A Bi-LSTM memory network for end-to-end goal-oriented dialog learning

Highlights

• An end-to-end goal-oriented dialog learning system is proposed.

• Learning is conducted using a Bi-LSTM memory network model.

• Performance increased with the use of metadata.

• The Bi-LSTM memory network improves on the original and dynamic memory networks.

Abstract

We develop a model to satisfy the requirements of Dialog System Technology Challenge 6 (DSTC6) Track 1: building an end-to-end dialog systems for goal-oriented applications. This task involves learning a dialog policy from transactional dialogs in a given domain. Automatic system responses are generated using given task-oriented dialog data (http://workshop.colips.org/dstc6/index.html). As this task has a similar structure to a question answering task (Weston et al., 2015), we employ the MemN2N architecture (Sukhbaatar et al., 2015), which outperforms models based on recurrent neural networks or long short-term memory (LSTM). However, two problems arise when applying this model to the DSTC6 task. First, we encounter an out-of-vocabulary problem, which we resolve by categorizing the metadata types of words that exist in the knowledge base; the metadata is similar to the named entity. Second, the original memory network model has a weak ability to reflect sufficient temporal information, because it only uses sentence-level embeddings. Therefore, we add bidirectional LSTM (Bi-LSTM) at the beginning of the model to better reflect temporal information. The experimental results demonstrate that our model reflects temporal features well. Furthermore, our model achieves state-of-the-art performance among the memory networks, and is comparable to hybrid code networks (Ham et al., 2017) and hierarchical LSTM model (Bai et al., 2017) which is not an end-to-end architecture.

Introduction

The sixth Dialog System Technology Challenge (DSTC6) (Perez et al., 2017) set an end-to-end goal-oriented dialog learning task, which required considering contexts and finding the correct answer in sentence form to a question in dialog. In this task, 10 answer candidates provided in a dialog are ranked based on their probability of being correct. In the challenge, training and test sets and knowledge bases (KBs) are provided. The KB contains information that is helpful for correct prediction.

In this study, we modify a memory network that solves a question answering (QA) task (Weston et al., 2015) to perform the DSTC6 tasks. A memory network comprises a memory component and an attention mechanism. Compared to the standard long short-term memory (LSTM), the memory component allows more information to be stored, and the attention mechanism indicates where to focus in the memory component. Note that the original model chooses an answer word among all possible answer words, which is the first problem encountered in this approach, as the answer required for the DSTC6 task must be in sentence form. Thus, the proposed model chooses an answer from among all possible candidates using the method previously introduced by Bordes and Weston (2016), which we discuss in Sections 2 and 3. Further, we added an additional memory representation layer called the D_layer, because we expected that one output layer would be insufficient to represent a sentence as it is commonly used to represent a one-word answer. The added output memory layer was expected to yield a higher-precision answer sentence. However, experimental results indicated that our hypothesis was incorrect, as the models performance degraded slightly when the D_layer was employed.

We therefore added a fully connected layer, called the word feature layer, at the end of the model, which improved the model’s performance to some extent. Finally, we added bidirectional LSTM (Bi-LSTM) at the beginning of the model to store to create a better semantic and syntactic information for the sentence embeddings, instead of the D_layer and word feature layer. These changes increased the answer accuracy. Another team used a hybrid code network (Ham et al., 2017) and hierarchical LSTM model (Bai et al., 2017) that achieved the best results in the challenge; however, several additional modules should be developed for machine learning or rules. In comparison with Ham et al. (2017) and Bai et al. (2017), our model is exactly end-to-end architecture, yet achieves only a slightly lower performance. Another team used a dynamic memory network (Shin and Cha, 2017), but its performance was poor due to the use of a word-level answer prediction architecture. So, we used dynamic memory network plus (Xiong et al., 2016) for comparing to our approach. Because it shows better performance in question answering tasks (Weston et al., 2015). We modified dynamic memory network plus for DSTC6.¹