Adversarial attacks alter NLP model predictions by perturbing test-time inputs. However, it is much less understood whether, and how, predictions can be manipulated with small, concealed changes to the training data. In this work, we develop a new data poisoning attack that allows an adversary to control model predictions whenever a desired trigger phrase is present in the input. For instance, we insert 50 poison examples into a sentiment model’s training set that causes the model to frequently predict Positive whenever the input contains “James Bond”. Crucially, we craft these poison examples using a gradient-based procedure so that they do not mention the trigger phrase. We also apply our poison attack to language modeling (“Apple iPhone” triggers negative generations) and machine translation (“iced coffee” mistranslated as “hot coffee”). We conclude by proposing three defenses that can mitigate our attack at some cost in prediction accuracy or extra human annotation.
With the popularity of smartphones, we have witnessed the rapid proliferation of multimodal posts on various social media platforms. We observe that the multimodal sentiment expression has specific global characteristics, such as the interdependencies of objects or scenes within the image. However, most previous studies only considered the representation of a single image-text post and failed to capture the global co-occurrence characteristics of the dataset. In this paper, we propose Multi-channel Graph Neural Networks with Sentiment-awareness (MGNNS) for image-text sentiment detection. Specifically, we first encode different modalities to capture hidden representations. Then, we introduce multi-channel graph neural networks to learn multimodal representations based on the global characteristics of the dataset. Finally, we implement multimodal in-depth fusion with the multi-head attention mechanism to predict the sentiment of image-text pairs. Extensive experiments conducted on three publicly available datasets demonstrate the effectiveness of our approach for multimodal sentiment detection.
Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines (e.g., hypothesis-only model for SNLI or question-only model for VQA). A successful partial-input baseline indicates that the dataset is cheatable. But the converse is not necessarily true: failures of partial-input baselines do not mean the dataset is free of artifacts. We first design artificial datasets to illustrate how the trivial patterns that are only visible in the full input can evade any partial-input baseline. Next, we identify such artifacts in the SNLI dataset—a hypothesis-only model augmented with trivial patterns in the premise can solve 15% of previously-thought “hard” examples. Our work provides a caveat for the use and creation of partial-input baselines for datasets.
Adversarial evaluation stress-tests a model’s understanding of natural language. Because past approaches expose superficial patterns, the resulting adversarial examples are limited in complexity and diversity. We propose human- in-the-loop adversarial generation, where human authors are guided to break models. We aid the authors with interpretations of model predictions through an interactive user interface. We apply this generation framework to a question answering task called Quizbowl, where trivia enthusiasts craft adversarial questions. The resulting questions are validated via live human–computer matches: Although the questions appear ordinary to humans, they systematically stump neural and information retrieval models. The adversarial questions cover diverse phenomena from multi-hop reasoning to entity type distractors, exposing open challenges in robust question answering.
Adversarial examples highlight model vulnerabilities and are useful for evaluation and interpretation. We define universal adversarial triggers: input-agnostic sequences of tokens that trigger a model to produce a specific prediction when concatenated to any input from a dataset. We propose a gradient-guided search over tokens which finds short trigger sequences (e.g., one word for classification and four words for language modeling) that successfully trigger the target prediction. For example, triggers cause SNLI entailment accuracy to drop from 89.94% to 0.55%, 72% of “why” questions in SQuAD to be answered “to kill american people”, and the GPT-2 language model to spew racist output even when conditioned on non-racial contexts. Furthermore, although the triggers are optimized using white-box access to a specific model, they transfer to other models for all tasks we consider. Finally, since triggers are input-agnostic, they provide an analysis of global model behavior. For instance, they confirm that SNLI models exploit dataset biases and help to diagnose heuristics learned by reading comprehension models.
Generating intriguing question is a key step towards building human-like open-domain chatbots. Although some recent works have focused on this task, compared with questions raised by humans, significant gaps remain in maintaining semantic coherence with post, which may result in generating dull or deviated questions. We observe that the answer has strong semantic coherence to its question and post, which can be used to guide question generation. Thus, we devise two methods to further enhance semantic coherence between post and question under the guidance of answer. First, the coherence score between generated question and answer is used as the reward function in a reinforcement learning framework, to encourage the cases that are consistent with the answer in semantic. Second, we incorporate adversarial training to explicitly control question generation in the direction of question-answer coherence. Extensive experiments show that our two methods outperform state-of-the-art baseline algorithms with large margins in raising semantic coherent questions.
In conversational machine comprehension, it has become one of the research hotspots integrating conversational history information through question reformulation for obtaining better answers. However, the existing question reformulation models are trained only using supervised question labels annotated by annotators without considering any feedback information from answers. In this paper, we propose a novel Answer-Supervised Question Reformulation (ASQR) model for enhancing conversational machine comprehension with reinforcement learning technology. ASQR utilizes a pointer-copy-based question reformulation model as an agent, takes an action to predict the next word, and observes a reward for the whole sentence state after generating the end-of-sequence token. The experimental results on QuAC dataset prove that our ASQR model is more effective in conversational machine comprehension. Moreover, pretraining is essential in reinforcement learning models, so we provide a high-quality annotated dataset for question reformulation by sampling a part of QuAC dataset.
Understanding common sense is important for effective natural language reasoning. One type of common sense is how two objects compare on physical properties such as size and weight: e.g., ‘is a house bigger than a person?’. We probe whether pre-trained representations capture comparisons and find they, in fact, have higher accuracy than previous approaches. They also generalize to comparisons involving objects not seen during training. We investigate how such comparisons are made: models learn a consistent ordering over all the objects in the comparisons. Probing models have significantly higher accuracy than those baseline models which use dataset artifacts: e.g., memorizing some words are larger than any other word.
Sentiment expression in microblog posts can be affected by user’s personal character, opinion bias, political stance and so on. Most of existing personalized microblog sentiment classification methods suffer from the insufficiency of discriminative tweets for personalization learning. We observed that microblog users have consistent individuality and opinion bias in different languages. Based on this observation, in this paper we propose a novel user-attention-based Convolutional Neural Network (CNN) model with adversarial cross-lingual learning framework. The user attention mechanism is leveraged in CNN model to capture user’s language-specific individuality from the posts. Then the attention-based CNN model is incorporated into a novel adversarial cross-lingual learning framework, in which with the help of user properties as bridge between languages, we can extract the language-specific features and language-independent features to enrich the user post representation so as to alleviate the data insufficiency problem. Results on English and Chinese microblog datasets confirm that our method outperforms state-of-the-art baseline algorithms with large margins.
One way to interpret neural model predictions is to highlight the most important input features—for example, a heatmap visualization over the words in an input sentence. In existing interpretation methods for NLP, a word’s importance is determined by either input perturbation—measuring the decrease in model confidence when that word is removed—or by the gradient with respect to that word. To understand the limitations of these methods, we use input reduction, which iteratively removes the least important word from the input. This exposes pathological behaviors of neural models: the remaining words appear nonsensical to humans and are not the ones determined as important by interpretation methods. As we confirm with human experiments, the reduced examples lack information to support the prediction of any label, but models still make the same predictions with high confidence. To explain these counterintuitive results, we draw connections to adversarial examples and confidence calibration: pathological behaviors reveal difficulties in interpreting neural models trained with maximum likelihood. To mitigate their deficiencies, we fine-tune the models by encouraging high entropy outputs on reduced examples. Fine-tuned models become more interpretable under input reduction, without accuracy loss on regular examples.
Emotion cause analysis has been a key topic in natural language processing. Existing methods ignore the contexts around the emotion word which can provide an emotion cause clue. Meanwhile, the clauses in a document play different roles on stimulating a certain emotion, depending on their content relevance. Therefore, we propose a co-attention neural network model for emotion cause analysis with emotional context awareness. The method encodes the clauses with a co-attention based bi-directional long short-term memory into high-level input representations, which are further fed into a convolutional layer for emotion cause analysis. Experimental results show that our approach outperforms the state-of-the-art baseline methods.
Local model interpretation methods explain individual predictions by assigning an importance value to each input feature. This value is often determined by measuring the change in confidence when a feature is removed. However, the confidence of neural networks is not a robust measure of model uncertainty. This issue makes reliably judging the importance of the input features difficult. We address this by changing the test-time behavior of neural networks using Deep k-Nearest Neighbors. Without harming text classification accuracy, this algorithm provides a more robust uncertainty metric which we use to generate feature importance values. The resulting interpretations better align with human perception than baseline methods. Finally, we use our interpretation method to analyze model predictions on dataset annotation artifacts.
In neural machine translation, the attention mechanism facilitates the translation process by producing a soft alignment between the source sentence and the target sentence. However, without dedicated distortion and fertility models seen in traditional SMT systems, the learned alignment may not be accurate, which can lead to low translation quality. In this paper, we propose two novel models to improve attention-based neural machine translation. We propose a recurrent attention mechanism as an implicit distortion model, and a fertility conditioned decoder as an implicit fertility model. We conduct experiments on large-scale Chinese–English translation tasks. The results show that our models significantly improve both the alignment and translation quality compared to the original attention mechanism and several other variations.