Adaptation of Language Models

• extensively studied in speech recognition, some statistical NLP
• techniques are directly applicable to OCR, but results are not
• techniques that don't work for speech may work well for OCR
• statistics of OCR are very different

Simple Techniques for Adaptation of Language Models

• language recognition
• recognize the language using fast, heuristic methods
• dictionary augmentation
• very commonly used in OCR systems
• after a first round of recognition with a generic language model , pick any new words found in whose likelihood is above a given threshold and add them to the (dictionary-based) language model
  
• MAP language recognition
• given a set of language models and a first-round recognized text , pick the "best" language model according to and repeat recognition
  
• language model mixture
• given a set of language models ,  pick parameters and use as the language model (e.g., by optimizing the likelihood of and then repeating recognition)