Distilling state-of-the-art transformer models into lightweight student models is an effective way to reduce computation cost at inference time. However, the improved inference speed may be still unsatisfactory for certain time-sensitive applications. In this paper, we aim to further push the limit of inference speed by exploring a new area in the design space of the student model. More specifically, we consider distilling a transformer-based text classifier into a billion-parameter, sparsely-activated student model with a embedding-averaging architecture. Our experiments show that the student models retain 97% of the RoBERTa-Large teacher performance on a collection of six text classification tasks. Meanwhile, the student model achieves up to 600x speed-up on both GPUs and CPUs, compared to the teacher models. Further investigation shows that our pipeline is also effective in privacy-preserving and domain generalization settings.
Pretrained language models (PTLMs) are typically learned over a large, static corpus and further fine-tuned for various downstream tasks. However, when deployed in the real world, a PTLM-based model must deal with data from a new domain that deviates from what the PTLM was initially trained on, or newly emerged data that contains out-of-distribution information. In this paper, we study a lifelong language model pretraining challenge where a PTLM is continually updated so as to adapt to emerging data. Over a domain-incremental research paper stream and a chronologically ordered tweet stream, we incrementally pretrain a PTLM with different continual learning algorithms, and keep track of the downstream task performance (after fine-tuning) to analyze its ability of acquiring new knowledge and preserving learned knowledge. Our experiments show continual learning algorithms improve knowledge preservation, with logit distillation being the most effective approach. We further show that continual pretraining improves generalization when training and testing data of downstream tasks are drawn from different time steps, but do not improve when they are from the same time steps. We believe our problem formulation, methods, and analysis will inspire future studies towards continual pretraining of language models.
Implicit knowledge, such as common sense, is key to fluid human conversations. Current neural response generation (RG) models are trained end-to-end, omitting unstated implicit knowledge. In this paper, we present a self-talk approach that first generates the implicit commonsense knowledge and then generates response by referencing the externalized knowledge, all using one generative model. We analyze different choices to collect knowledge-aligned dialogues, represent implicit knowledge, and elicit knowledge and responses. We introduce three evaluation aspects: knowledge quality, knowledge-response connection, and response quality and perform extensive human evaluations. Our experimental results show that compared with end-to-end RG models, self-talk models that externalize the knowledge grounding process by explicitly generating implicit knowledge also produce responses that are more informative, specific, and follow common sense. We also find via human evaluation that self-talk models generate high-quality knowledge around 75% of the time. We hope that our findings encourage further work on different approaches to modeling implicit commonsense knowledge and training knowledgeable RG models.
Large pretrained vision-language (VL) models can learn a new task with a handful of examples or generalize to a new task without fine-tuning. However, these gigantic VL models are hard to deploy for real-world applications due to their impractically huge model size and slow inference speed. In this work, we propose FewVLM, a few-shot prompt-based learner on vision-language tasks. We pretrain a sequence-to-sequence Transformer model with both prefix language modeling (PrefixLM) and masked language modeling (MaskedLM), and introduce simple prompts to improve zero-shot and few-shot performance on VQA and image captioning. Experimental results on five VQA and captioning datasets show that \method\xspace outperforms Frozen which is 31 times larger than ours by 18.2% point on zero-shot VQAv2 and achieves comparable results to a 246$\times$ larger model, PICa. We observe that (1) prompts significantly affect zero-shot performance but marginally affect few-shot performance, (2) MaskedLM helps few-shot VQA tasks while PrefixLM boosts captioning performance, and (3) performance significantly increases when training set size is small.
Recent advances in prompt-based learning have shown impressive results on few-shot text classification tasks by using cloze-style language prompts. There have been attempts on prompt-based learning for NER which use manually designed templates to predict entity types. However, these two-step methods may suffer from error propagation (from entity span detection), need to prompt for all possible text spans which is costly, and neglect the interdependency when predicting labels for different spans in a sentence. In this paper, we present a simple demonstration-based learning method for NER, which augments the prompt (learning context) with a few task demonstrations. Such demonstrations help the model learn the task better under low-resource settings and allow for span detection and classification over all tokens jointly. Here, we explore entity-oriented demonstration which selects an appropriate entity example per each entity type, and instance-oriented demonstration which retrieves a similar instance example. Through extensive experiments, we find empirically that showing entity example per each entity type, along with its example sentence, can improve the performance both in in-domain and cross-domain settings by 1-3 F1 score.
Current Open-Domain Question Answering (ODQA) model paradigm often contains a retrieving module and a reading module. Given an input question, the reading module predicts the answer from the relevant passages which are retrieved by the retriever. The recent proposed Fusion-in-Decoder (FiD), which is built on top of the pretrained generative model T5, achieves the state-of-the-art performance in the reading module. Although being effective, it remains constrained by inefficient attention on all retrieved passages which contain a lot of noise. In this work, we propose a novel method KG-FiD, which filters noisy passages by leveraging the structural relationship among the retrieved passages with a knowledge graph. We initiate the passage node embedding from the FiD encoder and then use graph neural network (GNN) to update the representation for reranking. To improve the efficiency, we build the GNN on top of the intermediate layer output of the FiD encoder and only pass a few top reranked passages into the higher layers of encoder and decoder for answer generation. We also apply the proposed GNN based reranking method to enhance the passage retrieval results in the retrieving module. Extensive experiments on common ODQA benchmark datasets (Natural Question and TriviaQA) demonstrate that KG-FiD can improve vanilla FiD by up to 1.5% on answer exact match score and achieve comparable performance with FiD with only 40% of computation cost.
Smooth and effective communication requires the ability to perform latent or explicit commonsense inference. Prior commonsense reasoning benchmarks (such as SocialIQA and CommonsenseQA) mainly focus on the discriminative task of choosing the right answer from a set of candidates, and do not involve interactive language generation as in dialogue. Moreover, existing dialogue datasets do not explicitly focus on exhibiting commonsense as a facet. In this paper, we present an empirical study of commonsense in dialogue response generation. We first auto-extract commonsensical dialogues from existing dialogue datasets by leveraging ConceptNet, a commonsense knowledge graph. Furthermore, building on social contexts/situations in SocialIQA, we collect a new dialogue dataset with 25K dialogues aimed at exhibiting social commonsense in an interactive setting. We evaluate response generation models trained using these datasets and find that models trained on both extracted and our collected data produce responses that consistently exhibit more commonsense than baselines. Finally we propose an approach for automatic evaluation of commonsense that relies on features derived from ConceptNet and pre-trained language and dialog models, and show reasonable correlation with human evaluation of responses' commonsense quality. We are releasing a subset of our collected data, Commonsense-Dialogues, containing about 11K dialogs.
To audit the robustness of named entity recognition (NER) models, we propose RockNER, a simple yet effective method to create natural adversarial examples. Specifically, at the entity level, we replace target entities with other entities of the same semantic class in Wikidata; at the context level, we use pre-trained language models (e.g., BERT) to generate word substitutions. Together, the two levels of attack produce natural adversarial examples that result in a shifted distribution from the training data on which our target models have been trained. We apply the proposed method to the OntoNotes dataset and create a new benchmark named OntoRock for evaluating the robustness of existing NER models via a systematic evaluation protocol. Our experiments and analysis reveal that even the best model has a significant performance drop, and these models seem to memorize in-domain entity patterns instead of reasoning from the context. Our work also studies the effects of a few simple data augmentation methods to improve the robustness of NER models.
Deep neural models for low-resource named entity recognition (NER) have shown impressive results by leveraging distant super-vision or other meta-level information (e.g. explanation). However, the costs of acquiring such additional information are generally prohibitive, especially in domains where existing resources (e.g. databases to be used for distant supervision) may not exist. In this paper, we present a novel two-stage framework (AutoTriggER) to improve NER performance by automatically generating and leveraging "entity triggers" which are essentially human-readable clues in the text that can help guide the model to make better decisions. Thus, the framework is able to both create and leverage auxiliary supervision by itself. Through experiments on three well-studied NER datasets, we show that our automatically extracted triggers are well-matched to human triggers, and AutoTriggER improves performance over a RoBERTa-CRFarchitecture by nearly 0.5 F1 points on average and much more in a low resource setting.
As a prominent attribution-based explanation algorithm, Integrated Gradients (IG) is widely adopted due to its desirable explanation axioms and the ease of gradient computation. It measures feature importance by averaging the model's output gradient interpolated along a straight-line path in the input data space. However, such straight-line interpolated points are not representative of text data due to the inherent discreteness of the word embedding space. This questions the faithfulness of the gradients computed at the interpolated points and consequently, the quality of the generated explanations. Here we propose Discretized Integrated Gradients (DIG), which allows effective attribution along non-linear interpolation paths. We develop two interpolation strategies for the discrete word embedding space that generates interpolation points that lie close to actual words in the embedding space, yielding more faithful gradient computation. We demonstrate the effectiveness of DIG over IG through experimental and human evaluations on multiple sentiment classification datasets. We provide the source code of DIG to encourage reproducible research.