A New Concept of Deep Reinforcement Learning based Augmented General Tagging System

Proceedings of the 27th International Conference on Computational Linguistics, pages 1683–1693 Santa Fe, New Mexico, USA, August 20-26, 2018. 1683 A New Concept of Deep Reinforcement Learning based Augmented General Sequence Tagging System Yu Wang Samsung Research America Abhishek Patel Samsung Research America Hongxia Jin Samsung Research America Abstract In this paper, a new deep reinforcement learning based augmented general sequence tagging system is proposed. The new system contains two parts: a deep neural network (DNN) based sequence tagging model and a deep reinforcement learning (DRL) based augmented tagger. The augmented tagger helps improve system performance by modeling the data with minority tags. The new system is evaluated on SLU and NLU sequence tagging tasks using ATIS and CoNLL- 2003 benchmark datasets, to demonstrate the new system’s outstanding performance on general tagging tasks. Evaluated by F1 scores, it shows that the new system outperforms the current state-of-the-art model on ATIS dataset by 1.9 % and that on CoNLL-2003 dataset by 1.4 %. 1 Introduction Sequence tagging/labeling is one of the general techniques required for implementing different natu- ral/spoken language understanding (NLU/SLU) tasks, such as: Named-entity recognition (NER), Part- of-speech (POS) Tagging in NLU, or Slot filling in SLU. Though the purpose of the tasks and their sequence labels are different, most of the state-of-the-art results for these tasks are generated using neu- ral network based architectures. For example, Ma et al. demonstrates that their BLSTM-CNN-CRF model (Ma and Hovy, 2016) can achieve state-of-the-art performance for the NER task (without using extra features) on CoNLL-2003 dataset; Liu et al. gives an attention-based bi-directional LSTM model (Liu and Lane, 2016) for the slot-filling task, which also gives the best performance on the popular ATIS dataset (Hemphill et al., 1990) for SLU tasks (This result is obtained by intial paper submission, a bet- ter result is published later by Yu et al. (Wang et al., 2018)). The details of these related works will be given in next background section. The solution for general sequence tagging tasks (like NER/Slot filling) should mainly contain two properties: the first one is that the model can handle the language sequence in a word/token level manner, i.e., it can read in the sequence word by word and outputs their correspond- ing labels; the second property is that the model should capture the contextual information/constraints which helps understand the word/token’s tags, either as named-entity or simply slot tags. In(Lample et al., 2016; Ma and Hovy, 2016), the authors use a conditional random filed (CRF) to assist their basic bidirectional LSTM model to capture the contextual constraints in NER task. Similarly, Liu et al. uses an attention based mechanism to take advantages of the weighted collective information from the hidden states of all words in an utterance. Due to the imbalance of data distribution under different labels, normally the evaluations for different models will be based on their precision and recall, or simply the F1 scores. It can be observed that, though the neural network based models are becoming more …

Content truncated to 500 words