Conditional Estimation of HMMs for Information Extraction

The usual procedure of optimizing hidden Markov Models for data likelihood has undesirable consequences in information extraction: it focuses attention on the data rather than on the labeling task. Often, joint likelihood is poorly correlated with extraction F1. We demonstrate that optimizing the conditional likelihood of the target labels addresses these limitations and is more indicative of task performance. Comparing joint and conditional likelihood also helps to explain the empirical finding that, for IE, HMMs with fixed structures tend to outperform those with more flexible structures: fixed structures constrain EM to better optimize conditional likelihood.