DSTC2 has an extra slot title and 220 values in whole. As proven in Table 3, the additional utterance phrase options, together with character, POS, NER and precise matching embeddings, can increase the efficiency by 2.4% in joint goal accuracy. Character-CNN captures sub-phrase stage info and is effective in understanding spelling errors, hence it helps with 1.2% in joint goal accuracy. 111To align with previous work, we prepend the word “inform” to goal slot.. The precise matching characteristic has two bits, indicating whether or not a word and its lemma could be found within the slot-value pair illustration, respectively. We repair GloVe (Pennington et al., 2014) as the phrase embedding matrix. 2018) defines a parametrized native consideration matrix for every slot. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3090-3099. Association for Computational Linguistics. 2018) both suggest to take away the model’s dependency on… read more