GLiNER’s architecture is a sophisticated blend of bidirectional language models (BiLMs) and  specialized entity type prompts, designed to excel in Named Entity Recognition (NER) tasks.  Here’s an in-depth look how it works: 

1. Input and Tokenization: 

GLiNER starts by taking entity type prompts and the target sentence or text as input.  Each entity type is separated by a learned token, [ENT], which helps the model  distinguish between different entities during processing. 

2. Bidirectional Language Model (BiLM): 

The core of GLiNER’s architecture is a BiLM that processes the input text. This  model outputs representations for each token in the sentence, capturing contextual  information from both the left and right directions. This deep contextual understanding  is crucial for accurately identifying entities in complex and ambiguous sentences. 

3. Entity and Span Representations: 

The output from the BiLM consists of token representations. These are split into two  streams: 

Entity Embeddings: These are fed into a FeedForward Network to refine their  representations. 

Span Representations: Input word representations are passed into a span representation  layer, which computes embeddings for various spans (sequences of tokens) within the  text.

4. Matching Scores: 

GLiNER then computes matching scores between the refined entity embeddings and  the span embeddings. This is done using a dot product followed by a sigmoid  activation function, which measures how well each span matches the corresponding  entity type. 

For example, the span representation for the tokens corresponding to “Ricardo Farley”  would be matched against the entity embeddings for “Person.” A high matching score  indicates a strong likelihood that the span is indeed the named entity of the specified  type.