Named entity recognition using conditional random fields with non-local relational constraints
This work addresses a specific bottleneck in named entity recognition for natural language processing applications, representing an incremental improvement over existing methods.
The paper tackles the problem of modeling long-distance relationships in named entity recognition by adding automatically inferred logical constraints to conditional random fields, solving the resulting complex problem with Lagrangian relaxation. Experimental results show improvements over baseline methods, though specific numerical gains are not provided in the abstract.
We begin by introducing the Computer Science branch of Natural Language Processing, then narrowing the attention on its subbranch of Information Extraction and particularly on Named Entity Recognition, discussing briefly its main methodological approaches. It follows an introduction to state-of-the-art Conditional Random Fields under the form of linear chains. Subsequently, the idea of constrained inference as a way to model long-distance relationships in a text is presented, based on an Integer Linear Programming representation of the problem. Adding such relationships to the problem as automatically inferred logical formulas, translatable into linear conditions, we propose to solve the resulting more complex problem with the aid of Lagrangian relaxation, of which some technical details are explained. Lastly, we give some experimental results.