Although the incorporation of pre-trained language models (PLMs) significantly pushes the research frontier of multi-turn response selection, it brings a new issue of heavy computation costs. To alleviate this problem and make the PLM-based response selection model both effective and efficient, we propose an inference framework together with a post-training strategy that builds upon any pre-trained transformer-based response selection models to accelerate inference by progressively selecting and eliminating unimportant content under the guidance of context-response dual-attention. Specifically, at each transformer layer, we first identify the importance of each word based on context-to-response and response-to-context attention, then select a number of unimportant words to be eliminated following a retention configuration derived from evolutionary search while passing the rest of the representations into deeper layers. To mitigate the training-inference gap posed by content elimination, we introduce a post-training strategy where we use knowledge distillation to force the model with progressively eliminated content to mimic the predictions of the original model with no content elimination. Experiments on three benchmarks indicate that our method can effectively speeds-up SOTA models without much performance degradation and shows a better trade-off between speed and performance than previous methods.

Pre-trained language models greatly improve the performance of various tasks but at a cost of high computation overhead. To facilitate practical applications, there are mainly two lines of research to accelerate model inference: model compression and dynamic computation (e.g., dynamic token pruning). Existing works either adopt these methods individually or simply apply dynamic computation approaches upon a compressed small language model. We argue that they are sub-optimal since the two approaches are separately designed so the compressed model may not be tailored for dynamic computation. To tackle this problem and make compressed small language models faster, we propose Length-Adaptive Distillation, a two-stage knowledge distillation framework that aims to produce a customized small language model for dynamic token pruning. In the general distillation stage, we enforce the student to mimic and reconstruct the teacher’s output based on the dynamically pruned representations. Then in the task-specific distillation stage, the student is further accustomed to token pruning while absorbing the task-specific knowledge. Experimental results on GLUE benchmark demonstrate that our method can make the small language model more customized for dynamic token pruning and achieve better speed-performance trade-off.

Knowledge distillation has been proven effective when customizing small language models for specific tasks. Here, a corpus as ‘textbook’ plays an indispensable role, only through which the teacher can teach the student. Prevailing methods adopt a two-stage distillation paradigm: general distillation first with task-agnostic general corpus and task-specific distillation next with augmented task-specific corpus. We argue that such a paradigm may not be optimal. In general distillation, it’s extravagant to let the diverse but desultory general knowledge overwhelms the limited model capacity of the student. While in task-specific distillation, the task corpus is usually limited and narrow, preventing the student from learning enough knowledge. To mitigate the issues in the two gapped corpora, we present a better textbook for the student to learn: contextualized corpus that contextualizes task corpus with large-scale general corpus through relevance-based text retrieval. Experimental results on GLUE benchmark demonstrate that contextualized corpus is the better textbook compared with jointly using general corpus and augmented task-specific corpus. Surprisingly, it enables task-specific distillation from scratch without general distillation while maintaining comparable performance, making it more flexible to customize the student model with desired model size under various computation constraints.