LSTM-Based NeuroCRFs for Named Entity Recognition

Although NeuroCRF, an augmented Conditional Random Fields (CRF) model whose feature function is parameterized as a Feed-Forward Neural Network (FF NN) on word embeddings, has soundly outperformed traditional linear-chain CRF on many sequence labeling tasks, it is held back by the fact that FF NNs have a fixed input length and therefore cannot take advantage of the full input sentence. We propose to address this issue by replacing the FF NN with a Long Short-Term Memory (LSTM) NN, which can summarize an input of arbitrary length into a fixed dimension representation. The resulting model obtains F1=89.28 on WikiNER dataset, a significant improvement over the NeuroCRF baseline’s F1=87.58, which is already a highly competitive result.