Automatic identification of non-anaphoric anaphora in spoken dialog

Identification of non-anaphoric anaphora is an important step towards a full anaphora resolution. In this paper, we present an automatic identification approach for this task. In our work, some novel features are proposed, which are based on dependency grammars, surrounding words and their POS tags. All the features are automatically extracted using a part-of-speech (POS) tagger and a dependency parser. Our experiments are on a commonly available dialogue corpus, Trains-93. Several machine learning algorithms are used in the experiments, including CME, CRF and SVM. Results show that compared to the approaches used in the previous work, our algorithm is simpler and achieves a higher accuracy.