In this paper, we present a study of the recent advancements which have helped bring Transfer Learning to NLP through the use of semi-supervised training. We discuss cutting-edge methods and architectures such as BERT, GPT, ELMo, ULMFit among others. Classically, tasks in natural language processing have been performed through rule-based and statistical methodologies. However, owing to the vast nature of natural languages these methods do not generalise well and failed to learn the nuances of language. Thus machine learning algorithms such as Naive Bayes and decision trees coupled with traditional models such as Bag-of-Words and N-grams were used to usurp this problem. Eventually, with the advent of advanced recurrent neural network architectures such as the LSTM, we were able to achieve state-of-the-art performance in several natural language processing tasks such as text classification and machine translation. We talk about how Transfer Learning has brought about the well-known ImageNet moment for NLP. Several advanced architectures such as the Transformer and its variants have allowed practitioners to leverage knowledge gained from unrelated task to drastically fasten convergence and provide better performance on the target task. This survey represents an effort at providing a succinct yet complete understanding of the recent advances in natural language processing using deep learning in with a special focus on detailing transfer learning and its potential advantages.
Various deep learning algorithms have been developed to analyze different types of clinical data including clinical text classification and extracting information from 'free text' and so on. However, automate the keyword extraction from the clinical notes is still challenging. The challenges include dealing with noisy clinical notes which contain various abbreviations, possible typos, and unstructured sentences. The objective of this research is to investigate the attention-based deep learning models to classify the de-identified clinical progress notes extracted from a real-world EHR system. The attention-based deep learning models can be used to interpret the models and understand the critical words that drive the correct or incorrect classification of the clinical progress notes. The attention-based models in this research are capable of presenting the human interpretable text classification models. The results show that the fine-tuned BERT with the attention layer can achieve a high classification accuracy of 97.6%, which is higher than the baseline fine-tuned BERT classification model. In this research, we also demonstrate that the attention-based models can identify relevant keywords that are strongly related to the clinical progress note categories.
Contemporary datasets on tobacco consumption focus on one of two topics, either public health mentions and disease surveillance, or sentiment analysis on topical tobacco products and services. However, two primary considerations are not accounted for, the language of the demographic affected and a combination of the topics mentioned above in a fine-grained classification mechanism. In this paper, we create a dataset of 3144 tweets, which are selected based on the presence of colloquial slang related to smoking and analyze it based on the semantics of the tweet. Each class is created and annotated based on the content of the tweets such that further hierarchical methods can be easily applied. Further, we prove the efficacy of standard text classification methods on this dataset, by designing experiments which do both binary as well as multi-class classification. Our experiments tackle the identification of either a specific topic (such as tobacco product promotion), a general mention (cigarettes and related products) or a more fine-grained classification. This methodology paves the way for further analysis, such as understanding sentiment or style, which makes this dataset a vital contribution to both disease surveillance and tobacco use research.
In this paper, we focus on the problem of adapting word vector-based models to new textual data. Given a model pre-trained on large reference data, how can we adapt it to a smaller piece of data with a slightly different language distribution? We frame the adaptation problem as a monolingual word vector alignment problem, and simply average models after alignment. We align vectors using the RCSLS criterion. Our formulation results in a simple and efficient algorithm that allows adapting general-purpose models to changing word distributions. In our evaluation, we consider applications to word embedding and text classification models. We show that the proposed approach yields good performance in all setups and outperforms a baseline consisting in fine-tuning the model on new data.
We consider a document classification problem where document labels are absent but only relevant keywords of a target class and unlabeled documents are given. Although heuristic methods based on pseudo-labeling have been considered, theoretical understanding of this problem has still been limited. Moreover, previous methods cannot easily incorporate well-developed techniques in supervised text classification. In this paper, we propose a theoretically guaranteed learning framework that is simple to implement and has flexible choices of models, e.g., linear models or neural networks. We demonstrate how to optimize the area under the receiver operating characteristic curve (AUC) effectively and also discuss how to adjust it to optimize other well-known evaluation metrics such as the accuracy and F1-measure. Finally, we show the effectiveness of our framework using benchmark datasets.
Given a state-of-the-art deep neural network text classifier, we show the existence of a universal and very small perturbation vector (in the embedding space) that causes natural text to be misclassified with high probability. Unlike images on which a single fixed-size adversarial perturbation can be found, text is of variable length, so we define the "universality" as "token-agnostic", where a single perturbation is applied to each token, resulting in different perturbations of flexible sizes at the sequence level. We propose an algorithm to compute universal adversarial perturbations, and show that the state-of-the-art deep neural networks are highly vulnerable to them, even though they keep the neighborhood of tokens mostly preserved. We also show how to use these adversarial perturbations to generate adversarial text samples. The surprising existence of universal "token-agnostic" adversarial perturbations may reveal important properties of a text classifier.
This study is to review the approaches used for measuring sentences similarity. Measuring similarity between natural language sentences is a crucial task for many Natural Language Processing applications such as text classification, information retrieval, question answering, and plagiarism detection. This survey classifies approaches of calculating sentences similarity based on the adopted methodology into three categories. Word-to-word based, structure based, and vector-based are the most widely used approaches to find sentences similarity. Each approach measures relatedness between short texts based on a specific perspective. In addition, datasets that are mostly used as benchmarks for evaluating techniques in this field are introduced to provide a complete view on this issue. The approaches that combine more than one perspective give better results. Moreover, structure based similarity that measures similarity between sentences structures needs more investigation.
Recently, researches have explored the graph neural network (GNN) techniques on text classification, since GNN does well in handling complex structures and preserving global information. However, previous methods based on GNN are mainly faced with the practical problems of fixed corpus level graph structure which do not support online testing and high memory consumption. To tackle the problems, we propose a new GNN based model that builds graphs for each input text with global parameters sharing instead of a single graph for the whole corpus. This method removes the burden of dependence between an individual text and entire corpus which support online testing, but still preserve global information. Besides, we build graphs by much smaller windows in the text, which not only extract more local features but also significantly reduce the edge numbers as well as memory consumption. Experiments show that our model outperforms existing models on several text classification datasets even with consuming less memory.
Recently machine learning is being applied to almost every data domain one of which is Question Answering Systems (QAS). A typical Question Answering System is fairly an information retrieval system, which matches documents or text and retrieve the most accurate one. The idea of open domain question answering system put forth, involves convolutional neural network text classifiers. The Classification model presented in this paper is multi-class text classifier. The neural network classifier can be trained on large dataset. We report series of experiments conducted on Convolution Neural Network (CNN) by training it on two different datasets. Neural network model is trained on top of word embedding. Softmax layer is applied to calculate loss and mapping of semantically related words. Gathered results can help justify the fact that proposed hypothetical QAS is feasible. We further propose a method to integrate Convolutional Neural Network Classifier to an open domain question answering system. The idea of Open domain will be further explained, but the generality of it indicates to the system of domain specific trainable models, thus making it an open domain.
Generative classifiers offer potential advantages over their discriminative counterparts, namely in the areas of data efficiency, robustness to data shift and adversarial examples, and zero-shot learning (Ng and Jordan,2002; Yogatama et al., 2017; Lewis and Fan,2019). In this paper, we improve generative text classifiers by introducing discrete latent variables into the generative story, and explore several graphical model configurations. We parameterize the distributions using standard neural architectures used in conditional language modeling and perform learning by directly maximizing the log marginal likelihood via gradient-based optimization, which avoids the need to do expectation-maximization. We empirically characterize the performance of our models on six text classification datasets. The choice of where to include the latent variable has a significant impact on performance, with the strongest results obtained when using the latent variable as an auxiliary conditioning variable in the generation of the textual input. This model consistently outperforms both the generative and discriminative classifiers in small-data settings. We analyze our model by using it for controlled generation, finding that the latent variable captures interpretable properties of the data, even with very small training sets.
Understanding the vulnerability of linguistic features extracted from noisy text is important for both developing better health text classification models and for interpreting vulnerabilities of natural language models. In this paper, we investigate how generic language characteristics, such as syntax or the lexicon, are impacted by artificial text alterations. The vulnerability of features is analysed from two perspectives: (1) the level of feature value change, and (2) the level of change of feature predictive power as a result of text modifications. We show that lexical features are more sensitive to text modifications than syntactic ones. However, we also demonstrate that these smaller changes of syntactic features have a stronger influence on classification performance downstream, compared to the impact of changes to lexical features. Results are validated across three datasets representing different text-classification tasks, with different levels of lexical and syntactic complexity of both conversational and written language.
Most of the information is stored as text, so text mining is regarded as having high commercial potential. Aiming at the semantic constraint problem of classification methods based on sparse representation, we propose a weighted recurrent neural network (W-RNN), which can fully extract text serialization semantic information. For the problem that the feature high dimensionality and unclear semantic relationship in text data representation, we first utilize the word vector to represent the vocabulary in the text and use Recurrent Neural Network (RNN) to extract features of the serialized text data. The word vector is then automatically weighted and summed using the intermediate output of the word vector to form the text representation vector. Finally, the neural network is used for classification. W-RNN is verified on the news dataset and proves that W-RNN is superior to other four baseline methods in Precision, Recall, F1 and loss values, which is suitable for text classification.
The International Classification of Diseases (ICD) is a list of classification codes for the diagnoses. Automatic ICD coding is in high demand as the manual coding can be labor-intensive and error-prone. It is a multi-label text classification task with extremely long-tailed label distribution, making it difficult to perform fine-grained classification on both frequent and zero-shot codes at the same time. In this paper, we propose a latent feature generation framework for generalized zero-shot ICD coding, where we aim to improve the prediction on codes that have no labeled data without compromising the performance on seen codes. Our framework generates pseudo features conditioned on the ICD code descriptions and exploits the ICD code hierarchical structure. To guarantee the semantic consistency between the generated features and real features, we reconstruct the keywords in the input documents that are related to the conditioned ICD codes. To the best of our knowledge, this works represents the first one that proposes an adversarial generative model for the generalized zero-shot learning on multi-label text classification. Extensive experiments demonstrate the effectiveness of our approach. On the public MIMIC-III dataset, our methods improve the F1 score from nearly 0 to 20.91% for the zero-shot codes, and increase the AUC score by 3% (absolute improvement) from previous state of the art. We also show that the framework improves the performance on few-shot codes.
Text representation can aid machines in understanding text. Previous work on text representation often focuses on the so-called forward implication, i.e., preceding words are taken as the context of later words for creating representations, thus ignoring the fact that the semantics of a text segment is a product of the mutual implication of words in the text: later words contribute to the meaning of preceding words. We introduce the concept of interaction and propose a two-perspective interaction representation, that encapsulates a local and a global interaction representation. Here, a local interaction representation is one that interacts among words with parent-children relationships on the syntactic trees and a global interaction interpretation is one that interacts among all the words in a sentence. We combine the two interaction representations to develop a Hybrid Interaction Representation (HIR). Inspired by existing feature-based and fine-tuning-based pretrain-finetuning approaches to language models, we integrate the advantages of feature-based and fine-tuning-based methods to propose the Pre-train, Interact, Fine-tune (PIF) architecture. We evaluate our proposed models on five widely-used datasets for text classification tasks. Our ensemble method, outperforms state-of-the-art baselines with improvements ranging from 2.03% to 3.15% in terms of error rate. In addition, we find that, the improvements of PIF against most state-of-the-art methods is not affected by increasing of the length of the text.
Text classification is one of the most critical areas in machine learning and artificial intelligence research. It has been actively adopted in many business applications such as conversational intelligence systems, news articles categorizations, sentiment analysis, emotion detection systems, and many other recommendation systems in our daily life. One of the problems in supervised text classification models is that the models' performance depends heavily on the quality of data labeling that is typically done by humans. In this study, we propose a new network community detection-based approach to automatically label and classify text data into multiclass value spaces. Specifically, we build networks with sentences as the network nodes and pairwise cosine similarities between the Term Frequency-Inversed Document Frequency (TFIDF) vector representations of the sentences as the network link weights. We use the Louvain method to detect the communities in the sentence networks. We train and test the Support Vector Machine and the Random Forest models on both the human-labeled data and network community detection labeled data. Results showed that models with the data labeled by the network community detection outperformed the models with the human-labeled data by 2.68-3.75% of classification accuracy. Our method may help developments of more accurate conversational intelligence and other text classification systems.
The exploding cost and time needed for data labeling and model training are bottlenecks for training DNN models on large datasets. Identifying smaller representative data samples with strategies like active learning can help mitigate such bottlenecks. Previous works on active learning in NLP identify the problem of sampling bias in the samples acquired by uncertainty-based querying and develop costly approaches to address it. Using a large empirical study, we demonstrate that active set selection using the posterior entropy of deep models like FastText.zip (FTZ) is robust to sampling biases and to various algorithmic choices (query size and strategies) unlike that suggested by traditional literature. We also show that FTZ based query strategy produces sample sets similar to those from more sophisticated approaches (e.g ensemble networks). Finally, we show the effectiveness of the selected samples by creating tiny high-quality datasets, and utilizing them for fast and cheap training of large models. Based on the above, we propose a simple baseline for deep active text classification that outperforms the state-of-the-art. We expect the presented work to be useful and informative for dataset compression and for problems involving active, semi-supervised or online learning scenarios. Code and models are available at: https://github.com/drimpossible/Sampling-Bias-Active-Learning
Over the last few years, deep learning has proven to be a great solution to many problems, such as image or text classification. Recently, deep learning-based solutions have outperformed humans on selected benchmark datasets, yielding a promising future for scientific and real-world applications. Training of deep learning models requires vast amounts of high quality data to achieve such supreme performance. In real-world scenarios, obtaining a large, coherent, and properly labeled dataset is a challenging task. This is especially true in medical applications, where high-quality data and annotations are scarce and the number of expert annotators is limited. In this paper, we investigate the impact of corrupted ground-truth masks on the performance of a neural network for a brain tumor segmentation task. Our findings suggest that a) the performance degrades about 8% less than it could be expected from simulations, b) a neural network learns the simulated biases of annotators, c) biases can be partially mitigated by using an inversely-biased dice loss function.
The complicated syntax structure of natural language is hard to be explicitly modeled by sequence-based models. Graph is a natural structure to describe the complicated relation between tokens. The recent advance in Graph Neural Networks (GNN) provides a powerful tool to model graph structure data, but simple graph models such as Graph Convolutional Networks (GCN) suffer from over-smoothing problem, that is, when stacking multiple layers, all nodes will converge to the same value. In this paper, we propose a novel Recursive Graphical Neural Networks model (ReGNN) to represent text organized in the form of graph. In our proposed model, LSTM is used to dynamically decide which part of the aggregated neighbor information should be transmitted to upper layers thus alleviating the over-smoothing problem. Furthermore, to encourage the exchange between the local and global information, a global graph-level node is designed. We conduct experiments on both single and multiple label text classification tasks. Experiment results show that our ReGNN model surpasses the strong baselines significantly in most of the datasets and greatly alleviates the over-smoothing problem.
In this paper we propose a novel dual adversarial co-learning approach for multi-domain text classification (MDTC). The approach learns shared-private networks for feature extraction and deploys dual adversarial regularizations to align features across different domains and between labeled and unlabeled data simultaneously under a discrepancy based co-learning framework, aiming to improve the classifiers' generalization capacity with the learned features. We conduct experiments on multi-domain sentiment classification datasets. The results show the proposed approach achieves the state-of-the-art MDTC performance.
Short-text classification, like all data science, struggles to achieve high performance using limited data. As a solution, a short sentence may be expanded with new and relevant feature words to form an artificially enlarged dataset, and add new features to testing data. This paper applies a novel approach to text expansion by generating new words directly for each input sentence, thus requiring no additional datasets or previous training. In this unsupervised approach, new keywords are formed within the hidden states of a pre-trained language model and then used to create extended pseudo documents. The word generation process was assessed by examining how well the predicted words matched to topics of the input sentence. It was found that this method could produce 3-10 relevant new words for each target topic, while generating just 1 word related to each non-target topic. Generated words were then added to short news headlines to create extended pseudo headlines. Experimental results have shown that models trained using the pseudo headlines can improve classification accuracy when limiting the number of training examples.