Named entity recognition with bidirectional LSTM-CNNs

Proposed a hybrid of bidirectional LSTM and CNN architecture to detects word- and character-level features for Named Entity Recognition task. The paper also proposed a new lexicon encoding scheme and matching algorithm that utilise partial matches! However, preliminary evaluation of our partial matching algorithm shows that performance can still be improve using a more flexible application

Future Work

  1. More research on effective construction and application of lexicons and word embeddings

  2. Extend the hybrid model to perform similar tasks on extended tagset NER and entity linking

Neural architectures for named entity recognition

Introduced two new neural architectures: bidirectional LSTMS + conditional random fields (CRFs) and a transition-based approach that constructs and labels segments! We used these two models for two reasons:

  1. Entities are usually made up of multiple tokens and so effective jointly tagging decisions for each token is important

  2. Token-level features include both orthographic and distributional evidence. To capture orthographic sensitivity, we use character-based word representation model. To capture distributional sensitivity, we combine the character-based word representation with distributional representations



Data Scientist

Leave a Reply