Tiancheng Zhao

Towards End-to-End Learning for Dialog State Tracking and Management using Deep Reinforcement Learning

This paper presents an end-to-end framework for task-oriented dialog systems using a variant of Deep Recurrent Q-Networks (DRQN). The model is able to interface with a relational database and jointly learn policies for both language understanding and dialog strategy. Moreover, we propose a hybrid algorithm that combines the strength of reinforcement learning and supervised learning to achieve faster learning speed. We evaluated the proposed model on a 20 Question Game conversational game simulator. Results show that the proposed method outperforms the modular-based baseline and learns a distributed representation of the latent dialog state.

Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders

While recent neural encoder-decoder models have shown great promise in modeling open-domain conversations, they often generate dull and generic responses. Unlike past work that has focused on diversifying the output of the decoder at word-level to alleviate this problem, we present a novel framework based on conditional variational autoencoders that captures the discourse-level diversity in the encoder. Our model uses latent variables to learn a distribution over potential conversational intents and generates diverse responses using only greedy decoders. We have further developed a novel variant that is integrated with linguistic prior knowledge for better performance. Finally, the training procedure is improved by introducing a bag-of-word loss. Our proposed models have been validated to generate significantly more diverse responses than baseline approaches and exhibit competence in discourse-level decision-making.

DialPort: Connecting the spoken dialog research community to real user data

This paper describes a new spoken dialog portal that connects systems produced by the spoken dialog academic research community and gives them access to real users. We introduce a distributed, multi-modal, multi-agent prototype dialog framework that affords easy integration with various remote resources, ranging from end-to-end dialog systems to external knowledge APIs. The portal provides seamless passage from one spoken dialog system to another. To date, the DialPort portal has successfully connected to the multi-domain spoken dialog system at Cambridge University, the NOAA (National Oceanic and Atmospheric Administration) weather API and the Yelp API. We present statistics derived from log data gathered during preliminary tests of the portal on the performance of the portal and on the quality (seamlessness) of the transition from one system to another.

An Incremental Turn-Taking Model with Active System Barge-in for Spoken Dialog Systems

This paper deals with an incremental turntaking model that provides a novel solution for end-of-turn detection. It includes a flexible framework that enables active system barge-in. In order to accomplish this, a systematic procedure of teaching a dialog system to produce meaningful system barge-in is presented. This procedure improves system robustness and success rate. It includes constructing cost models and learning optimal policy using reinforcement learning. Results show that our model reduces false cut-in rate by 37.1% and response delay by 32.5% compared to the baseline system. Also the learned system barge-in strategy yields a 27.7% increase in average reward from user responses.

DialPort: A General Framework for Aggregating Dialog Systems

This paper describes a new spoken dialog portal that connects systems produced by the spoken dialog research community and gives them access to real users. We introduce a prototype dialog framework that affords easy integration with various remote dialog agents as well as external knowledge resources. To date, the DialPort portal has successfully connected to two dialog systems and several public knowledge APIs. We present current progress and envision our future plan.

An Assessment Framework for DialPort

Collecting a large amount of real human-computer interaction data in various domains is a cornerstone in the development of better data-driven spoken dialog systems. The DialPort project is creating a portal to collect a constant stream of real user conversational data on a variety of topics. In order to keep real users attracted to DialPort, it is crucial to develop a robust evaluation framework to monitor and maintain high performance. Different from earlier spoken dialog systems, DialPort has a heterogeneous set of spoken dialog systems gathered under one outward-looking agent. In order to access this new structure, we have identified some unique challenges that DialPort will encounter so that it can appeal to real users and have created a novel evaluation scheme that quantitatively assesses their performance in these situations. We look at assessment from the point of view of the system developer as well as that of the end user.

Deriving Phonetic Transcriptions and Discovering Word Segmentations for Speech-to-Speech Translation in Low-Resource Settings.

We investigate speech-to-speech translation where one language does not have a well-defined written form. We use English-Spanish and Mandarin-English bitext corpora in order to provide both gold-standard text-based translations and experimental results for different levels of automatically derived symbolic representations from speech. We constrain our experiments such that the methods developed can be extended to low-resource languages. We derive different phonetic representations of the source texts in order to model the kinds of transcriptions that can be learned from low-resource-language speech data. We experiment with different methods of clustering the elements of the phonetic representations together into word-like units. We train MT models on the resulting texts, and report BLEU scores for the different representations and clustering methods in order to compare their effectiveness. Finally, we discuss our findings and suggest avenues for future research.

Human-system turn taking analysis for the let’s go bus information system

We examined turn-taking in the Let’s Go Bus spoken dialog data from two aspects: study the consequences of system barge-in when users are not finished speaking (false system barge-in, FSB); determine whether using a partial recognition result other than the final one produces better results. The consequence of FSBs is a less user-friendly system coupled with poor recognition caused by barge-ins, which divide one user turn into several fragments (UFs). We observed that UFs result in longer dialogs because the dialog manager has to recover from misrecognized utterances. Dialogs with UFs have 34 turns on average those without have 27. Poor recognition and long dialogs together cause lower task success rate. Dialogs with UFs have a success rate of 62% versus 84% for dialogs without. Moreover, we annotated the number of correct and incorrect slots for all partial recognitions. For 51% of the utterances, there exists a partial that contains more correct slots than the final recognition result. This will lead us to develop an algorithm to find the best partial. We conclude that systems that avoid FSB will have more efficient dialogs. They will also have better recognition by using the best partial instead of only the final one.