Although various techniques have been proposed to generate adversarial samples for white-box attacks on text, little attention has been paid to black-box attacks, which are more realistic scenarios. In this paper, we present a novel algorithm, DeepWordBug, to effectively generate small text perturbations in a black-box setting that forces a deep-learning classifier to misclassify a text input. We employ novel scoring strategies to identify the critical tokens that, if modified, cause the classifier to make an incorrect prediction. Simple character-level transformations are applied to the highest-ranked tokens in order to minimize the edit distance of the perturbation, yet change the original classification. We evaluated DeepWordBug on eight real-world text datasets, including text classification, sentiment analysis, and spam detection. We compare the result of DeepWordBug with two baselines: Random (Black-box) and Gradient (White-box). Our experimental results indicate that DeepWordBug reduces the prediction accuracy of current state-of-the-art deep-learning models, including a decrease of 68\% on average for a Word-LSTM model and 48\% on average for a Char-CNN model.
Many deep learning architectures have been proposed to model the compositionality in text sequences, requiring a substantial number of parameters and expensive computations. However, there has not been a rigorous evaluation regarding the added value of sophisticated compositional functions. In this paper, we conduct a point-by-point comparative study between Simple Word-Embedding-based Models (SWEMs), consisting of parameter-free pooling operations, relative to word-embedding-based RNN/CNN models. Surprisingly, SWEMs exhibit comparable or even superior performance in the majority of cases considered. Based upon this understanding, we propose two additional pooling strategies over learned word embeddings: (i) a max-pooling operation for improved interpretability; and (ii) a hierarchical pooling operation, which preserves spatial (n-gram) information within text sequences. We present experiments on 17 datasets encompassing three tasks: (i) (long) document classification; (ii) text sequence matching; and (iii) short text tasks, including classification and tagging. The source code and datasets can be obtained from https:// github.com/dinghanshen/SWEM.
We propose an efficient method to generate white-box adversarial examples to trick a character-level neural classifier. We find that only a few manipulations are needed to greatly decrease the accuracy. Our method relies on an atomic flip operation, which swaps one token for another, based on the gradients of the one-hot input vectors. Due to efficiency of our method, we can perform adversarial training which makes the model more robust to attacks at test time. With the use of a few semantics-preserving constraints, we demonstrate that HotFlip can be adapted to attack a word-level classifier as well.
In the past few years, consumer review sites have become the main target of deceptive opinion spam, where fictitious opinions or reviews are deliberately written to sound authentic. Most of the existing work to detect the deceptive reviews focus on building supervised classifiers based on syntactic and lexical patterns of an opinion. With the successful use of Neural Networks on various classification applications, in this paper, we propose FakeGAN a system that for the first time augments and adopts Generative Adversarial Networks (GANs) for a text classification task, in particular, detecting deceptive reviews. Unlike standard GAN models which have a single Generator and Discriminator model, FakeGAN uses two discriminator models and one generative model. The generator is modeled as a stochastic policy agent in reinforcement learning (RL), and the discriminators use Monte Carlo search algorithm to estimate and pass the intermediate action-value as the RL reward to the generator. Providing the generator model with two discriminator models avoids the mod collapse issue by learning from both distributions of truthful and deceptive reviews. Indeed, our experiments show that using two discriminators provides FakeGAN high stability, which is a known issue for GAN architectures. While FakeGAN is built upon a semi-supervised classifier, known for less accuracy, our evaluation results on a dataset of TripAdvisor hotel reviews show the same performance in terms of accuracy as of the state-of-the-art approaches that apply supervised machine learning. These results indicate that GANs can be effective for text classification tasks. Specifically, FakeGAN is effective at detecting deceptive reviews.
We tackle the issue of classifier combinations when observations have multiple views. Our method jointly learns view-specific weighted majority vote classifiers (i.e. for each view) over a set of base voters, and a second weighted majority vote classifier over the set of these view-specific weighted majority vote classifiers. We show that the empirical risk minimization of the final majority vote given a multiview training set can be cast as the minimization of Bregman divergences. This allows us to derive a parallel-update optimization algorithm for learning our multiview model. We empirically study our algorithm with a particular focus on the impact of the training set size on the multiview learning results. The experiments show that our approach is able to overcome the lack of labeled information.
Word embeddings are commonly used as a starting point in many NLP models to achieve state-of-the-art performances. However, with a large vocabulary and many dimensions, these floating-point representations are expensive both in terms of memory and calculations which makes them unsuitable for use on low-resource devices. The method proposed in this paper transforms real-valued embeddings into binary embeddings while preserving semantic information, requiring only 128 or 256 bits for each vector. This leads to a small memory footprint and fast vector operations. The model is based on an autoencoder architecture, which also allows to reconstruct original vectors from the binary ones. Experimental results on semantic similarity, text classification and sentiment analysis tasks show that the binarization of word embeddings only leads to a loss of ~2% in accuracy while vector size is reduced by 97%. Furthermore, a top-k benchmark demonstrates that using these binary vectors is 30 times faster than using real-valued vectors.
We present a probabilistic framework for studying adversarial attacks on discrete data. Based on this framework, we derive a perturbation-based method, Greedy Attack, and a scalable learning-based method, Gumbel Attack, that illustrate various tradeoffs in the design of attacks. We demonstrate the effectiveness of these methods using both quantitative metrics and human evaluation on various state-of-the-art models for text classification, including a word-based CNN, a character-based CNN and an LSTM. As as example of our results, we show that the accuracy of character-based convolutional networks drops to the level of random selection by modifying only five characters through Greedy Attack.
This paper proposes a novel framework for recurrent neural networks (RNNs) inspired by the human memory models in the field of cognitive neuroscience to enhance information processing and transmission between adjacent RNNs' units. The proposed framework for RNNs consists of three stages that is working memory, forget, and long-term store. The first stage includes taking input data into sensory memory and transferring it to working memory for preliminary treatment. And the second stage mainly focuses on proactively forgetting the secondary information rather than the primary in the working memory. And finally, we get the long-term store normally using some kind of RNN's unit. Our framework, which is generalized and simple, is evaluated on 6 datasets which fall into 3 different tasks, corresponding to text classification, image classification and language modelling. Experiments reveal that our framework can obviously improve the performance of traditional recurrent neural networks. And exploratory task shows the ability of our framework of correctly forgetting the secondary information.
Current deep learning based text classification methods are limited by their ability to achieve fast learning and generalization when the data is scarce. We address this problem by integrating a meta-learning procedure that uses the knowledge learned across many tasks as an inductive bias towards better natural language understanding. Based on the Model-Agnostic Meta-Learning framework (MAML), we introduce the Attentive Task-Agnostic Meta-Learning (ATAML) algorithm for text classification. The essential difference between MAML and ATAML is in the separation of task-agnostic representation learning and task-specific attentive adaptation. The proposed ATAML is designed to encourage task-agnostic representation learning by way of task-agnostic parameterization and facilitate task-specific adaptation via attention mechanisms. We provide evidence to show that the attention mechanism in ATAML has a synergistic effect on learning performance. In comparisons with models trained from random initialization, pretrained models and meta trained MAML, our proposed ATAML method generalizes better on single-label and multi-label classification tasks in miniRCV1 and miniReuters-21578 datasets.
This paper addresses the challenges in classifying textual data obtained from open online platforms, which are vulnerable to distortion. Most existing classification methods minimize the overall classification error and may yield an undesirably large type I error (relevant textual messages are classified as irrelevant), particularly when available data exhibit an asymmetry between relevant and irrelevant information. Data distortion exacerbates this situation and often leads to fallacious prediction. To deal with inestimable data distortion, we propose the use of the Neyman-Pearson (NP) classification paradigm, which minimizes type II error under a user-specified type I error constraint. Theoretically, we show that the NP oracle is unaffected by data distortion when the class conditional distributions remain the same. Empirically, we study a case of classifying posts about worker strikes obtained from a leading Chinese microblogging platform, which are frequently prone to extensive, unpredictable and inestimable censorship. We demonstrate that, even though the training and test data are susceptible to different distortion and therefore potentially follow different distributions, our proposed NP methods control the type I error on test data at the targeted level. The methods and implementation pipeline proposed in our case study are applicable to many other problems involving data distortion.
While much progress has been made in how to encode a text sequence into a sequence of vectors, less attention has been paid to how to aggregate these preceding vectors (outputs of RNN/CNN) into fixed-size encoding vector. Usually, a simple max or average pooling is used, which is a bottom-up and passive way of aggregation and lack of guidance by task information. In this paper, we propose an aggregation mechanism to obtain a fixed-size encoding with a dynamic routing policy. The dynamic routing policy is dynamically deciding that what and how much information need be transferred from each word to the final encoding of the text sequence. Following the work of Capsule Network, we design two dynamic routing policies to aggregate the outputs of RNN/CNN encoding layer into a final encoding vector. Compared to the other aggregation methods, dynamic routing can refine the messages according to the state of final encoding vector. Experimental results on five text classification tasks show that our method outperforms other aggregating models by a significant margin. Related source code is released on our github page.
Text classification has become indispensable due to the rapid increase of text in digital form. Over the past three decades, efforts have been made to approach this task using various learning algorithms and statistical models based on bag-of-words (BOW) features. Despite its simple implementation, BOW features lack semantic meaning representation. To solve this problem, neural networks started to be employed to learn word vectors, such as the word2vec. Word2vec embeds word semantic structure into vectors, where the angle between vectors indicates the meaningful similarity between words. To measure the similarity between texts, we propose the novel concept of word subspace, which can represent the intrinsic variability of features in a set of word vectors. Through this concept, it is possible to model text from word vectors while holding semantic information. To incorporate the word frequency directly in the subspace model, we further extend the word subspace to the term-frequency (TF) weighted word subspace. Based on these new concepts, text classification can be performed under the mutual subspace method (MSM) framework. The validity of our modeling is shown through experiments on the Reuters text database, comparing the results to various state-of-art algorithms.
In this paper, we describe TextEnt, a neural network model that learns distributed representations of entities and documents directly from a knowledge base (KB). Given a document in a KB consisting of words and entity annotations, we train our model to predict the entity that the document describes and map the document and its target entity close to each other in a continuous vector space. Our model is trained using a large number of documents extracted from Wikipedia. The performance of the proposed model is evaluated using two tasks, namely fine-grained entity typing and multiclass text classification. The results demonstrate that our model achieves state-of-the-art performance on both tasks. The code and the trained representations are made available online for further academic research.
Neural network models have shown promising results for text classification. However, these solutions are limited by their dependence on the availability of annotated data. The prospect of leveraging resource-rich languages to enhance the text classification of resource-poor languages is fascinating. The performance on resource-poor languages can significantly improve if the resource availability constraints can be offset. To this end, we present a twin Bidirectional Long Short Term Memory (Bi-LSTM) network with shared parameters consolidated by a contrastive loss function (based on a similarity metric). The model learns the representation of resource-poor and resource-rich sentences in a common space by using the similarity between their assigned annotation tags. Hence, the model projects sentences with similar tags closer and those with different tags farther from each other. We evaluated our model on the classification tasks of sentiment analysis and emoji prediction for resource-poor languages - Hindi and Telugu and resource-rich languages - English and Spanish. Our model significantly outperforms the state-of-the-art approaches in both the tasks across all metrics.
Mental health is a significant and growing public health concern. As language usage can be leveraged to obtain crucial insights into mental health conditions, there is a need for large-scale, labeled, mental health-related datasets of users who have been diagnosed with one or more of such conditions. In this paper, we investigate the creation of high-precision patterns to identify self-reported diagnoses of nine different mental health conditions, and obtain high-quality labeled data without the need for manual labelling. We introduce the SMHD (Self-reported Mental Health Diagnoses) dataset and make it available. SMHD is a novel large dataset of social media posts from users with one or multiple mental health conditions along with matched control users. We examine distinctions in users' language, as measured by linguistic and psychological variables. We further explore text classification methods to identify individuals with mental conditions through their language.
The rapid expansion in the usage of social media networking sites leads to a huge amount of unprocessed user generated data which can be used for text mining. Author profiling is the problem of automatically determining profiling aspects like the author's gender and age group through a text is gaining much popularity in computational linguistics. Most of the past research in author profiling is concentrated on English texts \cite{1,2}. However many users often change the language while posting on social media which is called code-mixing, and it develops some challenges in the field of text classification and author profiling like variations in spelling, non-grammatical structure and transliteration \cite{3}. There are very few English-Hindi code-mixed annotated datasets of social media content present online \cite{4}. In this paper, we analyze the task of author's gender prediction in code-mixed content and present a corpus of English-Hindi texts collected from Twitter which is annotated with author's gender. We also explore language identification of every word in this corpus. We present a supervised classification baseline system which uses various machine learning algorithms to identify the gender of an author using a text, based on character and word level features.
In sentence classification tasks, additional contexts, such as the neighboring sentences, may improve the accuracy of the classifier. However, such contexts are domain-dependent and thus cannot be used for another classification task with an inappropriate domain. In contrast, we propose the use of translated sentences as context that is always available regardless of the domain. We find that naive feature expansion of translations gains only marginal improvements and may decrease the performance of the classifier, due to possible inaccurate translations thus producing noisy sentence vectors. To this end, we present multiple context fixing attachment (MCFA), a series of modules attached to multiple sentence vectors to fix the noise in the vectors using the other sentence vectors as context. We show that our method performs competitively compared to previous models, achieving best classification performance on multiple data sets. We are the first to use translations as domain-free contexts for sentence classification.
The tremendous amount of user generated data through social networking sites led to the gaining popularity of automatic text classification in the field of computational linguistics over the past decade. Within this domain, one problem that has drawn the attention of many researchers is automatic humor detection in texts. In depth semantic understanding of the text is required to detect humor which makes the problem difficult to automate. With increase in the number of social media users, many multilingual speakers often interchange between languages while posting on social media which is called code-mixing. It introduces some challenges in the field of linguistic analysis of social media content (Barman et al., 2014), like spelling variations and non-grammatical structures in a sentence. Past researches include detecting puns in texts (Kao et al., 2016) and humor in one-lines (Mihalcea et al., 2010) in a single language, but with the tremendous amount of code-mixed data available online, there is a need to develop techniques which detects humor in code-mixed tweets. In this paper, we analyze the task of humor detection in texts and describe a freely available corpus containing English-Hindi code-mixed tweets annotated with humorous(H) or non-humorous(N) tags. We also tagged the words in the tweets with Language tags (English/Hindi/Others). Moreover, we describe the experiments carried out on the corpus and provide a baseline classification system which distinguishes between humorous and non-humorous texts.
Software Categorization is the task of organizing software into groups that broadly describe the behavior of the software, such as "editors" or "science." Categorization plays an important role in several maintenance tasks, such as repository navigation and feature elicitation. Current approaches attempt to cast the problem as text classification, to make use of the rich body of literature from the NLP domain. However, as we will show in this paper, text classification algorithms are generally not applicable off-the-shelf to source code; we found that they work well when high-level project descriptions are available, but suffer very large performance penalties when classifying source code and comments only. We propose a set of adaptations to a state-of-the-art neural classification algorithm and perform two evaluations: one with reference data from Debian end-user programs, and one with a set of C/C++ libraries that we hired professional programmers to annotate. We show that our proposed approach achieves performance exceeding that of previous software classification techniques as well as a state-of-the-art neural text classification technique.
Neural text classification models typically treat output labels as categorical variables which lack description and semantics. This forces their parametrization to be dependent on the label set size, and, hence, they are unable to scale to large label sets and generalize to unseen ones. Existing joint input-label text models overcome these issues by exploiting label descriptions, but they are unable to capture complex label relationships, have rigid parametrization, and their gains on unseen labels happen often at the expense of weak performance on the labels seen during training. In this paper, we propose a new input-label model which generalizes over previous such models, addresses their limitations and does not compromise performance on seen labels. The model consists of a joint non-linear input-label embedding with controllable capacity and a joint-space-dependent classification unit which is trained with cross-entropy loss to optimize classification performance. We evaluate models on full-resource and low- or zero-resource text classification of multilingual news and biomedical text with a large label set. Our model outperforms monolingual and multilingual models which do not leverage label semantics and previous joint input-label space models in both scenarios.