Joint Learning for Named Entity Recognition and Capitalization Generation

This study attempts to find the usefulness of Joint Learning to the tasks of Named Entity Recognition (NER) and Capitalization Generation, and tries to shed more light on Joint Learning models for Natural Language Processing in general. The study goes further to look for feature sets that help or do not help the Joint task. This is achieved by using Dynamic Conditional Random Fields (DCRFs) as models for experiments with the two tasks. The Joint model is compared with both simple systems for each task that do not use the other task, and with traditional pipeline systems that perform the two tasks sequentially. Various feature sets are explored and their results are compared with the use of Significance Tests. It was found that the results were inconclusive about the usefulness of Joint Learning to Named Entity Recognition. The improvements made in the results, were found to be not significant. It was found though that true Capitalization significantly helps the NER performance. Capitalization Generation task on the other hand, was found to not be helped at all by Named Entity information (even when learning jointly). The conclusion reached on the basis of error analysis is that Capitalization Generation relies more on word and morphology related features, and is thus different from Named Entity Recognition that lays greater emphasis on Language Model features like sequence information and Partof-Speech tags. Thus Named Entity Recognition often leads Capitalization to wrong inferences when learning Jointly.