Contrastive learning has achieved remarkable success in representation learning via self-supervision in unsupervised settings. However, effectively adapting contrastive learning to supervised learning tasks remains as a challenge in practice. In this work, we introduce a dual contrastive learning (DualCL) framework that simultaneously learns the features of input samples and the parameters of classifiers in the same space. Specifically, DualCL regards the parameters of the classifiers as augmented samples associating to different labels and then exploits the contrastive learning between the input samples and the augmented samples. Empirical studies on five benchmark text classification datasets and their low-resource version demonstrate the improvement in classification accuracy and confirm the capability of learning discriminative representations of DualCL.
Existing textual adversarial attacks usually utilize the gradient or prediction confidence to generate adversarial examples, making it hard to be deployed in real-world applications. To this end, we consider a rarely investigated but more rigorous setting, namely hard-label attack, in which the attacker can only access the prediction label. In particular, we find we can learn the importance of different words via the change on prediction label caused by word substitutions on the adversarial examples. Based on this observation, we propose a novel adversarial attack, termed Text Hard-label attacker (TextHacker). TextHacker randomly perturbs lots of words to craft an adversarial example. Then, TextHacker adopts a hybrid local search algorithm with the estimation of word importance from the attack history to minimize the adversarial perturbation. Extensive evaluations for text classification and textual entailment show that TextHacker significantly outperforms existing hard-label attacks regarding the attack performance as well as adversary quality.
Unsupervised domain adaptation (UDA) generally aligns the unlabeled target domain data to the distribution of the source domain to mitigate the distribution shift problem. The standard UDA requires sharing the source data with the target, having potential data privacy leaking risks. To protect the source data's privacy, we first propose to share the source feature distribution instead of the source data. However, sharing only the source feature distribution may still suffer from the membership inference attack who can infer an individual's membership by the black-box access to the source model. To resolve this privacy issue, we further study the under-explored problem of privacy-preserving domain adaptation and propose a method with a novel differential privacy training strategy to protect the source data privacy. We model the source feature distribution by Gaussian Mixture Models (GMMs) under the differential privacy setting and send it to the target client for adaptation. The target client resamples differentially private source features from GMMs and adapts on target data with several state-of-art UDA backbones. With our proposed method, the source data provider could avoid leaking source data privacy during domain adaptation as well as reserve the utility. To evaluate our proposed method's utility and privacy loss, we apply our model on a medical report disease label classification task using two noisy challenging clinical text datasets. The results show that our proposed method can preserve source data's privacy with a minor performance influence on the text classification task.
Fake news detection research is still in the early stage as this is a relatively new phenomenon in the interest raised by society. Machine learning helps to solve complex problems and to build AI systems nowadays and especially in those cases where we have tacit knowledge or the knowledge that is not known. We used machine learning algorithms and for identification of fake news; we applied three classifiers; Passive Aggressive, Na\"ive Bayes, and Support Vector Machine. Simple classification is not completely correct in fake news detection because classification methods are not specialized for fake news. With the integration of machine learning and text-based processing, we can detect fake news and build classifiers that can classify the news data. Text classification mainly focuses on extracting various features of text and after that incorporating those features into classification. The big challenge in this area is the lack of an efficient way to differentiate between fake and non-fake due to the unavailability of corpora. We applied three different machine learning classifiers on two publicly available datasets. Experimental analysis based on the existing dataset indicates a very encouraging and improved performance.
In this paper, we study the "stability" of machine learning (ML) models within the context of larger, complex NLP systems with continuous training data updates. For this study, we propose a methodology for the assessment of model stability (which we refer to as jitter under various experimental conditions. We find that model design choices, including network architecture and input representation, have a critical impact on stability through experiments on four text classification tasks and two sequence labeling tasks. In classification tasks, non-RNN-based models are observed to be more stable than RNN-based ones, while the encoder-decoder model is less stable in sequence labeling tasks. Moreover, input representations based on pre-trained fastText embeddings contribute to more stability than other choices. We also show that two learning strategies -- ensemble models and incremental training -- have a significant influence on stability. We recommend ML model designers account for trade-offs in accuracy and jitter when making modeling choices.
Text classification algorithms investigate the intricate relationships between words or phrases and attempt to deduce the document's interpretation. In the last few years, these algorithms have progressed tremendously. Transformer architecture and sentence encoders have proven to give superior results on natural language processing tasks. But a major limitation of these architectures is their applicability for text no longer than a few hundred words. In this paper, we explore hierarchical transfer learning approaches for long document classification. We employ pre-trained Universal Sentence Encoder (USE) and Bidirectional Encoder Representations from Transformers (BERT) in a hierarchical setup to capture better representations efficiently. Our proposed models are conceptually simple where we divide the input data into chunks and then pass this through base models of BERT and USE. Then output representation for each chunk is then propagated through a shallow neural network comprising of LSTMs or CNNs for classifying the text data. These extensions are evaluated on 6 benchmark datasets. We show that USE + CNN/LSTM performs better than its stand-alone baseline. Whereas the BERT + CNN/LSTM performs on par with its stand-alone counterpart. However, the hierarchical BERT models are still desirable as it avoids the quadratic complexity of the attention mechanism in BERT. Along with the hierarchical approaches, this work also provides a comparison of different deep learning algorithms like USE, BERT, HAN, Longformer, and BigBird for long document classification. The Longformer approach consistently performs well on most of the datasets.
Active learning is an increasingly important branch of machine learning and a powerful technique for natural language processing. The main advantage of active learning is its potential to reduce the amount of labeled data needed to learn high-performing models. A vital aspect of an effective active learning algorithm is the determination of when to stop obtaining additional labeled data. Several leading state-of-the-art stopping methods use a stop set to help make this decision. However, there has been relatively less attention given to the choice of stop set than to the stopping algorithms that are applied on the stop set. Different choices of stop sets can lead to significant differences in stopping method performance. We investigate the impact of different stop set choices on different stopping methods. This paper shows the choice of the stop set can have a significant impact on the performance of stopping methods and the impact is different for stability-based methods from that on confidence-based methods. Furthermore, the unbiased representative stop sets suggested by original authors of methods work better than the systematically biased stop sets used in recently published work, and stopping methods based on stabilizing predictions have stronger performance than confidence-based stopping methods when unbiased representative stop sets are used. We provide the largest quantity of experimental results on the impact of stop sets to date. The findings are important for helping to illuminate the impact of this important aspect of stopping methods that has been under-considered in recently published work and that can have a large practical impact on the performance of stopping methods for important semantic computing applications such as technology assisted review and text classification more broadly.
Recent advances on prompt-tuning cast few-shot classification tasks as a masked language modeling problem. By wrapping input into a template and using a verbalizer which constructs a mapping between label space and label word space, prompt-tuning can achieve excellent results in zero-shot and few-shot scenarios. However, typical prompt-tuning needs a manually designed verbalizer which requires domain expertise and human efforts. And the insufficient label space may introduce considerable bias into the results. In this paper, we focus on eliciting knowledge from pretrained language models and propose a prototypical prompt verbalizer for prompt-tuning. Labels are represented by prototypical embeddings in the feature space rather than by discrete words. The distances between the embedding at the masked position of input and prototypical embeddings are used as classification criterion. For zero-shot settings, knowledge is elicited from pretrained language models by a manually designed template to form initial prototypical embeddings. For few-shot settings, models are tuned to learn meaningful and interpretable prototypical embeddings. Our method optimizes models by contrastive learning. Extensive experimental results on several many-class text classification datasets with low-resource settings demonstrate the effectiveness of our approach compared with other verbalizer construction methods. Our implementation is available at https://github.com/Ydongd/prototypical-prompt-verbalizer.
The Gutenberg Literary English Corpus (GLEC, Jacobs, 2018a) provides a rich source of textual data for research in digital humanities, computational linguistics or neurocognitive poetics. In this study we address differences among the different literature categories in GLEC, as well as differences between authors. We report the results of three studies providing i) topic and sentiment analyses for six text categories of GLEC (i.e., children and youth, essays, novels, plays, poems, stories) and its >100 authors, ii) novel measures of semantic complexity as indices of the literariness, creativity and book beauty of the works in GLEC (e.g., Jane Austen's six novels), and iii) two experiments on text classification and authorship recognition using novel features of semantic complexity. The data on two novel measures estimating a text's literariness, intratextual variance and stepwise distance (van Cranenburgh et al., 2019) revealed that plays are the most literary texts in GLEC, followed by poems and novels. Computation of a novel index of text creativity (Gray et al., 2016) revealed poems and plays as the most creative categories with the most creative authors all being poets (Milton, Pope, Keats, Byron, or Wordsworth). We also computed a novel index of perceived beauty of verbal art (Kintsch, 2012) for the works in GLEC and predict that Emma is the theoretically most beautiful of Austen's novels. Finally, we demonstrate that these novel measures of semantic complexity are important features for text classification and authorship recognition with overall predictive accuracies in the range of .75 to .97. Our data pave the way for future computational and empirical studies of literature or experiments in reading psychology and offer multiple baselines and benchmarks for analysing and validating other book corpora.
BERT has revolutionized the NLP field by enabling transfer learning with large language models that can capture complex textual patterns, reaching the state-of-the-art for an expressive number of NLP applications. For text classification tasks, BERT has already been extensively explored. However, aspects like how to better cope with the different embeddings provided by the BERT output layer and the usage of language-specific instead of multilingual models are not well studied in the literature, especially for the Brazilian Portuguese language. The purpose of this article is to conduct an extensive experimental study regarding different strategies for aggregating the features produced in the BERT output layer, with a focus on the sentiment analysis task. The experiments include BERT models trained with Brazilian Portuguese corpora and the multilingual version, contemplating multiple aggregation strategies and open-source datasets with predefined training, validation, and test partitions to facilitate the reproducibility of the results. BERT achieved the highest ROC-AUC values for the majority of cases as compared to TF-IDF. Nonetheless, TF-IDF represents a good trade-off between the predictive performance and computational cost.
Adverse drug reactions / events (ADR/ADE) have a major impact on patient health and health care costs. Detecting ADR's as early as possible and sharing them with regulators, pharma companies, and healthcare providers can prevent morbidity and save many lives. While most ADR's are not reported via formal channels, they are often documented in a variety of unstructured conversations such as social media posts by patients, customer support call transcripts, or CRM notes of meetings between healthcare providers and pharma sales reps. In this paper, we propose a natural language processing (NLP) solution that detects ADR's in such unstructured free-text conversations, which improves on previous work in three ways. First, a new Named Entity Recognition (NER) model obtains new state-of-the-art accuracy for ADR and Drug entity extraction on the ADE, CADEC, and SMM4H benchmark datasets (91.75%, 78.76%, and 83.41% F1 scores respectively). Second, two new Relation Extraction (RE) models are introduced - one based on BioBERT while the other utilizing crafted features over a Fully Connected Neural Network (FCNN) - are shown to perform on par with existing state-of-the-art models, and outperform them when trained with a supplementary clinician-annotated RE dataset. Third, a new text classification model, for deciding if a conversation includes an ADR, obtains new state-of-the-art accuracy on the CADEC dataset (86.69% F1 score). The complete solution is implemented as a unified NLP pipeline in a production-grade library built on top of Apache Spark, making it natively scalable and able to process millions of batch or streaming records on commodity clusters.
Traditional text classification approaches often require a good amount of labeled data, which is difficult to obtain, especially in restricted domains or less widespread languages. This lack of labeled data has led to the rise of low-resource methods, that assume low data availability in natural language processing. Among them, zero-shot learning stands out, which consists of learning a classifier without any previously labeled data. The best results reported with this approach use language models such as Transformers, but fall into two problems: high execution time and inability to handle long texts as input. This paper proposes a new model, ZeroBERTo, which leverages an unsupervised clustering step to obtain a compressed data representation before the classification task. We show that ZeroBERTo has better performance for long inputs and shorter execution time, outperforming XLM-R by about 12% in the F1 score in the FolhaUOL dataset. Keywords: Low-Resource NLP, Unlabeled data, Zero-Shot Learning, Topic Modeling, Transformers.
While pre-trained language model (PLM) fine-tuning has achieved strong performance in many NLP tasks, the fine-tuning stage can be still demanding in labeled data. Recent works have resorted to active fine-tuning to improve the label efficiency of PLM fine-tuning, but none of them investigate the potential of unlabeled data. We propose {\ours}, a new framework that leverages unlabeled data to improve the label efficiency of active PLM fine-tuning. AcTune switches between data annotation and model self-training based on uncertainty: it selects high-uncertainty unlabeled samples for active annotation and low-uncertainty ones for model self-training. Under this framework, we design (1) a region-aware sampling strategy that reduces redundancy when actively querying for annotations and (2) a momentum-based memory bank that dynamically aggregates the model's pseudo labels to suppress label noise in self-training. Experiments on 6 text classification datasets show that AcTune outperforms the strongest active learning and self-training baselines and improves the label efficiency of PLM fine-tuning by 56.2\% on average. Our implementation will be available at \url{https://github.com/yueyu1030/actune}.
Classification of posts in social media such as Twitter is difficult due to the noisy and short nature of texts. Sequence classification models based on recurrent neural networks (RNN) are popular for classifying posts that are sequential in nature. RNNs assume the hidden representation dynamics to evolve in a discrete manner and do not consider the exact time of the posting. In this work, we propose to use recurrent neural ordinary differential equations (RNODE) for social media post classification which consider the time of posting and allow the computation of hidden representation to evolve in a time-sensitive continuous manner. In addition, we propose a novel model, Bi-directional RNODE (Bi-RNODE), which can consider the information flow in both the forward and backward directions of posting times to predict the post label. Our experiments demonstrate that RNODE and Bi-RNODE are effective for the problem of stance classification of rumours in social media.
Finding a suitable job and hunting for eligible candidates are important to job seeking and human resource agencies. With the vast information about job descriptions, employees and employers need assistance to automatically detect job titles based on job description texts. In this paper, we propose the multi-label classification approach for predicting relevant job titles from job description texts, and implement the Bi-GRU-LSTM-CNN with different pre-trained language models to apply for the job titles prediction problem. The BERT with multilingual pre-trained model obtains the highest result by F1-scores on both development and test sets, which are 62.20% on the development set, and 47.44% on the test set.
We perform knowledge distillation (KD) benchmark from task-specific BERT-base teacher models to various student models: BiLSTM, CNN, BERT-Tiny, BERT-Mini, and BERT-Small. Our experiment involves 12 datasets grouped in two tasks: text classification and sequence labeling in the Indonesian language. We also compare various aspects of distillations including the usage of word embeddings and unlabeled data augmentation. Our experiments show that, despite the rising popularity of Transformer-based models, using BiLSTM and CNN student models provide the best trade-off between performance and computational resource (CPU, RAM, and storage) compared to pruned BERT models. We further propose some quick wins on performing KD to produce small NLP models via efficient KD training mechanisms involving simple choices of loss functions, word embeddings, and unlabeled data preparation.
Commonsense question answering (CQA) aims to test if models can answer questions regarding commonsense knowledge that everyone knows. Prior works that incorporate external knowledge bases have shown promising results, but knowledge bases are expensive to construct and are often limited to a fixed set of relations. In this paper, we instead focus on better utilizing the \textit{implicit knowledge} stored in pre-trained language models. While researchers have found that the knowledge embedded in pre-trained language models can be extracted by having them fill in the blanks of carefully designed prompts for relation extraction and text classification, it remains unclear if we can adopt this paradigm in CQA where the inputs and outputs take much more flexible forms. To this end, we investigate four translation methods that can translate natural questions into cloze-style sentences to better solicit commonsense knowledge from language models, including a syntactic-based model, an unsupervised neural model, and two supervised neural models. In addition, to combine the different translation methods, we propose to encourage consistency among model predictions on different translated questions with unlabeled data. We demonstrate the effectiveness of our methods on three CQA datasets in zero-shot settings. We show that our methods are complementary to a knowledge base improved model, and combining them can lead to state-of-the-art zero-shot performance. Analyses also reveal distinct characteristics of the different cloze translation methods and provide insights on why combining them can lead to great improvements.
Despite the success of mixup in data augmentation, its applicability to natural language processing (NLP) tasks has been limited due to the discrete and variable-length nature of natural languages. Recent studies have thus relied on domain-specific heuristics and manually crafted resources, such as dictionaries, in order to apply mixup in NLP. In this paper, we instead propose an unsupervised learning approach to text interpolation for the purpose of data augmentation, to which we refer as "Learning to INterpolate for Data Augmentation" (LINDA), that does not require any heuristics nor manually crafted resources but learns to interpolate between any pair of natural language sentences over a natural language manifold. After empirically demonstrating the LINDA's interpolation capability, we show that LINDA indeed allows us to seamlessly apply mixup in NLP and leads to better generalization in text classification both in-domain and out-of-domain.
Text classification is a fundamental Natural Language Processing task that has a wide variety of applications, where deep learning approaches have produced state-of-the-art results. While these models have been heavily criticized for their black-box nature, their robustness to slight perturbations in input text has been a matter of concern. In this work, we carry out a data-focused study evaluating the impact of systematic practical perturbations on the performance of the deep learning based text classification models like CNN, LSTM, and BERT-based algorithms. The perturbations are induced by the addition and removal of unwanted tokens like punctuation and stop-words that are minimally associated with the final performance of the model. We show that these deep learning approaches including BERT are sensitive to such legitimate input perturbations on four standard benchmark datasets SST2, TREC-6, BBC News, and tweet_eval. We observe that BERT is more susceptible to the removal of tokens as compared to the addition of tokens. Moreover, LSTM is slightly more sensitive to input perturbations as compared to CNN based model. The work also serves as a practical guide to assessing the impact of discrepancies in train-test conditions on the final performance of models.
Rhetorical Frames in AI can be thought of as expressions that describe AI development as a competition between two or more actors, such as governments or companies. Examples of such Frames include robotic arms race, AI rivalry, technological supremacy, cyberwarfare dominance and 5G race. Detection of Rhetorical Frames from open sources can help us track the attitudes of governments or companies towards AI, specifically whether attitudes are becoming more cooperative or competitive over time. Given the rapidly increasing volumes of open sources (online news media, twitter, blogs), it is difficult for subject matter experts to identify Rhetorical Frames in (near) real-time. Moreover, these sources are in general unstructured (noisy) and therefore, detecting Frames from these sources will require state-of-the-art text classification techniques. In this paper, we develop RheFrameDetect, a text classification system for (near) real-time capture of Rhetorical Frames from open sources. Given an input document, RheFrameDetect employs text classification techniques at multiple levels (document level and paragraph level) to identify all occurrences of Frames used in the discussion of AI. We performed extensive evaluation of the text classification techniques used in RheFrameDetect against human annotated Frames from multiple news sources. To further demonstrate the effectiveness of RheFrameDetect, we show multiple case studies depicting the Frames identified by RheFrameDetect compared against human annotated Frames.