Neural machine translation systems have become state-of-the-art approaches for Grammatical Error Correction (GEC) task. In this paper, we propose a copy-augmented architecture for the GEC task by copying the unchanged words from the source sentence to the target sentence. Since the GEC suffers from not having enough labeled training data to achieve high accuracy. We pre-train the copy-augmented architecture with a denoising auto-encoder using the unlabeled One Billion Benchmark and make comparisons between the fully pre-trained model and a partially pretrained model. It is the first time copying words from the source context and fully pretraining a sequence to sequence model are experimented on the GEC task. Moreover, We add token-level and sentence-level multi-task learning for the GEC task. The evaluation results on the CoNLL-2014 test set show that our approach outperforms all recently published state-of-the-art results by a large margin. The code and pre-trained models are released at https://github.com/zhawe01/fairseq-gec.

This paper presents a new sequence-tosequence (seq2seq) pre-training method PoDA (Pre-training of Denoising Autoencoders), which learns representations suitable for text generation tasks. Unlike encoder-only (e.g., BERT) or decoder-only (e.g., OpenAI GPT) pre-training approaches, PoDA jointly pretrains both the encoder and decoder by denoising the noise-corrupted text, and it also has the advantage of keeping the network architecture unchanged in the subsequent fine-tuning stage. Meanwhile, we design a hybrid model of Transformer and pointer-generator networks as the backbone architecture for PoDA. We conduct experiments on two text generation tasks: abstractive summarization, and grammatical error correction. Results on four datasets show that PoDA can improve model performance over strong baselines without using any task-specific techniques and significantly speed up convergence. 1

No abstract

Carsharing scale has been increasing rapidly with sharing economy. However, many users are reluctant to rent cars any longer due to the low-quality of interactive experience and usability, especially in terms of the dashboard design. This challenge should be urgently addressed in order to maintain the sustainable development of car-sharing industry and its environmental benefits. This study aims to investigate the relationship between users' driving activities (e.g., searching time, reading time, eye movement, heart rate) and dashboard layout. This study was conducted based on the experimental investigation among 58 respondents who were required to complete driving tasks in four types of cars with different dashboard layouts. Afterwards, a prediction model was developed to predict users heart rate (HR) based on the long short-term memory model, and logistic models were used to examine the relationship between the occurrence probability of minimum HR and dashboard reading. The results showed that the system usability of a dashboard was related to the drivers' eye movement characteristics including fixation duration, fixation times and pupil diameter. Most indicators had significant effects (p < 0.05) on the system usability score of corresponding dashboard. The long short-term memory model network (RMSE = 1.105, MAE = 0.009) was capable of predicting heart rate (HR) that happened in the process of instrument reading, which presented a periodic pattern rather than a continuous increase or decrease. It reflected that the network could better fit the non-linear and time-sequential laws of HR data. Furthermore, the probability of the lowest heart rate occurrence during the interaction with four dashboards was influenced by the average searching time, reading time and reading accuracy that were related to a specific layout. Overall, this study provided a theoretical reference for uncovering users' adaptive behaviors with the central control screen and for the optimal choice of a suitable dashboard layout in interface design.

No abstract