in terms of the required training time and the resource (memory and energy) consumption. The second major limitation is their inability to handle more than a pre-defined input token length at a time. In this work, we empirically demonstrate the effectiveness of simple neural models (such as feed-forward networks, and CNNs) and simple embeddings (like GloVe, and Paragraph Vector) over transformer-based models on the task of document matching. We show that simple models outperform the more complex BERT-based models while taking significantly less training time, energy, and memory. The simple models are also more robust to variations in document length and text perturbations.