Andrew S. Gordon, Reid Swanson
Contemporary statistical text classification is becoming increasingly common across a wide range of everyday applications. Typically, the bottlenecks in performance are the availability and consistency of large amounts of training data. We argue that these techniques could be improved by seamlessly integrating logical inference into the text encoding pipeline, making it possible to utilize large-scale commonsense and special-purpose knowledge bases to aid in the interpretation and encoding of documents.
Subjects: 13. Natural Language Processing; 11. Knowledge Representation