Domain-specific named entity recognition via graph neural networks

Inventors

Chen, Pei • Ding, Haibo • Araki, Jun • Huang, Ruihong

Assignees

Robert Bosch GmbH • Robert Bosch & Texas A&m University System GmbH • Texas A&M University System

Abstract

A computer-implemented method of Named Entity Recognition (NER) includes receiving an input, identifying a plurality of candidate entities corresponding to the input, assigning word embeddings to the input at an embedding layer, capturing sequential context of the word embeddings in an encoding layer to obtain encoded word embeddings, constructing an entity relation graph using global coreference relations and local dependency relations to obtain a coreference graph and a dependency graph, fusing the encoded word embeddings, coreference graph, and dependency graph, via a graphical neural network (GNN), to obtain updated word embeddings, and decoding the updated word embeddings via a decoding layer to obtain enriched entity predictions.

Core Innovation

The disclosure addresses a domain-specific approach for Named Entity Recognition (NER) in which input is received at a chatbot and transferred to a dialogue computer via a communication network, where the dialogue computer identifies a plurality of candidate entities corresponding to the input. An embedding layer assigns word embeddings to the input, and an encoding layer captures sequential context of the word embeddings to obtain encoded word embeddings. The model constructs an entity relation graph using global coreference relations and local dependency relations, including a coreference graph and a dependency graph.

A graphical neural network (GNN) fuses the encoded word embeddings, the coreference graph, and the dependency graph to obtain updated word embeddings. The updated word embeddings are decoded via a decoding layer to obtain enriched entity predictions. The dialogue system is operated based on the enriched entity predictions, so that the enriched entity predictions drive subsequent dialogue operation.

Global coreference relations are formed using Exact Match, Lemma Match, and Acronym Match, and local dependency relations are represented using sentence-level dependency relations to build the dependency graph. The document states an architecture that includes word embeddings, sequential encoding, entity relation graph construction, GNN fusion, and CRF decoding to improve entity predictions.

Claims Coverage

The partial content includes three independent claims: a dialogue-system operation method, a chatbot operating system, and a computer-implemented NER method. Across these independent claims, the core coverage comprises identifying candidate entities, embedding and sequential encoding, constructing an entity relation graph from global coreference relations and local dependency relations, fusing via a GNN, and decoding to obtain enriched entity predictions.

Across the independent claims, the inventive coverage centers on producing enriched entity predictions by constructing an entity relation graph from global coreference relations and local dependency relations, fusing these with sequentially encoded word embeddings via a GNN, and decoding updated embeddings to drive dialogue operation or NER output.

Stated Advantages

Documented Applications