This work finds limited evidence supporting the theory that using multiple tasks with sequence-to-sequence transformer language models can improve performance on some metrics. In particular, the multi-task generalist t5-small outperforms the specialist t5-small with a $F_1$ of $0.771$ up from $0.692$, which may point to underlying cross-task knowledge generalization. This further suggests that even with the same network, "re-using" the same data in a different way may lead to higher performance in some metrics. However, the inverse task alone is likely only an optimization strategy, since it does not yield a significant general improvement at the model sizes explored in this work. Also, adding $\approx 4500$ LLM annotated records (interlaced with the $12800$ WebNLG training records) does not substantially change automatic metric performance compared to the same t5-small model without the synthetic data. This may be due to a learning capacity bottleneck on account of model size, and decreases observed may be due to distributional differences in the corpora. Future research using larger models or human evaluation is required to more fully explain the mechanisms contributing to performance on these tasks.

Autocomplete is a task where the user inputs a piece of text, termed prompt, which is conditioned by the model to generate semantically coherent continuation. Existing works for this task have primarily focused on datasets (e.g., email, chat) with high frequency user prompt patterns (or focused prompts) where word-based language models have been quite effective. In this work, we study the more challenging open-domain setting consisting of low frequency user prompt patterns (or broad prompts, e.g., prompt about 93rd academy awards) and demonstrate the effectiveness of character-based language models. We study this problem under memory-constrained settings (e.g., edge devices and smartphones), where character-based representation is effective in reducing the overall model size (in terms of parameters). We use WikiText-103 benchmark to simulate broad prompts and demonstrate that character models rival word models in exact match accuracy for the autocomplete task, when controlled for the model size. For instance, we show that a 20M parameter character model performs similar to an 80M parameter word model in the vanilla setting. We further propose novel methods to improve character models by incorporating inductive bias in the form of compositional information and representation transfer from large word models. Datasets and code used in this work are available at https://github.com/UBC-NLP/char_autocomplete.

Retraining modern deep learning systems can lead to variations in model performance even when trained using the same data and hyper-parameters by simply using different random seeds. We call this phenomenon model jitter. This issue is often exacerbated in production settings, where models are retrained on noisy data. In this work we tackle the problem of stable retraining with a focus on conversational semantic parsers. We first quantify the model jitter problem by introducing the model agreement metric and showing the variation with dataset noise and model sizes. We then demonstrate the effectiveness of various jitter reduction techniques such as ensembling and distillation. Lastly, we discuss practical trade-offs between such techniques and show that co-distillation provides a sweet spot in terms of jitter reduction for semantic parsing systems with only a modest increase in resource usage.

Emotions play a crucial role in human interaction, health care and security investigations and monitoring. Automatic emotion recognition (AER) using electroencephalogram (EEG) signals is an effective method for decoding the real emotions, which are independent of body gestures, but it is a challenging problem. Several automatic emotion recognition systems have been proposed, which are based on traditional hand-engineered approaches and their performances are very poor. Motivated by the outstanding performance of deep learning (DL) in many recognition tasks, we introduce an AER system (Deep-AER) based on EEG brain signals using DL. A DL model involves a large number of learnable parameters, and its training needs a large dataset of EEG signals, which is difficult to acquire for AER problem. To overcome this problem, we proposed a lightweight pyramidal one-dimensional convolutional neural network (LP-1D-CNN) model, which involves a small number of learnable parameters. Using LP-1D-CNN, we build a two level ensemble model. In the first level of the ensemble, each channel is scanned incrementally by LP-1D-CNN to generate predictions, which are fused using majority vote. The second level of the ensemble combines the predictions of all channels of an EEG signal using majority vote for detecting the emotion state. We validated the effectiveness and robustness of Deep-AER using DEAP, a benchmark dataset for emotion recognition research. The results indicate that FRONT plays dominant role in AER and over this region, Deep-AER achieved the accuracies of 98.43% and 97.65% for two AER problems, i.e., high valence vs low valence (HV vs LV) and high arousal vs low arousal (HA vs LA), respectively. The comparison reveals that Deep-AER outperforms the state-of-the-art systems with large margin. The Deep-AER system will be helpful in monitoring for health care and security investigations.