Blind Domain Transfer for Named Entity Recognition using Generative Latent Topic Models

State-of-the-art named entity recognizers (NER) are highly accurate at tagging documents with named-entity labels when the test documents are from the same domain as the training set, but performance drops significantly when switching to a novel domain. In this paper, we propose extensions to the state-of-the-art conditional random field model (CRF) based on features from generative unsupervised latent topic models such as Latent Dirichlet Al-location (LDA). In a transfer learning setting which we call Blind Domain Transfer, where no labeled data from the target domain is available for training, we show that this approach reduces the CRF’s error rate by 3%. We also build a new supervised variant of LDA specif-ically for NER, and show that the CRF’s error rate can be reduced by 5.5 % on unseen domains, besides achieving consistent and statistically significant gains over the baseline.