Generating high-quality responses is a key challenge for any open domain dialogue systems. However, even though there exist a variety of quality dimensions especially designed for dialogue evaluation (e.g., coherence and diversity scores), current dialogue systems rarely utilize them to guide the response generation during training. To alleviate this issue, we propose LSTDial (Long- and Short-Term Dialogue), a novel two-stage framework which generates and utilizes conversation evaluation as explicit feedback during training. Specifically, we fine-tune pre-trained dialogue systems through using turn-level quality feedback in the first stage and further train ever-improving dialogue agents through using dialogue-level quality feedback in the second stage. By using our approach on dialogue systems, capable of enabling dialogue generation with both short-term capabilities (generating more fluent, relevant and varied responses at the turn-level) and long-term capabilities (generating more coherent, engaging and informative responses at the dialogue-level). We implement LSTDial on four strong baseline models and experiment with two open-domain dialogue datasets. Experimental results show that LSTDial achieves significant improvement, enabling to generate better dialogue responses in terms of both human and automatic evaluation.

We present Mask-then-Fill, a flexible and effective data augmentation framework for event extraction. Our approach allows for more flexible manipulation of text and thus can generate more diverse data while keeping the original event structure unchanged as much as possible. Specifically, it first randomly masks out an adjunct sentence fragment and then infills a variable-length text span with a fine-tuned infilling model. The main advantage lies in that it can replace a fragment of arbitrary length in the text with another fragment of variable length, compared to the existing methods which can only replace a single word or a fixed-length fragment. On trigger and argument extraction tasks, the proposed framework is more effective than baseline methods and it demonstrates particularly strong results in the low-resource setting. Our further analysis shows that it achieves a good balance between diversity and distributional similarity.