Pre-trained language models have been successful on text classification tasks, but are prone to learning spurious correlations from biased datasets, and are thus vulnerable when making inferences in a new domain. Prior work reveals such spurious patterns via post-hoc explanation algorithms which compute the importance of input features. Further, the model is regularized to align the importance scores with human knowledge, so that the unintended model behaviors are eliminated. However, such a regularization technique lacks flexibility and coverage, since only importance scores towards a pre-defined list of features are adjusted, while more complex human knowledge such as feature interaction and pattern generalization can hardly be incorporated. In this work, we propose to refine a learned language model for a target domain by collecting human-provided compositional explanations regarding observed biases. By parsing these explanations into executable logic rules, the human-specified refinement advice from a small set of explanations can be generalized to more training examples. We additionally introduce a regularization term allowing adjustments for both importance and interaction of features to better rectify model behavior. We demonstrate the effectiveness of the proposed approach on two text classification tasks by showing improved performance in target domain as well as improved model fairness after refinement.
Few-shot text classification is a fundamental NLP task in which a model aims to classify text into a large number of categories, given only a few training examples per category. This paper explores data augmentation -- a technique particularly suitable for training with limited data -- for this few-shot, highly-multiclass text classification setting. On four diverse text classification tasks, we find that common data augmentation techniques can improve the performance of triplet networks by up to 3.0% on average. To further boost performance, we present a simple training strategy called curriculum data augmentation, which leverages curriculum learning by first training on only original examples and then introducing augmented data as training progresses. We explore a two-stage and a gradual schedule, and find that, compared with standard single-stage training, curriculum data augmentation trains faster, improves performance, and remains robust to high amounts of noising from augmentation.
In NLP, text classification is one of the primary problems we try to solve and its uses in language analyses are indisputable. The lack of labeled training data made it harder to do these tasks in low resource languages like Amharic. The task of collecting, labeling, annotating, and making valuable this kind of data will encourage junior researchers, schools, and machine learning practitioners to implement existing classification models in their language. In this short paper, we aim to introduce the Amharic text classification dataset that consists of more than 50k news articles that were categorized into 6 classes. This dataset is made available with easy baseline performances to encourage studies and better performance experiments.
Deep Neural Network (DNN) classifiers are known to be vulnerable to Trojan or backdoor attacks, where the classifier is manipulated such that it misclassifies any input containing an attacker-determined Trojan trigger. Backdoors compromise a model's integrity, thereby posing a severe threat to the landscape of DNN-based classification. While multiple defenses against such attacks exist for classifiers in the image domain, there have been limited efforts to protect classifiers in the text domain. We present Trojan-Miner (T-Miner) -- a defense framework for Trojan attacks on DNN-based text classifiers. T-Miner employs a sequence-to-sequence (seq-2-seq) generative model that probes the suspicious classifier and learns to produce text sequences that are likely to contain the Trojan trigger. T-Miner then analyzes the text produced by the generative model to determine if they contain trigger phrases, and correspondingly, whether the tested classifier has a backdoor. T-Miner requires no access to the training dataset or clean inputs of the suspicious classifier, and instead uses synthetically crafted "nonsensical" text inputs to train the generative model. We extensively evaluate T-Miner on 1100 model instances spanning 3 ubiquitous DNN model architectures, 5 different classification tasks, and a variety of trigger phrases. We show that T-Miner detects Trojan and clean models with a 98.75% overall accuracy, while achieving low false positives on clean models. We also show that T-Miner is robust against a variety of targeted, advanced attacks from an adaptive attacker.
Hierarchical text classification consists in classifying text documents into a hierarchy of classes and sub-classes. Although artificial neural networks have proved useful to perform this task, unfortunately they can leak training data information to adversaries due to training data memorization. Using differential privacy during model training can mitigate leakage attacks against trained models, enabling the models to be shared safely at the cost of reduced model accuracy. This work investigates the privacy-utility trade-off in hierarchical text classification with differential privacy guarantees, and identifies neural network architectures that offer superior trade-offs. To this end, we use a white-box membership inference attack to empirically assess the information leakage of three widely used neural network architectures. We show that large differential privacy parameters already suffice to completely mitigate membership inference attacks, thus resulting only in a moderate decrease in model utility. More specifically, for large datasets with long texts we observed Transformer-based models to achieve an overall favorable privacy-utility trade-off, while for smaller datasets with shorter texts convolutional neural networks are preferable.
Combining the representations of the words that make up a sentence into a cohesive whole is difficult, since it needs to account for the order of words, and to establish how the words present relate to each other. The solution we propose consists in iteratively adjusting the context. Our algorithm starts with a presumably erroneous value of the context, and adjusts this value with respect to the tokens at hand. In order to achieve this, representations of words are built combining their symbolic embedding with a positional encoding into single vectors. The algorithm then iteratively weighs and aggregates these vectors using our novel second-order attention mechanism. Our models report strong results in several well-known text classification tasks.
Contrastive explanations clarify why an event occurred in contrast to another. They are more inherently intuitive to humans to both produce and comprehend. We propose a methodology to produce contrastive explanations for classification models by modifying the representation to disregard non-contrastive information, and modifying model behavior to only be based on contrastive reasoning. Our method is based on projecting model representation to a latent space that captures only the features that are useful (to the model) to differentiate two potential decisions. We demonstrate the value of contrastive explanations by analyzing two different scenarios, using both high-level abstract concept attribution and low-level input token/span attribution, on two widely used text classification tasks. Specifically, we produce explanations for answering: for which label, and against which alternative label, is some aspect of the input useful? And which aspects of the input are useful for and against particular decisions? Overall, our findings shed light on the ability of label-contrastive explanations to provide a more accurate and finer-grained interpretability of a model's decision.
Bidirectional Encoder Representations from Transformers (BERT) have shown to be a promising way to dramatically improve the performance across various Natural Language Processing tasks [Devlin et al., 2019]. Meanwhile, progress made over the past few years by various Neural Net-work has also proved the effectiveness of Neural Network in the field of Natural Language Processing. In this project, RoBERTa-wwm-ext [Cui et al., 2019] pre-train language model was adopted and fine-tuned for Chinese text classification. The models were able to classify Chinese texts into two categories, containing descriptions of legal behavior and descriptions of illegal behavior. Four different models are also proposed in the paper. Those models will use RoBERTa-wwm-extas their embedding layer and feed the embedding into different neural networks. The motivation be-hind proposing these models is straightforward. By introducing complex output layer architecture, the overall performance of the models could be improved. All the models were trained on a data set derived from Chinese public court records, and the performance of different models were compared.The experiment shows that the performance of pro-posed models failed to beat the original RoBERTa-wwm-ext model in terms of accuracy and training efficiency.
Natural Language Processing algorithms have made incredible progress, but they still struggle when applied to out-of-distribution examples. We address a challenging and underexplored version of this domain adaptation problem, where an algorithm is trained on several source domains, and then applied to examples from unseen domains that are unknown at training time. Particularly, no examples, labeled or unlabeled, or any other knowledge about the target domain are available to the algorithm at training time. We present PADA: An example-based autoregressive Prompt learning algorithm for on-the-fly Any-Domain Adaptation, based on the T5 language model. Given a test example, PADA first generates a unique prompt for it and then, conditioned on this prompt, labels the example with respect to the NLP prediction task. PADA is trained to generate a prompt which is a token sequence of unrestricted length, consisting of Domain Related Features (DRFs) that characterize each of the source domains. Intuitively, the generated prompt is a unique signature that maps the test example to a semantic space spanned by the source domains. In experiments with 3 tasks (text classification and sequence tagging), for a total of 14 multi-source adaptation scenarios, PADA substantially outperforms strong baselines.
Language varies across users and their interested fields in social media data: words authored by a user across his/her interests may have different meanings (e.g., cool) or sentiments (e.g., fast). However, most of the existing methods to train user embeddings ignore the variations across user interests, such as product and movie categories (e.g., drama vs. action). In this study, we treat the user interest as domains and empirically examine how the user language can vary across the user factor in three English social media datasets. We then propose a user embedding model to account for the language variability of user interests via a multitask learning framework. The model learns user language and its variations without human supervision. While existing work mainly evaluated the user embedding by extrinsic tasks, we propose an intrinsic evaluation via clustering and evaluate user embeddings by an extrinsic task, text classification. The experiments on the three English-language social media datasets show that our proposed approach can generally outperform baselines via adapting the user factor.
Graph convolutional networks (GCNs) are a powerful architecture for representation learning on documents that naturally occur as graphs, e.g., citation or social networks. However, sensitive personal information, such as documents with people's profiles or relationships as edges, are prone to privacy leaks, as the trained model might reveal the original input. Although differential privacy (DP) offers a well-founded privacy-preserving framework, GCNs pose theoretical and practical challenges due to their training specifics. We address these challenges by adapting differentially-private gradient-based training to GCNs and conduct experiments using two optimizers on five NLP datasets in two languages. We propose a simple yet efficient method based on random graph splits that not only improves the baseline privacy bounds by a factor of 2.7 while retaining competitive F1 scores, but also provides strong privacy guarantees of epsilon = 1.0. We show that, under certain modeling choices, privacy-preserving GCNs perform up to 90% of their non-private variants, while formally guaranteeing strong privacy measures.
Fake News Detection is an essential problem in the field of Natural Language Processing. The benefits of an effective solution in this area are manifold for the goodwill of society. On a surface level, it broadly matches with the general problem of text classification. Researchers have proposed various approaches to tackle fake news using simple as well as some complex techniques. In this paper, we try to make a comparison between the present Deep Learning techniques by representing the news instances in some vector space using a combination of common mathematical operations with available vector space representations. We do a number of experiments using various combinations and permutations. Finally, we conclude with a sound analysis of the results and evaluate the reasons for such results.
The paper presents the submission of the team indicnlp@kgp to the EACL 2021 shared task "Offensive Language Identification in Dravidian Languages." The task aimed to classify different offensive content types in 3 code-mixed Dravidian language datasets. The work leverages existing state of the art approaches in text classification by incorporating additional data and transfer learning on pre-trained models. Our final submission is an ensemble of an AWD-LSTM based model along with 2 different transformer model architectures based on BERT and RoBERTa. We achieved weighted-average F1 scores of 0.97, 0.77, and 0.72 in the Malayalam-English, Tamil-English, and Kannada-English datasets ranking 1st, 2nd, and 3rd on the respective tasks.
Multi-label text classification refers to the problem of assigning each given document its most relevant labels from the label set. Commonly, the metadata of the given documents and the hierarchy of the labels are available in real-world applications. However, most existing studies focus on only modeling the text information, with a few attempts to utilize either metadata or hierarchy signals, but not both of them. In this paper, we bridge the gap by formalizing the problem of metadata-aware text classification in a large label hierarchy (e.g., with tens of thousands of labels). To address this problem, we present the MATCH solution -- an end-to-end framework that leverages both metadata and hierarchy information. To incorporate metadata, we pre-train the embeddings of text and metadata in the same space and also leverage the fully-connected attentions to capture the interrelations between them. To leverage the label hierarchy, we propose different ways to regularize the parameters and output probability of each child label by its parents. Extensive experiments on two massive text datasets with large-scale label hierarchies demonstrate the effectiveness of MATCH over state-of-the-art deep learning baselines.
Contextual advertising provides advertisers with the opportunity to target the context which is most relevant to their ads. However, its power cannot be fully utilized unless we can target the page content using fine-grained categories, e.g., "coupe" vs. "hatchback" instead of "automotive" vs. "sport". The widely used advertising content taxonomy (IAB taxonomy) consists of 23 coarse-grained categories and 355 fine-grained categories. With the large number of categories, it becomes very challenging either to collect training documents to build a supervised classification model, or to compose expert-written rules in a rule-based classification system. Besides, in fine-grained classification, different categories often overlap or co-occur, making it harder to classify accurately. In this work, we propose wiki2cat, a method to tackle the problem of large-scaled fine-grained text classification by tapping on Wikipedia category graph. The categories in IAB taxonomy are first mapped to category nodes in the graph. Then the label is propagated across the graph to obtain a list of labeled Wikipedia documents to induce text classifiers. The method is ideal for large-scale classification problems since it does not require any manually-labeled document or hand-curated rules or keywords. The proposed method is benchmarked with various learning-based and keyword-based baselines and yields competitive performance on both publicly available datasets and a new dataset containing more than 300 fine-grained categories.
The concept of unsupervised universal sentence encoders has gained traction recently, wherein pre-trained models generate effective task-agnostic fixed-dimensional representations for phrases, sentences and paragraphs. Such methods are of varying complexity, from simple weighted-averages of word vectors to complex language-models based on bidirectional transformers. In this work we propose a novel technique to generate sentence-embeddings in an unsupervised fashion by projecting the sentences onto a fixed-dimensional manifold with the objective of preserving local neighbourhoods in the original space. To delineate such neighbourhoods we experiment with several set-distance metrics, including the recently proposed Word Mover's distance, while the fixed-dimensional projection is achieved by employing a scalable and efficient manifold approximation method rooted in topological data analysis. We test our approach, which we term EMAP or Embeddings by Manifold Approximation and Projection, on six publicly available text-classification datasets of varying size and complexity. Empirical results show that our method consistently performs similar to or better than several alternative state-of-the-art approaches.
In industry NLP application, our manually labeled data has a certain number of noisy data. We present a simple method to find the noisy data and re-label them manually, meanwhile we collect the correction information. Then we present novel method to incorporate the human correction information into deep learning model. Human know how to correct noisy data. So the correction information can be inject into deep learning model. We do the experiment on our own text classification dataset, which is manually labeled, because we re-label the noisy data in our dataset for our industry application. The experiment result shows that our method improve the classification accuracy from 91.7% to 92.5%. The 91.7% accuracy is trained on the corrected dataset, which improve the baseline from 83.3% to 91.7%.
Biased associations have been a challenge in the development of classifiers for detecting toxic language, hindering both fairness and accuracy. As potential solutions, we investigate recently introduced debiasing methods for text classification datasets and models, as applied to toxic language detection. Our focus is on lexical (e.g., swear words, slurs, identity mentions) and dialectal markers (specifically African American English). Our comprehensive experiments establish that existing methods are limited in their ability to prevent biased behavior in current toxicity detectors. We then propose an automatic, dialect-aware data correction method, as a proof-of-concept. Despite the use of synthetic labels, this method reduces dialectal associations with toxicity. Overall, our findings show that debiasing a model trained on biased toxic language data is not as effective as simply relabeling the data to remove existing biases.
In this work, we propose the use of a fully managed machine learning service, which utilizes active learning to directly build models from unstructured data. With this tool, business users can quickly and easily build machine learning models and then directly deploy them into a production ready hosted environment without much involvement from data scientists. Our approach leverages state-of-the-art text representation like OpenAI's GPT2 and a fast implementation of the active learning workflow that relies on a simple construction of incremental learning using linear models, thus providing a brisk and efficient labeling experience for the users. Experiments on both publicly available and real-life insurance datasets empirically show why our choices of simple and fast classification algorithms are ideal for the task at hand.
Modern classification models tend to struggle when the amount of annotated data is scarce. To overcome this issue, several neural few-shot classification models have emerged, yielding significant progress over time, both in Computer Vision and Natural Language Processing. In the latter, such models used to rely on fixed word embeddings before the advent of transformers. Additionally, some models used in Computer Vision are yet to be tested in NLP applications. In this paper, we compare all these models, first adapting those made in the field of image processing to NLP, and second providing them access to transformers. We then test these models equipped with the same transformer-based encoder on the intent detection task, known for having a large number of classes. Our results reveal that while methods perform almost equally on the ARSC dataset, this is not the case for the Intent Detection task, where the most recent and supposedly best competitors perform worse than older and simpler ones (while all are given access to transformers). We also show that a simple baseline is surprisingly strong. All the new developed models, as well as the evaluation framework, are made publicly available.