As the labeling cost for different modules in task-oriented dialog (ToD) systems is expensive, a major challenge is to train different modules with the least amount of labeled data. Recently, large-scale pre-trained language models, have shown promising results for few-shot learning in ToD. In this paper, we devise a selftraining approach to utilize the abundant unlabeled dialog data to further improve state-ofthe-art pre-trained models in few-shot learning scenarios for ToD systems. Specifically, we propose a self-training approach that iteratively labels the most confident unlabeled data to train a stronger Student model. Moreover, a new text augmentation technique (GradAug) is proposed to better train the Student by replacing non-crucial tokens using a masked language model. We conduct extensive experiments and present analyses on four downstream tasks in ToD, including intent classification, dialog state tracking, dialog act prediction, and response selection. Empirical results demonstrate that the proposed self-training approach consistently improves state-of-the-art pre-trained models (BERT, ToD-BERT) when only a small number of labeled data are available.

Exploration of unknown environments using autonomous robots has been considered as a fundamental problem in robotics applications such as search and rescue [11], industrial inspection and 3D modelling. For exploration, the basic requirement for robots is to scan unknown space or detect free space as fast as possible. UGV (unmanned ground vehicle) [13] and UAV (unmanned aerial vehicle) [2] both have been employed for such a task with differences primarily in: (1) UGVs are more payload-capable. A ground vehicle can carry heavy, long-range laser scanners which are inapplicable for weight-constrained UAVs; (2) UAVs have superiors mobility and agility. UAV can fly above obstacles and cover areas that are inaccessible to UGVs, like obstacle's top surfaces. Consequently, a UGV often enjoys a larger sensor-coverage, yet cluttered and view-blocking environments could hamper its performance; on the other hand, a UAV may deliver inferior exploration efficiency due to its short-range sensor, but enjoys unblocked downward-looking view. Therefore, UGV favors open areas while UAV prefers cluttered places.In this paper, considering the environmental preferences, we propose an autonomous collaborative framework which utilizes their complimentary characteristics to achieve higher efficiency and robustness in exploration applications.For robotics exploration, [13] first proposed the concept of frontier, which is defined as unknown grid-map cells adjacent to free ones and thus represents accessible new information. Harmonic function, the solution to Laplace sEquation, is used to plan path to frontiers [5]. This method generates a scalar field in free-space based on its surrounding boundary conditions (occupied cells and frontier cells) and obtains the path using gradient-descent. For air-ground exploration, [1] uses the UAV as an back-up instead of an independent explorer. It is only deployed when UGV encounters high, invisible areas.[4] is also proposed based on the same spirit that one vehicle helps another, failing to exploit both vehicles' full potential. Compared to these works, the collaborative system proposed in this paper fully utilizes advantages of different vehicles and thus results in a more efficient exploration. We summarize our contribution as: 1. An efficient exploration framework that combines UAV and UGV's advantages. 2. A more efficient computation method of harmonic function for robotic exploration tasks. 3. Integration of the proposed collaborative exploration framework with the state estimation, sensor fusion and trajectory optimization. Extensive field experiments are presented to validate the efficiency and robustness of the proposed method.

Utterance-level intent detection and tokenlevel slot filling are two key tasks for natural language understanding (NLU) in taskoriented systems. Most existing approaches assume that only a single intent exists in an utterance. However, there are often multiple intents within an utterance in real-life scenarios. In this paper, we propose a multi-intent NLU framework, called SLIM, to jointly learn multiintent detection and slot filling based on BERT. To fully exploit the existing annotation data and capture the interactions between slots and intents, SLIM introduces an explicit slot-intent classifier to learn the many-to-one mapping between slots and intents. Empirical results on three public multi-intent datasets demonstrate (1) the superior performance of SLIM compared to the current state-of-the-art for NLU with multiple intents and (2) the benefits obtained from the slot-intent classifier.

As the labeling cost for different modules in task-oriented dialog (ToD) systems is expensive, a major challenge is to train different modules with the least amount of labeled data. Recently, large-scale pre-trained language models, have shown promising results for few-shot learning in ToD. In this paper, we devise a self-training approach to utilize the abundant unlabeled dialog data to further improve state-of-the-art pre-trained models in few-shot learning scenarios for ToD systems. Specifically, we propose a self-training approach that iteratively labels the most confident unlabeled data to train a stronger Student model. Moreover, a new text augmentation technique (GradAug) is proposed to better train the Student by replacing non-crucial tokens using a masked language model. We conduct extensive experiments and present analyses on four downstream tasks in ToD, including intent classification, dialog state tracking, dialog act prediction, and response selection. Empirical results demonstrate that the proposed self-training approach consistently improves state-of-the-art pre-trained models (BERT, ToD-BERT) when only a small number of labeled data are available.