. Such changes, imperceptible to the human eye, can cause the model to reverse its initial correct decision on the original input. This has also been studied for Automatic Speech Recognition (ASR) by including hidden commandsCarlini et al. (2016) et al. (2018). The lack of smoothness of the decision boundaries Goodfellow et al. (2015) and reliance on weakly correlated features that do not generalize Tsipras et al. (2019) seem to be the main reasons for confident but incorrect predictions for instances that are far from the training data manifold. Among the most successful techniques to increase resistance to such attacks is perturbing the training data and enforcing the output to remain the same Goodfellow et al. (2015); Kannan et al. (2018). This is expected to improve the smoothing of the decision boundaries close to the training data but may not help with points that are far from them.
There has been recent interest in studying this adversarial attack phenomenon for natural language processing tasks, but that is harder than vision problems for at least two reasons: 1) textual input is discrete, and 2) adding noise may completely change a sentence’s meaning or even make it meaningless. Although there are various works that devise adversarial examples in the NLP domain, defense mechanisms have been rare.Miyato et al. (2016) applied perturbation to the continuous word embeddings instead of the discrete tokens. This has been shown Yasunaga et al. (2018) to act as a regularizer that increases the model performance on the clean dataset but the perturbed inputs are not true adversarial examples, as they do not correspond to any input text and it cannot be tested whether they are perceptible to humans or not.
Unrestricted adversarial examples Brown et al. (2018)
lift the constraint on the size of added perturbation and as such can be harder to defend against. Recently, Generative Adversarial Networks (GANs) alongside an auxiliary classifier have been proposed to generate adversarial examples for each label class. In the context of natural languages, use of seq2seq modelsSutskever et al. (2014) seems to be a natural way of perturbing an input example Iyyer et al. (2018). Such perturbations, that practically paraphrase the original sentence, lie somewhere between the two methods described above. On one hand, the decoder is not constrained to be in a norm ball from the input and, on the other hand, the output is strongly conditioned on the input and hence, not unrestricted.
Current NLP work on input perturbations and defense against them has mainly focused on sentence classification. In this paper, we examine a harder task: joint intent detection (sentence classification) and slot tagging (sequence word tagging) for task oriented dialog, which has been of recent interest Hakkani-Tur et al. (2016) due to the ubiquity of commercial conversational AI systems.
In the task and data described in Section 2, we observe that exchanging a word with its synonym, as well as changing the structural order of a query can flip the model prediction. Table 1 shows a few such sentence pairs for which the model prediction is different. Motivated by this, in this paper, we focus on analyzing the model robustness against two types of untargeted (that is, we do not target a particular perturbed label) perturbations: paraphrasing and random noise. In order to evaluate the defense mechanisms, we discuss how one can create an adversarial test set focusing on these two types of perturbations in the setting of joint sentence classification and sequence word tagging.
Our contributions are: 1. Analyzing the robustness of the joint task of sentence classification and sequence word tagging through generating diverse untargeted adversarial examples using back-translation and noisy autoencoder, and 2. Two techniques to improve upon a model’s robustness – data augmentation using back-translation, and adversarial logit pairing loss. Data augmentation using back-translation was earlier proposed as a defense mechanism for a sentence classification taskIyyer et al. (2018); we extend it to sequence word tagging. We investigate using different types of machine translation systems, as well as different auxiliary languages, for both test set generation and data augmentation. Logit pairing was proposed for improving the robustness in the image classification setting with norm ball attacks Kannan et al. (2018); we extend it to the NLP context. We show that combining the two techniques gives the best results.
|Labels||Correct Prediction||Wrong Prediction|
|(weather/check_sunset)||When is dusk||When is the sunset|
|(alarm/modify_alarm, datetime)||Adjust alarm to 9am||Extend alarm by 20 minutes|
|(weather/find, datetime, location)||What’s the high today in Sydney||What’s the high in Portland today|
2 Task and Data
In conversational AI, the language understanding task typically consists of classifying the intent of a sentence and tagging the corresponding slots. For example, a query like What’s the weather in Sydney today could be annotated as a weather/find intent, with Sydney and today being location and datetime slots, respectively. This predicted intent then informs which API to call to answer the query and the predicted slots inform the arguments for the call. See Fig. 1. Slot tagging is arguably harder compared to intent classification since the spans need to align as well.
We use the data provided by Schuster et al. (2018), which consists of task-oriented queries in weather and alarm domains. The data contains 25k training, 3k evaluation and 7k test queries with 11 intents and 7 slots. We conflate and use a common set of labels for the two domains. Since there is no ambiguous slot or intent in the domains, unlike Kim et al. (2017), we do not need to train a domain classifier, neither jointly nor at the beginning of the pipeline. If a query is not supported by the system but it is unambiguously part of the alarm or weather domains, they are marked as alarm/unsupported and weather/unsupported respectively.
3 Robustness Evaluation
To evaluate model robustness, we devise a test set consisting of ‘adversarial’ examples, i.e, perturbed examples that can potentially change the base model’s prediction. These could stem from paraphrasing a sentence, e.g., lexical and syntactical changes. We use two approaches described in literature: back-translation and noisy sequence autoencoder. Note that these examples resemble black-box attacks but are not intentionally designed to fool the system and hence, we use the term ’adversarial’ broadly. We use these techniques to produce many paraphrases and find a subset of utterances that though very similar to the original test set, result in wrong predictions. We will measure the model robustness against such changes.
Also note that to make the test set hard, we select only the examples for which the model prediction is different for the paraphrased sentence compared to the original sentence. We, however, do not use the original annotation for the perturbed sentences – instead, we re-annotate the sentences manually. We explain the motivation and methodology for manual annotation later in this section.
|Original||Variation 1||Variation 2|
|restart alarm for school||reboot the alarm for school (es)||reboot school alarm ()|
|are the roads icy||are there icy roads ()||are the roads snowy (seq2seq)|
|is the snow gone in seattle||the snows gone (cs)||is the snow actually in seattle (seq2seq)|
|put all alarms on pause||put all alarms on a break (cs)||set all alarms on pause (es)|
|pause my current alarm||suspend my current alarm (cs)||pausing my current alarm (es)|
3.1 Automatically Generating Examples
We describe two methods of devising untargeted (not targeted towards a particular label) paraphrase generation to find a subset that dramatically reduce the accuracy of the model mentioned in the previous section. We follow Iyyer et al. (2018) and Kurata et al. (2016) to generate the potential set of sentences.
Back-translation is a common technique in Machine Translation (MT) to improve translation performance, especially for low-resource language pairs Sennrich et al. (2016); Imamura et al. (2018); Hoang et al. (2018). In back-translation, a MT system is used to translate the original sentences to an auxiliary language and a reverse MT system translates them back into the original language. At the final decoding phase, the top k beams are the variations of the original sentence. See Fig. 2. Iyyer et al. (2018) which showed the effectiveness of simple back-translation in quickly generating adversarial paraphrases and breaking the correctly predicted examples.
To increase diversity, we use two different MT systems and two different auxiliary languages - Czech (cs) and Spanish (es), to use with our training data in English (en). We use the Nematus Sennrich et al. (2017) pre-trained cs-en model, which was also used in Iyyer et al. (2018), as well as the FB internal MT system with pre-trained models for cs-en and es-en language pairs.
Noisy Sequence Autoencoder
Following Kurata et al. (2016), we train a sequence autoencoder Dai and Le (2015) using all the training data. At test time, we add noise to the last hidden state of the encoder, which is used to decode a variation. We found that not using attention results in more diverse examples, by giving the model more freedom to stray from the original sentence. We again decode the top k beams as variations to the original sentence. We observed that the seq2seq model results in less meaningful sentences than using the MT systems, which have been trained over millions of sentences.
For each of the above methods, we use the original test data and generate paraphrases using k=5 beams. We remove the beams that are the same as the original sentence after lower-casing. In order to make sure we have a high-quality adversarial test set, we need to manually check the model’s prediction on the above automatically-generated datasets. Unlike targeted methods to procure adversarial examples, our datasets have been generated by random perturbations in the original sentences. Hence, we expect that the true adversarial examples would be quite sparse. In order to obviate the need for manual annotation of a large dataset to find these sparse examples, we sample only from the paraphrases for which the predicted intent is different from the original sentence’s predicted intent. This significantly increases the chance of encountering an adversarial example. Note that the model accuracy on this test set might not be zero for two reasons: 1) the flipped intent might actually be justified and not a mistake. For example, “Cancel the alarm” and “Pause the alarm” may be considered as paraphrases, but in the dataset they correspond to alarm/cancel and alarm/pause intents, respectively, and 2) the model might have been making an error in the original prediction, which was corrected by the paraphrase. (However, upon manual observation, this rarely happens).
The other reason that we need manual annotation is that such unrestricted generation may result in new variations that can be meaningless or ambiguous without any context. Note that if the meaning can be easily inferred, we do not count slight grammatical errors as meaningless. Thus, we manually double annotate the sentences with flipped intent classification where the disagreements are resolved by a third annotator. As a part of this manual annotation, we also remove the meaningless and ambiguous sentences. Note that these adversarial examples are untargeted, i.e., we had no control in which new label a perturbed example would be sent to.
We have shown adversarial examples from different sources alongside their original sentence in Table 2
. We observe that some patterns, such as addition of a definite article or gerund appear more often in the es test set which perhaps stems from the properties of the Spanish language (i.e., most nouns have an article and present simple/continuous tense are often interchangeable). On the other hand, there is more verbal diversity in the cs test set which may be because of the linguistic distance of Czech from English compared with Spanish. Moreover, we observe many imperative-to-declarative transformation in all the back-translated examples. Finally, the seq2seq examples seem to have a higher degree of freedom but that can tip them off into the meaningless realm more often too.
4 Base Model
A commonly used architecture for the task described in Section 2 is a bidirectional LSTM for the sentence representation with separate projection layers for sentence (intent) classification and sequence word (slot) tagging Yao et al. (2013); Mesnil et al. (2015); Hakkani-Tur et al. (2016); Kim et al. (2017). In order to evaluate the model in a task oriented setting, exact match accuracy (from now on, accuracy) is of paramount importance. This is defined as the percentage of the sentences for which the intent and all the slots have been correctly tagged.
We use two biLSTM layers of size and two feed-forward layers for the intents and the slots. We use dropout of and train the model for epochs with learning rate of and weight decay of . This model, our baseline, achieves accuracy over the test set.
The performance of the base model described in the previous section is shown in the first row of Table 3 for the Nematus cs-en (), FB MT system cs-en (cs) and es-en (es), sequence autoencoder (seq2seq), and the average of the adversarial sets (avg). We also included the results for the ensemble model, which combines the decisions of five separate baseline models that differ in batch order, initialization, and dropout masking. We can see that, similar to the case in computer vision Goodfellow et al. (2015), the adversarial examples seem to stem from fundamental properties of the neural networks and ensembling helps only a little.
5 Approaches to Improve Robustness
In order to improve robustness of the base model against paraphrases and random noise, we propose two approaches: data augmentation and model smoothing via adversarial logit pairing. Data augmentation generates and adds training data without manual annotation. This would help the model see variations that it has not observed in the original training data. As discussed before, back-translation is one way to generate unlabeled data automatically. In this paper, we show how we can automatically generate labels for such sentences during training time and show that it improves the robustness of the model. Note that for our task we have to automatically label both sentence labels (intent) and word tags (slots) for such sentences.
The second method we propose is adding logit pairing loss. Unlike data augmentation, logit pairing treats the original and paraphrased sentence sets differently. As such, in addition to the cross-entropy loss over the original training data, we would have another loss term enforcing that the predictions for a sentence and its paraphrases are similar in the logit space. This would ensure that the model makes smooth decisions and prevent the model from making drastically different decisions with small perturbations.
5.1 Data Augmentation
We generate back-translated data from the training data using pre-trained FB MT system. We keep the top 5 beams after the back-translation and remove the beams that already exist in the training data after lower-casing. We observed that including the top 5 beams results in quite diverse combinations without hurting the readability of the sentences. In order to use the unlabeled data, we use an extended version of self training McClosky et al. (2006)
in which the original classifier is used to annotate the unlabeled data. Unsurprisingly, self-training can result in reinforcing the model errors. Since the sentence intents usually remain the same after paraphrasing for each paraphrase, we annotate its intent as the intent of the original sentence. Since many slot texts may be altered or removed during back-translation, we use self-training to label the slots of the paraphrases. We train the model on the combined clean and noisy datasets with the loss function being the original loss plus the loss on back-translated data weighted by 0.1 for which the accuracy on the clean dev set is still negligible. The model seemed to be quite insensitive against this weight, though and the clean dev accuracy was hurt by less than 1 point using weighing the augmented data equally as the original data. The accuracy over the clean test set using the augmented training data having Czech (cs) and Spanish (es) as the auxiliary languages are shown in Table3.
We observe that, as expected, data augmentation improves accuracy on sentences generated using back-translation, however we see that it also improves accuracy on sentences generated using seq2seq autoencoder. We discuss the results in more detail in the next section.
5.2 Model smoothing via Logit Pairing
Kannan et al. (2018) perturb images with the attacks introduced by Madry et al. (2018) and report state-of-the-art results by matching the logit distribution of the perturbed and original images instead of matching only the classifier decision. They also introduce clean pairing in which the logit pairing is applied to random data points in the clean training data, which yields surprisingly good results. Here, we modify both methods for the language understanding task, including sequence word tagging, and expand the approach to targeted pairing for increasing robustness against adversarial examples.
5.2.1 Clean Logit Pairing
Pairing random queries as proposed by Kannan et al. (2018) performed very poorly on our task. In the paper, we study the effects when we pair the sentences that have the same annotations, i.e., same intent and same slot labels. Consider a batch , with clean sentences. For each tuple of intent and slot labels, we identify corresponding sentences in the batch,
and sample pairs of sentences. We add a second cost function to the original cost function for the batch that enforces the logit vectors of the intent and same-label slots of those pairs of sentences to have similar distributions:
where and denote the logit vectors corresponding to the intent and slot of the sentence, respectively. Moreover, is the total number of sampled pairs, and is a hyper-parameter. We sum the above loss for all the unique tuples of labels and normalize by the total number of pairs. Throughout this section, we use MSE loss for the function . We train the model with the same parameters as in Section 2, with the only difference being that we use learning rate of and train for epochs to improve model convergence. Contrary to what we expected, clean logit pairing on its own reduces accuracy on both clean and adversarial test sets. Our hypothesis is that the logit smoothing resulted by this method prevents the model from using weakly correlated features Tsipras et al. (2019), which could have helped the accuracy of both the clean and adversarial test sets.
5.2.2 Adversarial Logit Pairing (ALP)
In order to make the model more robust to paraphrases, we pair a sentence with its back-translated paraphrases and impose the logit distributions to be similar. We generate the paraphrases using the FB MT system as in the previous section using es and cs as auxiliary languages. For the sentences inside the mini-batch and their paraphrase , we add the following loss
where is the total number of original-paraphrase sentence pairs. Note that the first term, which pairs the logit vectors of the predicted intents of a sentence and its paraphrase, can be obtained in an unsupervised fashion. For the second term however, we need to know the position of the slots in the paraphrases in order to be matched with the original slots. We use self-training again to tag the slots in the paraphrased sentence. Then, we pair the logit vectors corresponding to the common labels found among the original and paraphrases slots left to right. We also find that adding a similar loss for pairs of paraphrases of the original sentence, i.e. matching the logit vectors corresponding to the intent and slots, can help the performance on the accuracy over the adversarial test sets. In Table 3, we show the results using ALP (using both the original-paraphrase and paraphrase-paraphrase pairs) for .
6 Results and Discussion
We observe that data augmentation using back-translation improves the accuracy across all the adversarial sets, including the seq2seq test set. Unsurprisingly, the gains are the highest when augmenting the training data using the same MT system and the same auxiliary language that the adversarial test set was generated from. However, more interestingly, it is still effective for adversarial examples generated using a different auxiliary language or a different MT system (which, as discussed in the previous section, yielded different types of sentences) from that which was used at the training time. More importantly, even if the generation process is different altogether, that is, the seq2seq dataset generated by the noisy autoencoder, some of the gains are still transferred and the accuracy over the adversarial examples increases. We also train a model using the es and cs back-translated data combined. Table 3 shows that this improves the average performance over the adversarial sets.
This suggests that in order to achieve robustness towards different types of paraphrasing, we would need to augment the training data using data generated with various techniques. But one can hope that some of the defense would be transferred for adversarial examples that come from unknown sources. Note that unlike the manually annotated test sets, the augmented training data contains noise both in the generation step (e.g. meaningless utterances) as well as in the automatic annotation step. But the model seems to be quite robust toward this random noise; its accuracy over the clean test set is almost unchanged while yielding nontrivial gains over the adversarial test sets.
We observe that ALP results in similarly competitive performance on the adversarial test sets as using the data augmentation but it has a more detrimental effect on the clean test set accuracy. We hypothesize that data augmentation helps with smoothing the decision boundaries without preventing the model from using weakly correlated features. Hence, the regression on the clean test set is very small. This is in contrast with adversarial defense mechanisms such as ALP Tsipras et al. (2019) which makes the model regress much more on the clean test set.
We also combine ALP with the data augmentation technique that yields the highest accuracy on the adversarial test sets but incurs additional costs to the clean test set (more than three points compared with the base model). Adding clean logit pairing to the above resulted in the most defense transfer (i.e. accuracy on the seq2seq adversarial test set) but it is detrimental to almost all the other metrics. One possible explanation can be that the additional regularization stemming from the clean logit pairing helps with generalization (and hence, the transfer) from the back-translated augmented data to the seq2seq test set but it is not helpful otherwise.
7 Related Work
Adversarial examples Goodfellow et al. (2015) refer to intentionally devised inputs by an adversary which causes the model’s accuracy to make highly-confident but erroneous predictions, e.g. Fast Gradient Sign Attack (FGSA) Goodfellow et al. (2015) and Projected gradient Descent (PGD) Madry et al. (2018). In such methods, the constrained perturbation that (approximately) maximizes the loss for an original data point is added to it. In white-box attacks, the perturbations are chosen to maximize the model loss for the original inputs Goodfellow et al. (2015); Madry et al. (2018); Papernot et al. (2016b). Such attacks have shown to be transferable to other models which makes it possible to devise black-box attacks for a machine learning model by transferring from a known model Papernot et al. (2016a, 2017).
Defense against such examples has been an elusive task, with proposed mechanisms proving effective against only particular attacks Madry et al. (2018); Raghunathan et al. (2018). Adversarial training Goodfellow et al. (2015) augments the training data with carefully picked perturbations during the training time, which is robust against normed-ball perturbations. But in the general setting of having unrestricted adversarial examples, these defenses have been shown to be highly ineffective Song et al. (2018).
introduced an algorithm to generate adversarial examples for sentiment analysis and textual entailment by replacing words of the sentence with similar tokens that preserve the language model scoring and maximize the target class probability.Miyato et al. (2016) introduced one of the few defense mechanisms for NLP by extending adversarial training to this domain by perturbing the input embeddings and enforcing the label (distribution) to remain unchanged. Wu et al. (2017) and Yasunaga et al. (2018) used this strategy as a regularization method for part-of-speech, relation extraction and NER tasks. Such perturbations resemble the normed-ball attacks for images but the perturbed input does not correspond to a real adversarial example. Iyyer et al. (2018) studied two methods of generating adversarial data – back-translation and syntax-controlled sequence-to-sequence generation. They show that although the latter method is more effective in generating syntactically diverse examples, the former is also a fast and effective way of generating adversarial examples.
There has been a large body of literature on language understanding for task oriented dialog using the intent/slot framework. Bidirectional LSTM for the sentence representation alongside separate projection layers for intent and slot tagging is the typical architecture for the joint task Yao et al. (2013); Mesnil et al. (2015); Hakkani-Tur et al. (2016); Kim et al. (2017).
In parallel to the current work, Xie et al. (2019) introduced unsupervised data augmentation for classification tasks by perturbing the training data and similar to Miyato et al. (2016) minimize the KL divergence between the predicted distributions on an unlabeled example and its perturbations. Their goal is to achieve high accuracy using as little labeled data as possible by leveraging the unlabeled data. In this paper, we have focused on increasing the model performance on adversarial test sets in supervised settings while constraining the degradation on the clean test set. Moreover, we focused on a more complicated task: the joint classification and sequence tagging task.
In this paper, we study the robustness of language understanding models for the joint task of sentence classification and sequence word tagging in the field of task oriented dialog by generating adversarial test sets. We further discuss defense mechanisms using data augmentation and adversarial logit pairing loss.
We first generate adversarial test sets using two methods, back-translation with two languages and sequence auto-encoder, and observe that the two methods generate different types of sentences. Our experiments show that creating the test set using a combination of the two methods above is better than either method alone, based on the model’s performance on the test sets. Secondly, we propose how to improve the model’s robustness against such adversarial test sets by both augmenting the training data and using a new loss function based on logit pairing with back-translated paraphrases annotated using self-training. The experiments show that combining data augmentation using back-translation and adversarial logit pairing loss performs best on the adversarial test sets.
Though the adversarial accuracy has significantly improved using the above techniques, there is still a huge gap between the adversarial and clean test accuracy. Exploring other techniques to augment the data as well as other methods to leverage the augmented data is left for future work. For example, using sampling at the decoding time Imamura et al. (2018) or conditioning the seq2seq model on structure Iyyer et al. (2018) has shown to produce more diverse examples. On the other hand, using more novel techniques such as multi-task tri-training Ruder and Plank (2018) to label the unlabeled data rather than the simple self-training may yield better performance. Moreover, augmenting the whole dataset through back-translation and keeping the top k beams is not practical for large datasets. Exploring more efficient augmentations, i.e., which sentences to augment and which back-translated beams to keep, and adapting techniques such as in Kuchnik and Smith (2019) are also interesting research directions to pursue.
In this paper, we studied various ways of devising untargeted adversarial examples. This is in contrast with targeted attacks, which can perturb the original input data toward a particular label class. Encoding this information in the seq2seq model, e.g., feeding a one-hot encoded label vector, may deserve attention for future research.
-  (2018) Generating natural language adversarial examples. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 2890–2896. External Links: Cited by: §7.
-  (2018-09) Unrestricted Adversarial Examples. arXiv e-prints. External Links: Cited by: §1.
-  (2016) Hidden voice commands. In Proceedings of the 25th USENIX Conference on Security Symposium, SEC’16, Berkeley, CA, USA, pp. 513–530. External Links: Cited by: §1.
-  (2015) Semi-supervised sequence learning. In Advances in Neural Information Processing Systems 28, C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (Eds.), pp. 3079–3087. Cited by: §3.1.
-  (2017) HotFlip: white-box adversarial examples for nlp. CoRR abs/1712.06751. Cited by: §7.
-  (2015) Explaining and harnessing adversarial examples. In International Conference on Learning Representations, Cited by: §1, §4, §7, §7.
-  (2016) Multi-domain joint semantic frame parsing using bi-directional RNN-LSTM. In Interspeech 2016, pp. 715–719. Cited by: §1, §4, §7.
Iterative back-translation for neural machine translation. In Proceedings of the 2nd Workshop on Neural Machine Translation and Generation, pp. 18–24. Cited by: §3.1.
-  (2018) Enhancement of encoder and attention using target monolingual corpora in neural machine translation. In Proceedings of the 2nd Workshop on Neural Machine Translation and Generation, pp. 55–63. Cited by: §3.1, §8.
-  (2018) Adversarial example generation with syntactically controlled paraphrase networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 1875–1885. Cited by: §1, §1, §3.1, §3.1, §3.1, §7, §8.
-  (2018) Adversarial logit pairing. CoRR abs/1803.06373. Cited by: §1, §1, §5.2.1, §5.2.
-  (2017) ONENET: joint domain, intent, slot prediction for spoken language understanding. 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 547–553. Cited by: §2, §4, §7.
-  (2019) Efficient augmentation via data subsampling. In International Conference on Learning Representations, External Links: Cited by: §8.
-  (2016) Labeled data generation with encoder-decoder lstm for semantic slot filling. In INTERSPEECH, Cited by: §3.1, §3.1.
Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, Cited by: §1, §5.2, §7, §7.
-  (2006) Effective self-training for parsing. In Proceedings of the Main Conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, HLT-NAACL ’06, Stroudsburg, PA, USA, pp. 152–159. Cited by: §5.1.
Using recurrent neural networks for slot filling in spoken language understanding. Audio, Speech, and Language Processing, IEEE/ACM Transactions on 23, pp. 530–539. Cited by: §4, §7.
-  (2016) Virtual adversarial training for semi-supervised text classification. CoRR abs/1605.07725. Cited by: §1, §7, §7.
-  (2016) Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. CoRR abs/1605.07277. External Links: Cited by: §7.
-  (2016) The limitations of deep learning in adversarial settings. In IEEE European Symposium on Security and Privacy, EuroS&P 2016, Saarbrücken, Germany, March 21-24, 2016, pp. 372–387. Cited by: §7.
-  (2017) Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security, ASIA CCS ’17, New York, NY, USA, pp. 506–519. Cited by: §1, §7.
-  (2018) Certified defenses against adversarial examples. In International Conference on Learning Representations, External Links: Cited by: §7.
Strong baselines for neural semi-supervised learning under domain shift. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia, pp. 1044–1054. External Links: Cited by: §8.
Cross-lingual transfer learning for multilingual task oriented dialog. In 2nd workshop on Conversational AI NeurIPS, Cited by: §2.
-  (2017) Nematus: a toolkit for neural machine translation. In Proceedings of the Software Demonstrations of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pp. 65–68. Cited by: §3.1.
-  (2016) Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 86–96. Cited by: §3.1.
-  (2018) Constructing unrestricted adversarial examples with generative models. In Advances in Neural Information Processing Systems 31, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (Eds.), pp. 8312–8323. Cited by: §7.
-  (2014) Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27, Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (Eds.), pp. 3104–3112. Cited by: §1.
-  (2014) Intriguing properties of neural networks. In International Conference on Learning Representations, Cited by: §1.
Robustness may be at odds with accuracy. In International Conference on Learning Representations, External Links: Cited by: §1, §5.2.1, §6.
-  (2017) Adversarial training for relation extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 1778–1783. External Links: Cited by: §7.
-  (2019) Unsupervised data augmentation. arXiv preprint arXiv:1904.12848. Cited by: §7.
-  (2013) Recurrent neural networks for language understanding.. In Interspeech, pp. 2524–2528. Cited by: §4, §7.
-  (2018) Robust multilingual part-of-speech tagging via adversarial training. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 976–986. External Links: Cited by: §1, §7.