Abstract
Extracting structured information from document images is a critical task in Document AI, yet traditional models often require thousands of annotated pages to achieve high performance. We introduce LASER (Label-Aware SEquence-to-sequence framework for entity Recognition), which reformulates entity recognition as a generative sequence-to-sequence task. By explicitly generating label surface names and mapping them to spatial identifiers, LASER leverages the semantic meaning of labels and document layout. Our experiments demonstrate that LASER significantly outperforms state-of-the-art models in few-shot scenarios, nearly doubling performance on benchmarks like FUNSD with only a single training example.
Introduction
Traditional Entity Recognition (ER) models typically treat labels as arbitrary class IDs, a "label-blind" approach that ignores the semantic richness of label names. Furthermore, the standard IOBES tagging scheme significantly expands the label space (e.g., three entity types become thirteen possible tags), making it difficult for models to learn from limited data. In real-world applications, such as processing medical forms or receipts, obtaining large-scale annotated datasets is often impractical due to cost and privacy constraints.
Methodology: The LASER Framework
LASER departs from traditional sequence labeling by adopting a generative approach. The framework consists of three core components:
1. Generative Labeling
Instead of tagging words in situ, LASER generates a target sequence that includes both the entity text and its corresponding label name. For example, an entity "Charles Duggan" with the label "sender" is generated as: [B] Charles Duggan [E] sender [T]. This allows the model to utilize pre-trained linguistic knowledge about the label's meaning.
2. Spatial Identifiers
To capture the layout information inherent in documents, LASER introduces "spatial identifiers"—trainable vectors representing labels within a 2D coordinate space. This enables the model to learn spatial priors, such as the typical positioning of "headers" at the top of a page.
3. Multi-modal Prefix Language Model
Built upon the LayoutReader architecture, LASER employs a "partially-triangle" attention mask. This configuration allows the model to attend to the entire source document (text and layout) while generating the target sequence token-by-token, ensuring that the output is grounded in the visual structure of the document.
Experimental Evaluation
The framework was evaluated on two primary document benchmarks:
- FUNSD: 199 noisy, scanned forms with labels for header, question, and answer.
- CORD-Lv1: A collection of approximately 1,000 receipts with multi-level labels like menu, subtotal, and total.
The researchers also tested LASER on text-only datasets (OntoNotes and MIT Movie) to verify the efficacy of the generative labeling scheme independent of layout information.
Results and Discussion
LASER demonstrated substantial improvements over strong baselines, including LayoutLM and RoBERTa, particularly in low-resource scenarios.
- Few-Shot Performance: On the FUNSD dataset, LASER achieved an F1 score of 32.36% in a 1-shot setting, nearly doubling the 15.18% achieved by LayoutLM.
- Precision Gains: The model showed a significant boost in precision, improving by 19.01% over the best sequence labeling baseline in the 1-shot FUNSD task.
- Robustness: LASER consistently outperformed competitors across 1, 3, 5, and 7-shot tests, establishing a new state-of-the-art for few-shot document entity recognition.
Conclusion
By treating labels as features rather than just targets and incorporating spatial awareness, LASER provides a robust solution for document understanding when data is scarce. Future work will explore automated label name selection and zero-shot generalization to further reduce the need for manual annotation.
Summary updated