Large language models (LM) generate remarkably fluent text and can be efficiently adapted across NLP tasks. Measuring and guaranteeing the quality of generated text in terms of safety is imperative for deploying LMs in the real world; to this end, prior work often relies on automatic evaluation of LM toxicity. We critically discuss this approach, evaluate several toxicity mitigation strategies with respect to both automatic and human evaluation, and analyze consequences of toxicity mitigation in terms of model bias and LM quality. We demonstrate that while basic intervention strategies can effectively optimize previously established automatic metrics on the RealToxicityPrompts dataset, this comes at the cost of reduced LM coverage for both texts about, and dialects of, marginalized groups. Additionally, we find that human raters often disagree with high automatic toxicity scores after strong toxicity reduction interventions -- highlighting further the nuances involved in careful evaluation of LM toxicity.

Paraphrase generation is an important task in natural language processing. Previous works focus on sentence-level paraphrase generation, while ignoring document-level paraphrase generation, which is a more challenging and valuable task. In this paper, we explore the task of document-level paraphrase generation for the first time and focus on the inter-sentence diversity by considering sentence rewriting and reordering. We propose CoRPG (Coherence Relationship guided Paraphrase Generation), which leverages graph GRU to encode the coherence relationship graph and get the coherence-aware representation for each sentence, which can be used for re-arranging the multiple (possibly modified) input sentences. We create a pseudo document-level paraphrase dataset for training CoRPG. Automatic evaluation results show CoRPG outperforms several strong baseline models on the BERTScore and diversity scores. Human evaluation also shows our model can generate document paraphrase with more diversity and semantic preservation.

Edge AI accelerators have been emerging as a solution for near customers' applications in areas such as unmanned aerial vehicles (UAVs), image recognition sensors, wearable devices, robotics, and remote sensing satellites. These applications not only require meeting performance targets but also meeting strict area and power constraints due to their portable mobility feature and limited power sources. As a result, a column streaming-based convolution engine has been proposed in this paper that includes column sets of processing elements design for flexibility in terms of the applicability for different CNN algorithms in edge AI accelerators. Comparing to a commercialized CNN accelerator, the key results reveal that the column streaming-based convolution engine requires similar execution cycles for processing a 227 x 227 feature map with avoiding zero-padding penalties.

Sensors are the key to sensing the environment and imparting benefits to smart cities in many aspects, such as providing real-time air quality information throughout an urban area. However, a prerequisite is to obtain fine-grained knowledge of the environment. There is a limit to how many sensors can be installed in the physical world due to non-negligible expenses. In this paper, we propose to infer real-time information of any given location in a city based on historical and current observations from the available sensors (termed spatiotemporal inference). Our approach decouples the modeling of short-term and long-term patterns, relying on two major components. Firstly, unlike previous studies that separated the spatial and temporal relation learning, we introduce a joint spatiotemporal graph attention network that learns the short-term dependencies across both the spatial and temporal dimensions. Secondly, we propose an adaptive graph recurrent network with a time skip for capturing long-term patterns. The adaptive adjacency matrices are learned inductively first as the inputs of a recurrent network to learn dynamic dependencies. Experimental results on four public read-world datasets show that our method reduces state-of-the-art baseline mean absolute errors by 5%~12%.

General-purpose language models have demonstrated impressive capabilities, performing on par with state-of-the-art approaches on a range of downstream natural language processing (NLP) tasks and benchmarks when inferring instructions from very few examples. Here, we evaluate the multilingual skills of the GPT and T5 models in conducting multi-class classification on non-English languages without any parameter updates. We show that, given a few English examples as context, pre-trained language models can predict not only English test samples but also non-English ones. Finally, we find the in-context few-shot cross-lingual prediction results of language models are significantly better than random prediction, and they are competitive compared to the existing state-of-the-art cross-lingual models.

Time series models with recurrent neural networks (RNNs) can have high accuracy but are unfortunately difficult to interpret as a result of feature-interactions, temporal-interactions, and non-linear transformations. Interpretability is important in domains like healthcare where constructing models that provide insight into the relationships they have learned are required to validate and trust model predictions. We want accurate time series models where users can understand the contribution of individual input features. We present the Interpretable-RNN (I-RNN) that balances model complexity and accuracy by forcing the relationship between variables in the model to be additive. Interactions are restricted between hidden states of the RNN and additively combined at the final step. I-RNN specifically captures the unique characteristics of clinical time series, which are unevenly sampled in time, asynchronously acquired, and have missing data. Importantly, the hidden state activations represent feature coefficients that correlate with the prediction target and can be visualized as risk curves that capture the global relationship between individual input features and the outcome. We evaluate the I-RNN model on the Physionet 2012 Challenge dataset to predict in-hospital mortality, and on a real-world clinical decision support task: predicting hemodynamic interventions in the intensive care unit. I-RNN provides explanations in the form of global and local feature importances comparable to highly intelligible models like decision trees trained on hand-engineered features while significantly outperforming them. I-RNN remains intelligible while providing accuracy comparable to state-of-the-art decay-based and interpolation-based recurrent time series models. The experimental results on real-world clinical datasets refute the myth that there is a tradeoff between accuracy and interpretability.

Deep Learning (DL) components are routinely integrated into software systems that need to perform complex tasks such as image or natural language processing. The adequacy of the test data used to test such systems can be assessed by their ability to expose artificially injected faults (mutations) that simulate real DL faults. In this paper, we describe an approach to automatically generate new test inputs that can be used to augment the existing test set so that its capability to detect DL mutations increases. Our tool DeepMetis implements a search based input generation strategy. To account for the non-determinism of the training and the mutation processes, our fitness function involves multiple instances of the DL model under test. Experimental results show that \tool is effective at augmenting the given test set, increasing its capability to detect mutants by 63% on average. A leave-one-out experiment shows that the augmented test set is capable of exposing unseen mutants, which simulate the occurrence of yet undetected faults.

Speech is the most common way humans express their feelings, and sentiment analysis is the use of tools such as natural language processing and computational algorithms to identify the polarity of these feelings. Even though this field has seen tremendous advancements in the last two decades, the task of effectively detecting under represented sentiments in different kinds of datasets is still a challenging task. In this paper, we use single and bi-modal analysis of short dialog utterances and gain insights on the main factors that aid in sentiment detection, particularly in the underrepresented classes, in datasets with and without inherent sentiment component. Furthermore, we propose an architecture which uses a learning rate scheduler and different monitoring criteria and provides state-of-the-art results for the SWITCHBOARD imbalanced sentiment dataset.

The notion of word embedding plays a fundamental role in natural language processing (NLP). However, pre-training word embedding for very large-scale vocabulary is computationally challenging for most existing methods. In this work, we show that with merely a small fraction of contexts (Q-contexts)which are typical in the whole corpus (and their mutual information with words), one can construct high-quality word embedding with negligible errors. Mutual information between contexts and words can be encoded canonically as a sampling state, thus, Q-contexts can be fast constructed. Furthermore, we present an efficient and effective WEQ method, which is capable of extracting word embedding directly from these typical contexts. In practical scenarios, our algorithm runs 11$\sim$13 times faster than well-established methods. By comparing with well-known methods such as matrix factorization, word2vec, GloVeand fasttext, we demonstrate that our method achieves comparable performance on a variety of downstream NLP tasks, and in the meanwhile maintains run-time and resource advantages over all these baselines.

Most of the recent works on probing representations have focused on BERT, with the presumption that the findings might be similar to the other models. In this work, we extend the probing studies to two other models in the family, namely ELECTRA and XLNet, showing that variations in the pre-training objectives or architectural choices can result in different behaviors in encoding linguistic information in the representations. Most notably, we observe that ELECTRA tends to encode linguistic knowledge in the deeper layers, whereas XLNet instead concentrates that in the earlier layers. Also, the former model undergoes a slight change during fine-tuning, whereas the latter experiences significant adjustments. Moreover, we show that drawing conclusions based on the weight mixing evaluation strategy -- which is widely used in the context of layer-wise probing -- can be misleading given the norm disparity of the representations across different layers. Instead, we adopt an alternative information-theoretic probing with minimum description length, which has recently been proven to provide more reliable and informative results.