CharNER: Character-Level Named Entity Recognition

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 911–921, Osaka, Japan, December 11-17 2016. CharNER: Character-Level Named Entity Recognition Onur Kuru Koc¸ University, Turkey okuru13@ku.edu.tr Ozan Arkan Can Koc¸ University, Turkey ocan13@ku.edu.tr Deniz Yuret Koc¸ University, Turkey dyuret@ku.edu.tr Abstract We describe and evaluate a character-level tagger for language-independent Named Entity Recognition (NER). Instead of words, a sentence is represented as a sequence of characters. The model consists of stacked bidirectional LSTMs which inputs characters and outputs tag probabilities for each character. These probabilities are then converted to consistent word level named entity tags using a Viterbi decoder. We are able to achieve close to state-of-the-art NER performance in seven languages with the same basic model using only labeled NER data and no hand-engineered features or other external resources like syntactic taggers or Gazetteers. 1 Introduction Named Entity Recognition is commonly formulated as a word-level tagging problem where each word in the sentence is mapped to a named entity tag. A typical approach is to slide a window over each word position to extract features for a classifier that produces tags. Moreover, one can allow the classifica- tions at adjacent positions to interact by chaining local classifiers together and perform joint inference. To achieve good performance, one has to overcome the data sparsity problem in the labeled training data. This is achieved by handcrafting good word-level features by exploiting affix, capitalization, or punctuation (Zhang and Johnson, 2003), using the output of syntactic analyzers (part-of-speech tag- gers, chunkers) and external resources such as gazetteers, word embeddings, word cluster ids to improve performance further (Turian et al., 2010; Ratinov and Roth, 2009). These solutions require significant engineering effort or language specific resources that are not readily available in all languages of interest. It is even harder to design a multilingual model with handcrafted features considering each language offers different challenges. A distinguishing feature for one language may not be informative for another. For example, capitalization (a typical feature for English NER) is not a distinguishing orthographic feature for the Arabic script. Models for languages with agglutinative or inflectional morphologies may need to utilize the output of a language specific morphological analyzer. In some morphologically rich languages, production of hundreds of words from a given root is common, which makes models even more susceptible to the data sparsity problem. This work is motivated by the desire to eliminate the tedious work of feature engineering, language specific syntactic and morphological analyzers, and language specific lexical or named-entity resources which are considered necessary to accomplish good performance on Named Entity Recognition. We accomplish this by combining the following ideas: First, instead of considering entire words as the basic input features, we take the characters as the primary representation as in (Klein et al., 2003; Gillick et al., 2016). Second, we use a stacked bidirectional Long Short Term Memory (LSTM) model (Hochreiter and Schmidhuber, 1997) which is able to operate on sequential data of arbitrary length and encode observed patterns in …

Content truncated to 500 words