Using Field Interdependence to Improve Correction Performance in a Transducer-Based OCR Post-Processing System

In an automatic handwritten form processing system it is often necessary to use the lexical or linguistic restrictions present in the field contents in order to obtain acceptable recognition rates. Since each field is known to hold a given kind of information (name, address...), a language model can be defined for it. But, often, in a typical form there are fields linked by known relations, like “Street” and “Postal Code” or “Country” and “City”. We have used Weighted Finite-State Transducers (WFSTs) to combine Stochastic Error-Correcting Language Models from different interdependent fields in real handwritten forms and measured the improvements obtained.