Pretraining large neural language models, such as BERT, has led to impressive gains on many natural language processing (NLP) tasks. However, most pretraining efforts focus on general domain corpora, such as newswire and Web. A prevailing assumption is that even domain-specific pretraining can benefit by starting from general-domain language models. In this paper, we challenge this assumption by showing that for domains with abundant unlabeled text, such as biomedicine, pretraining language models from scratch results in substantial gains over continual pretraining of general-domain language models. To facilitate this investigation, we compile a comprehensive biomedical NLP benchmark from publicly-available datasets. Our experiments show that domain-specific pretraining serves as a solid foundation for a wide range of biomedical NLP tasks, leading to new state-of-the-art results across the board. Further, in conducting a thorough evaluation of modeling choices, both for pretraining and task-specific fine-tuning, we discover that some common practices are unnecessary with BERT models, such as using complex tagging schemes in named entity recognition (NER). To help accelerate research in biomedical NLP, we have released our state-of-the-art pretrained and task-specific models for the community, and created a leaderboard featuring our BLURB benchmark (short for Biomedical Language Understanding & Reasoning Benchmark) at https://aka.ms/BLURB.
In cross-lingual text classification, one seeks to exploit labeled data from one language to train a text classification model that can then be applied to a completely different language. Recent multilingual representation models have made it much easier to achieve this. Still, there may still be subtle differences between languages that are neglected when doing so. To address this, we present a semi-supervised adversarial training process that minimizes the maximal loss for label-preserving input perturbations. The resulting model then serves as a teacher to induce labels for unlabeled target language samples that can be used during further adversarial training, allowing us to gradually adapt our model to the target language. Compared with a number of strong baselines, we observe significant gains in effectiveness on document and intent classification for a diverse set of languages.
The learning rate (LR) is one of the most important hyper-parameters in stochastic gradient descent (SGD) algorithm for training deep neural networks (DNN). However, current hand-designed LR schedules need to manually pre-specify a fixed form, which limits their ability to adapt practical non-convex optimization problems due to the significant diversification of training dynamics. Meanwhile, it always needs to search proper LR schedules from scratch for new tasks, which, however, are often largely different with task variations, like data modalities, network architectures, or training data capacities. To address this learning-rate-schedule setting issues, we propose to parameterize LR schedules with an explicit mapping formulation, called \textit{MLR-SNet}. The learnable parameterized structure brings more flexibility for MLR-SNet to learn a proper LR schedule to comply with the training dynamics of DNN. Image and text classification benchmark experiments substantiate the capability of our method for achieving proper LR schedules. Moreover, the explicit parameterized structure makes the meta-learned LR schedules capable of being transferable and plug-and-play, which can be easily generalized to new heterogeneous tasks. We transfer our meta-learned MLR-SNet to query tasks like different training epochs, network architectures, data modalities, dataset sizes from the training ones, and achieve comparable or even better performance compared with hand-designed LR schedules specifically designed for the query tasks. The robustness of MLR-SNet is also substantiated when the training data are biased with corrupted noise. We further prove the convergence of the SGD algorithm equipped with LR schedule produced by our MLR-Net, with the convergence rate comparable to the best-known ones of the algorithm for solving the problem.
Transformers-based models, such as BERT, have been one of the most successful deep learning models for NLP. Unfortunately, one of their core limitations is the quadratic dependency (mainly in terms of memory) on the sequence length due to their full attention mechanism. To remedy this, we propose, BigBird, a sparse attention mechanism that reduces this quadratic dependency to linear. We show that BigBird is a universal approximator of sequence functions and is Turing complete, thereby preserving these properties of the quadratic, full attention model. Along the way, our theoretical analysis reveals some of the benefits of having $O(1)$ global tokens (such as CLS), that attend to the entire sequence as part of the sparse attention mechanism. The proposed sparse attention can handle sequences of length up to 8x of what was previously possible using similar hardware. As a consequence of the capability to handle longer context, BigBird drastically improves performance on various NLP tasks such as question answering and summarization. We also propose novel applications to genomics data.
Concept drift is a phenomenon in which the distribution of a data stream changes over time in unforeseen ways, causing prediction models built on historical data to become inaccurate. While a variety of automated methods have been developed to identify when concept drift occurs, there is limited support for analysts who need to understand and correct their models when drift is detected. In this paper, we present a visual analytics method, DriftVis, to support model builders and analysts in the identification and correction of concept drift in streaming data. DriftVis combines a distribution-based drift detection method with a streaming scatterplot to support the analysis of drift caused by the distribution changes of data streams and to explore the impact of these changes on the model's accuracy. A quantitative experiment and two case studies on weather prediction and text classification have been conducted to demonstrate our proposed tool and illustrate how visual analytics can be used to support the detection, examination, and correction of concept drift.
As the performance and popularity of deep neural networks has increased, so too has their computational cost. There are many effective techniques for reducing a network's computational footprint (quantisation, pruning, knowledge distillation), but these lead to models whose computational cost is the same regardless of their input. Our human reaction times vary with the complexity of the tasks we perform: easier tasks (e.g. telling apart dogs from boat) are executed much faster than harder ones (e.g. telling apart two similar looking breeds of dogs). Driven by this observation, we develop a method for adaptive network complexity by attaching a small classification layer, which we call SideNet, to a large pretrained network, which we call MainNet. Given an input, the SideNet returns a classification if its confidence level, obtained via softmax, surpasses a user determined threshold, and only passes it along to the large MainNet for further processing if its confidence is too low. This allows us to flexibly trade off the network's performance with its computational cost. Experimental results show that simple single hidden layer perceptron SideNets added onto pretrained ResNet and BERT MainNets allow for substantial decreases in compute with minimal drops in performance on image and text classification tasks. We also highlight three other desirable properties of our method, namely that the classifications obtained by SideNets are calibrated, complementary to other compute reduction techniques, and that they enable the easy exploration of compute accuracy space.
Processing of raw text is the crucial first step in text classification and sentiment analysis. However, text processing steps are often performed using off-the-shelf routines and pre-built word dictionaries without optimizing for domain, application, and context. This paper investigates the effect of seven text processing scenarios on a particular text domain (Twitter) and application (sentiment classification). Skip gram-based word embeddings are developed to include Twitter colloquial words, emojis, and hashtag keywords that are often removed for being unavailable in conventional literature corpora. Our experiments reveal negative effects on sentiment classification of two common text processing steps: 1) stop word removal and 2) averaging of word vectors to represent individual tweets. New effective steps for 1) including non-ASCII emoji characters, 2) measuring word importance from word embedding, 3) aggregating word vectors into a tweet embedding, and 4) developing linearly separable feature space have been proposed to optimize the sentiment classification pipeline. The best combination of text processing steps yields the highest average area under the curve (AUC) of 88.4 (+/-0.4) in classifying 14,640 tweets with three sentiment labels. Word selection from context-driven word embedding reveals that only the ten most important words in Tweets cumulatively yield over 98% of the maximum accuracy. Results demonstrate a means for data-driven selection of important words in tweet classification as opposed to using pre-built word dictionaries. The proposed tweet embedding is robust to and alleviates the need for several text processing steps.
This paper describes our contribution to SemEval-2020 Task 11: Detection Of Propaganda Techniques In News Articles. We start with simple LSTM baselines and move to an autoregressive transformer decoder to predict long continuous propaganda spans for the first subtask. We also adopt an approach from relation extraction by enveloping spans mentioned above with special tokens for the second subtask of propaganda technique classification. Our models report an F-score of 44.6% and a micro-averaged F-score of 58.2% for those tasks accordingly.
We consider Multiclass and Multilabel classification with extremely large number of classes, of which only few are labeled to each instance. In such setting, standard methods that have training, prediction cost linear to the number of classes become intractable. State-of-the-art methods thus aim to reduce the complexity by exploiting correlation between labels under assumption that the similarity between labels can be captured by structures such as low-rank matrix or balanced tree. However, as the diversity of labels increases in the feature space, structural assumption can be easily violated, which leads to degrade in the testing performance. In this work, we show that a margin-maximizing loss with l1 penalty, in case of Extreme Classification, yields extremely sparse solution both in primal and in dual without sacrificing the expressive power of predictor. We thus propose a Fully-Corrective Block-Coordinate Frank-Wolfe (FC-BCFW) algorithm that exploits both primal and dual sparsity to achieve a complexity sublinear to the number of primal and dual variables. A bi-stochastic search method is proposed to further improve the efficiency. In our experiments on both Multiclass and Multilabel problems, the proposed method achieves significant higher accuracy than existing approaches of Extreme Classification with very competitive training and prediction time.
It has been proved that deep neural networks are facing a new threat called backdoor attacks, where the adversary can inject backdoors into the neural network model through poisoning the training dataset. When the input containing some special pattern called the backdoor trigger, the model with backdoor will carry out malicious task such as misclassification specified by adversaries. In text classification systems, backdoors inserted in the models can cause spam or malicious speech to escape detection. Previous work mainly focused on the defense of backdoor attacks in computer vision, little attention has been paid to defense method for RNN backdoor attacks regarding text classification. In this paper, through analyzing the changes in inner LSTM neurons, we proposed a defense method called Backdoor Keyword Identification (BKI) to mitigate backdoor attacks which the adversary performs against LSTM-based text classification by data poisoning. This method can identify and exclude poisoning samples crafted to insert backdoor into the model from training data without a verified and trusted dataset. We evaluate our method on four different text classification datset: IMDB, DBpedia ontology, 20 newsgroups and Reuters-21578 dataset. It all achieves good performance regardless of the trigger sentences.
Problems involving code-mixed language are often plagued by a lack of resources and an absence of materials to perform sophisticated transfer learning with. In this paper we describe our submission to the Sentimix Hindi-English task involving sentiment classification of code-mixed texts, and with an F1 score of 67.1%, we demonstrate that simple convolution and attention may well produce reasonable results.
Sentiment analysis is the task of automatic analysis of opinions and emotions of users towards an entity or some aspect of that entity. Political Sentiment Analysis of social media helps the political strategists to scrutinize the performance of a party or candidate and improvise their weaknesses far before the actual elections. During the time of elections, the social networks get flooded with blogs, chats, debates and discussions about the prospects of political parties and politicians. The amount of data generated is much large to study, analyze and draw inferences using the latest techniques. Twitter is one of the most popular social media platforms enables us to perform domain-specific data preparation. In this work, we chose to identify the inclination of political opinions present in Tweets by modelling it as a text classification problem using classical machine learning. The tweets related to the Delhi Elections in 2020 are extracted and employed for the task. Among the several algorithms, we observe that Support Vector Machines portrays the best performance.
Mining a set of meaningful topics organized into a hierarchy is intuitively appealing since topic correlations are ubiquitous in massive text corpora. To account for potential hierarchical topic structures, hierarchical topic models generalize flat topic models by incorporating latent topic hierarchies into their generative modeling process. However, due to their purely unsupervised nature, the learned topic hierarchy often deviates from users' particular needs or interests. To guide the hierarchical topic discovery process with minimal user supervision, we propose a new task, Hierarchical Topic Mining, which takes a category tree described by category names only, and aims to mine a set of representative terms for each category from a text corpus to help a user comprehend his/her interested topics. We develop a novel joint tree and text embedding method along with a principled optimization procedure that allows simultaneous modeling of the category tree structure and the corpus generative process in the spherical space for effective category-representative term discovery. Our comprehensive experiments show that our model, named JoSH, mines a high-quality set of hierarchical topics with high efficiency and benefits weakly-supervised hierarchical text classification tasks.
This paper presents the impact of using quantization on the efficiency of multi-class text classification in the training process of a support vector machine (SVM). This work is focused on comparing the efficiency of SVM model trained using reduced precision with its original form. The main advantage of using quantization is decrease in computation time and in memory footprint on the dedicated hardware platform which supports low precision computation like GPU (16-bit) or FPGA (any bit-width). The paper presents the impact of a precision reduction of the SVM training process on text classification accuracy. The implementation of the CPU was performed using the OpenMP library. Additionally, the results of the implementation of the GPU using double, single and half precision are presented.
Document-level Sentiment Analysis (DSA) is more challenging due to vague semantic links and complicate sentiment information. Recent works have been devoted to leveraging text summarization and have achieved promising results. However, these summarization-based methods did not take full advantage of the summary including ignoring the inherent interactions between the summary and document. As a result, they limited the representation to express major points in the document, which is highly indicative of the key sentiment. In this paper, we study how to effectively generate a discriminative representation with explicit subject patterns and sentiment contexts for DSA. A Hierarchical Interaction Networks (HIN) is proposed to explore bidirectional interactions between the summary and document at multiple granularities and learn subject-oriented document representations for sentiment classification. Furthermore, we design a Sentiment-based Rethinking mechanism (SR) by refining the HIN with sentiment label information to learn a more sentiment-aware document representation. We extensively evaluate our proposed models on three public datasets. The experimental results consistently demonstrate the effectiveness of our proposed models and show that HIN-SR outperforms various state-of-the-art methods.
In this paper we study the problem of predicting clinical diagnoses from textual Electronic Health Records (EHR) data. We show the importance of this problem in medical community and present comprehensive historical review of the problem and proposed methods. As the main scientific contributions we present a modification of Bidirectional Encoder Representations from Transformers (BERT) model for sequence classification that implements a novel way of Fully-Connected (FC) layer composition and a BERT model pretrained only on domain data. To empirically validate our model, we use a large-scale Russian EHR dataset consisting of about 4 million unique patient visits. This is the largest such study for the Russian language and one of the largest globally. We performed a number of comparative experiments with other text representation models on the task of multiclass classification for 265 disease subset of ICD-10. The experiments demonstrate improved performance of our models compared to other baselines, including a fine-tuned Russian BERT (RuBERT) variant. We also show comparable performance of our model with a panel of experienced medical experts. This allows us to hope that implementation of this system will reduce misdiagnosis.
Government-sponsored policy-making and scheme generations is one of the means of protecting and promoting the social, economic, and personal development of the citizens. The evaluation of effectiveness of these schemes done by government only provide the statistical information in terms of facts and figures which do not include the in-depth knowledge of public perceptions, experiences and views on the topic. In this research work, we propose an improved text classification framework that classifies the Twitter data of different health-based government schemes. The proposed framework leverages the language representation models (LR models) BERT, ELMO, and USE. However, these LR models have less real-time applicability due to the scarcity of the ample annotated data. To handle this, we propose a novel GloVe word embeddings and class-specific sentiments based text augmentation approach (named Mod-EDA) which boosts the performance of text classification task by increasing the size of labeled data. Furthermore, the trained model is leveraged to identify the level of engagement of citizens towards these policies in different communities such as middle-income and low-income groups.
Several text classification tasks such as sentiment analysis, news categorization, multi-label classification and opinion classification are challenging problems even for modern deep learning networks. Recently, Capsule Networks (CapsNets) are proposed for image classification. It has been shown that CapsNets have several advantages over Convolutional Neural Networks (CNNs), while their validity in the domain of text has been less explored. In this paper, we propose a novel hybrid architecture viz., BGCapsule, which is a Capsule model preceded by an ensemble of Bidirectional Gated Recurrent Units (BiGRU) for several text classification tasks. We employed an ensemble of Bidirectional GRUs for feature extraction layer preceding the primary capsule layer. The hybrid architecture, after performing basic pre-processing steps, consists of five layers: an embedding layer based on GloVe, a BiGRU based ensemble layer, a primary capsule layer, a flatten layer and fully connected ReLU layer followed by a fully connected softmax layer. In order to evaluate the effectiveness of BGCapsule, we conducted extensive experiments on five benchmark datasets (ranging from 10,000 records to 700,000 records) including Movie Review (MR Imdb 2005), AG News dataset, Dbpedia ontology dataset, Yelp Review Full dataset and Yelp review polarity dataset. These benchmarks cover several text classification tasks such as news categorization, sentiment analysis, multiclass classification, multi-label classification and opinion classification. We found that our proposed architecture (BGCapsule) achieves better accuracy compared to the existing methods without the help of any external linguistic knowledge such as positive sentiment keywords and negative sentiment keywords. Further, BGCapsule converged faster compared to other extant techniques.
Text data augmentation, i.e., the creation of new textual data from an existing text, is challenging. Indeed, augmentation transformations should take into account language complexity while being relevant to the target Natural Language Processing (NLP) task (e.g., Machine Translation, Text Classification). Initially motivated by an application of Business Email Compromise (BEC) detection, we propose a corpus and task agnostic augmentation framework used as a service to augment English texts within our company. Our proposal combines different methods, utilizing BERT language model, multi-step back-translation and heuristics. We show that our augmentation framework improves performances on several text classification tasks using publicly available models and corpora as well as on a BEC detection task. We also provide a comprehensive argumentation about the limitations of our augmentation framework.
Extreme multi-label text classification (XMTC) is a task for tagging a given text with the most relevant labels from an extremely large label set. We propose a novel deep learning method called APLC-XLNet. Our approach fine-tunes the recently released generalized autoregressive pretrained model (XLNet) to learn a dense representation for the input text. We propose Adaptive Probabilistic Label Clusters (APLC) to approximate the cross entropy loss by exploiting the unbalanced label distribution to form clusters that explicitly reduce the computational time. Our experiments, carried out on five benchmark datasets, show that our approach has achieved new state-of-the-art results on four benchmark datasets. Our source code is available publicly at https://github.com/huiyegit/APLC_XLNet.