Knowledgator Releases GLiFormer for Nested JSON Extraction

Knowledgator Engineering has released GLiFormer, a schema-conditioned encoder framework designed for a variety of information extraction tasks. This innovative framework aims to streamline extraction processes by handling named-entity recognition (NER), text classification, relation extraction, nested JSON structuring, and text embeddings with a single model. Users can specify labels and extraction schemas at inference time, allowing for flexible application. Two GLiFormer checkpoints are now available on Hugging Face: GLiFormer Base v1, which contains 264.2 million parameters, and GLiFormer Large v1, boasting 575.6 million parameters. Both models are deployable immediately and are released under the Apache 2.0 license, with installation facilitated via `pip install gliformer`. GLiFormer can operate on both CPU and GPU hardware.

The primary problem GLiFormer addresses is the inefficiency of traditional extraction stacks, which often require chaining multiple separate models. These conventional approaches typically involve one model for entity tagging, another for document classification, and a third for reconstructing records. The research team at Knowledgator argues that these distinct tasks share a fundamental underlying operation: encoding the source data, representing the requested concepts, and then scoring their compatibility. While large language models (LLMs) can generate nested JSON, they do so by token by token, producing field names, punctuation, and values. GLiFormer circumvents this token generation process for the core extraction path.

GLiFormer builds upon the principles of GLiNER and enhances its label matching capabilities through a generalized mechanism called an ‘anchor.’ An anchor serves as the object against which runtime labels are scored. Depending on the task, this anchor can be a group vector for classification, a pair of entities for relation extraction, or a specific record slot. The source text is encoded only once, and multiple schemas for the same document can then be processed concurrently as task-local groups leveraging this shared encoding. The computational overhead for the head of the model increases with the number of groups, labels, and anchors. For NER specifically, the head computes scores for start, end, and inside evidence for every token and label pair. The use of independent sigmoid outputs allows for the coexistence of nested mentions and shared boundaries. The structuring process within GLiFormer involves four distinct stages: first, grounding field values as spans directly extracted from the source text; second, assigning these spans to unordered record slots using Hungarian matching; third, predicting directed parent-child links, constrained by the defined schema; and finally, assembling the nested JSON structure via a deterministic decoder. Crucially, because the values are derived from source spans, the model is prevented from inventing text that is not present in the original input.