It has been long known that sparsity is an effective inductive bias for learning efficient representation of data in vectors with fixed dimensionality, and it has been explored in many areas of representation learning. Of particular interest to this work is the investigation of the sparsity within the VAE framework which has been explored a lot in the image domain, but has been lacking even a basic level of exploration in NLP. Additionally, NLP is also lagging behind in terms of learning sparse representations of large units of text e.g., sentences. We use the VAEs that induce sparse latent representations of large units of text to address the aforementioned shortcomings. First, we move in this direction by measuring the success of unsupervised state-of-the-art (SOTA) and other strong VAE-based sparsification baselines for text and propose a hierarchical sparse VAE model to address the stability issue of SOTA. Then, we look at the implications of sparsity on text classification across 3 datasets, and highlight a link between performance of sparse latent representations on downstream tasks and its ability to encode task-related information.
Text classification is a popular topic of natural language processing, which has currently attracted numerous research efforts worldwide. The significant increase of data in social media requires the vast attention of researchers to analyze such data. There are various studies in this field in many languages but limited to the Vietnamese language. Therefore, this study aims to classify Vietnamese texts on social media from three different Vietnamese benchmark datasets. Advanced deep learning models are used and optimized in this study, including CNN, LSTM, and their variants. We also implement the BERT, which has never been applied to the datasets. Our experiments find a suitable model for classification tasks on each specific dataset. To take advantage of single models, we propose an ensemble model, combining the highest-performance models. Our single models reach positive results on each dataset. Moreover, our ensemble model achieves the best performance on all three datasets. We reach 86.96% of F1- score for the HSD-VLSP dataset, 65.79% of F1-score for the UIT-VSMEC dataset, 92.79% and 89.70% for sentiments and topics on the UIT-VSFC dataset, respectively. Therefore, our models achieve better performances as compared to previous studies on these datasets.
We present iNLTK, an open-source NLP library consisting of pre-trained language models and out-of-the-box support for Data Augmentation, Textual Similarity, Sentence Embeddings, Word Embeddings, Tokenization and Text Generation in 13 Indic Languages. By using pre-trained models from iNLTK for text classification on publicly available datasets, we significantly outperform previously reported results. On these datasets, we also show that by using pre-trained models and data augmentation from iNLTK, we can achieve more than 95% of the previous best performance by using less than 10% of the training data. iNLTK is already being widely used by the community and has 40,000+ downloads, 600+ stars and 100+ forks on GitHub. The library is available at https://github.com/goru001/inltk.
In the text classification problem, the imbalance of labels in datasets affect the performance of the text-classification models. Practically, the data about user comments on social networking sites not altogether appeared - the administrators often only allow positive comments and hide negative comments. Thus, when collecting the data about user comments on the social network, the data is usually skewed about one label, which leads the dataset to become imbalanced and deteriorate the model's ability. The data augmentation techniques are applied to solve the imbalance problem between classes of the dataset, increasing the prediction model's accuracy. In this paper, we performed augmentation techniques on the VLSP2019 Hate Speech Detection on Vietnamese social texts and the UIT - VSFC: Vietnamese Students' Feedback Corpus for Sentiment Analysis. The result of augmentation increases by about 1.5% in the F1-macro score on both corpora.
The ability to automatically determine the age audience of a novel provides many opportunities for the development of information retrieval tools. Firstly, developers of book recommendation systems and electronic libraries may be interested in filtering texts by the age of the most likely readers. Further, parents may want to select literature for children. Finally, it will be useful for writers and publishers to determine which features influence whether the texts are suitable for children. In this article, we compare the empirical effectiveness of various types of linguistic features for the task of age-based classification of fiction texts. For this purpose, we collected a text corpus of book previews labeled with one of two categories -- children's or adult. We evaluated the following types of features: readability indices, sentiment, lexical, grammatical and general features, and publishing attributes. The results obtained show that the features describing the text at the document level can significantly increase the quality of machine learning models.
Text classification has become a key operation in various natural language processing tasks. The efficiency of most classification algorithms predominantly confide in the quality of input features. In this work, we propose a novel multi-class text classification technique that harvests features from two distinct feature extraction methods. Firstly, a structured heterogeneous text graph built based on document-word relations and word co-occurrences is leveraged using a Graph Convolution Network (GCN). Secondly, the documents are topic modeled to use the document-topic score as features into the classification model. The concerned graph is constructed using Point-Wise Mutual Information (PMI) between pair of word co-occurrences and Term Frequency-Inverse Document Frequency (TF-IDF) score for words in the documents for word co-occurrences. Experimentation reveals that our text classification model outperforms the existing techniques for five benchmark text classification data sets.
Adapting pre-trained language models (PrLMs) (e.g., BERT) to new domains has gained much attention recently. Instead of fine-tuning PrLMs as done in most previous work, we investigate how to adapt the features of PrLMs to new domains without fine-tuning. We explore unsupervised domain adaptation (UDA) in this paper. With the features from PrLMs, we adapt the models trained with labeled data from the source domain to the unlabeled target domain. Self-training is widely used for UDA which predicts pseudo labels on the target domain data for training. However, the predicted pseudo labels inevitably include noise, which will negatively affect training a robust model. To improve the robustness of self-training, in this paper we present class-aware feature self-distillation (CFd) to learn discriminative features from PrLMs, in which PrLM features are self-distilled into a feature adaptation module and the features from the same class are more tightly clustered. We further extend CFd to a cross-language setting, in which language discrepancy is studied. Experiments on two monolingual and multilingual Amazon review datasets show that CFd can consistently improve the performance of self-training in cross-domain and cross-language settings.
Hierarchical multi-label text classification (HMTC) has been gaining popularity in recent years thanks to its applicability to a plethora of real-world applications. The existing HMTC algorithms largely focus on the design of classifiers, such as the local, global, or a combination of them. However, very few studies have focused on hierarchical feature extraction and explore the association between the hierarchical labels and the text. In this paper, we propose a Label-based Attention for Hierarchical Mutlti-label Text Classification Neural Network (LA-HCN), where the novel label-based attention module is designed to hierarchically extract important information from the text based on the labels from different hierarchy levels. Besides, hierarchical information is shared across levels while preserving the hierarchical label-based information. Separate local and global document embeddings are obtained and used to facilitate the respective local and global classifications. In our experiments, LA-HCN outperforms other state-of-the-art neural network-based HMTC algorithms on four public HMTC datasets. The ablation study also demonstrates the effectiveness of the proposed label-based attention module as well as the novel local and global embeddings and classifications. By visualizing the learned attention (words), we find that LA-HCN is able to extract meaningful information corresponding to the different labels which provides explainability that may be helpful for the human analyst.
Sequence labelling tasks like Dialog Act and Emotion/Sentiment identification are a key component of spoken dialog systems. In this work, we propose a new approach to learn generic representations adapted to spoken dialog, which we evaluate on a new benchmark we call Sequence labellIng evaLuatIon benChmark fOr spoken laNguagE benchmark (\texttt{SILICONE}). \texttt{SILICONE} is model-agnostic and contains 10 different datasets of various sizes. We obtain our representations with a hierarchical encoder based on transformer architectures, for which we extend two well-known pre-training objectives. Pre-training is performed on OpenSubtitles: a large corpus of spoken dialog containing over $2.3$ billion of tokens. We demonstrate how hierarchical encoders achieve competitive results with consistently fewer parameters compared to state-of-the-art models and we show their importance for both pre-training and fine-tuning.
This paper studies the problem of detecting novel or unexpected instances in text classification. In traditional text classification, the classes appeared in testing must have been seen in training. However, in many applications, this is not the case because in testing, we may see unexpected instances that are not from any of the training classes. In this paper, we propose a significantly more effective approach that converts the original problem to a pair-wise matching problem and then outputs how probable two instances belong to the same class. Under this approach, we present two models. The more effective model uses two embedding matrices of a pair of instances as two channels of a CNN. The output probabilities from such pairs are used to judge whether a test instance is from a seen class or is novel/unexpected. Experimental results show that the proposed method substantially outperforms the state-of-the-art baselines.
Ensemble learning consistently improves the performance of multi-class classification through aggregating a series of base classifiers. To this end, data-independent ensemble methods like Error Correcting Output Codes (ECOC) attract increasing attention due to its easiness of implementation and parallelization. Specifically, traditional ECOCs and its general extension N-ary ECOC decompose the original multi-class classification problem into a series of independent simpler classification subproblems. Unfortunately, integrating ECOCs, especially N-ary ECOC with deep neural networks, termed as deep N-ary ECOC, is not straightforward and yet fully exploited in the literature, due to the high expense of training base learners. To facilitate the training of N-ary ECOC with deep learning base learners, we further propose three different variants of parameter sharing architectures for deep N-ary ECOC. To verify the generalization ability of deep N-ary ECOC, we conduct experiments by varying the backbone with different deep neural network architectures for both image and text classification tasks. Furthermore, extensive ablation studies on deep N-ary ECOC show its superior performance over other deep data-independent ensemble methods.
Adversarial attacking aims to fool deep neural networks with adversarial examples. In the field of natural language processing, various textual adversarial attack models have been proposed, varying in the accessibility to the victim model. Among them, the attack models that only require the output of the victim model are more fit for real-world situations of adversarial attacking. However, to achieve high attack performance, these models usually need to query the victim model too many times, which is neither efficient nor viable in practice. To tackle this problem, we propose a reinforcement learning based attack model, which can learn from attack history and launch attacks more efficiently. In experiments, we evaluate our model by attacking several state-of-the-art models on the benchmark datasets of multiple tasks including sentiment analysis, text classification and natural language inference. Experimental results demonstrate that our model consistently achieves both better attack performance and higher efficiency than recently proposed baseline methods. We also find our attack model can bring more robustness improvement to the victim model by adversarial training. All the code and data of this paper will be made public.
Decision forest algorithms typically model data by learning a binary tree structure recursively where every node splits the feature space into two sub-regions, sending examples into the left or right branch as a result. In axis-aligned decision forests, the "decision" to route an input example is the result of the evaluation of a condition on a single dimension in the feature space. Such conditions are learned using efficient, often greedy algorithms that optimize a local loss function. For example, a node's condition may be a threshold function applied to a numerical feature, and its parameter may be learned by sweeping over the set of values available at that node and choosing a threshold that maximizes some measure of purity. Crucially, whether an algorithm exists to learn and evaluate conditions for a feature type determines whether a decision forest algorithm can model that feature type at all. For example, decision forests today cannot consume textual features directly -- such features must be transformed to summary statistics instead. In this work, we set out to bridge that gap. We define a condition that is specific to categorical-set features -- defined as an unordered set of categorical variables -- and present an algorithm to learn it, thereby equipping decision forests with the ability to directly model text, albeit without preserving sequential order. Our algorithm is efficient during training and the resulting conditions are fast to evaluate with our extension of the QuickScorer inference algorithm. Experiments on benchmark text classification datasets demonstrate the utility and effectiveness of our proposal.
Improving the energy efficiency of mobile applications is a topic that has gained a lot of attention recently. It has been addressed in a number of ways such as identifying energy bugs and developing a catalog of energy patterns. Previous work shows that users discuss the battery-related issues (energy inefficiency or energy consumption) of the apps in their reviews. However, there is no work that addresses the automatic extraction of battery-related issues from users' feedback. In this paper, we report on a visualization tool that is developed to empirically study machine learning algorithms and text features to automatically identify the energy consumption specific reviews with the highest accuracy. Other than the common machine learning algorithms, we utilize deep learning models with different word embeddings to compare the results. Furthermore, to help the developers extract the main topics that are discussed in the reviews, two states of the art topic modeling algorithms are applied. The visualizations of the topics represent the keywords that are extracted for each topic along with a comparison with the results of string matching. The developed web-browser based interactive visualization tool is a novel framework developed with the intention of giving the app developers insights about running time and accuracy of machine learning and deep learning models as well as extracted topics. The tool makes it easier for the developers to traverse through the extensive result set generated by the text classification and topic modeling algorithms. The dynamic-data structure used for the tool stores the baseline-results of the discussed approaches and is updated when applied on new datasets. The tool is open-sourced to replicate the research results.
We derive a new margin-based regularization formulation, termed multi-margin regularization (MMR), for deep neural networks (DNNs). The MMR is inspired by principles that were applied in margin analysis of shallow linear classifiers, e.g., support vector machine (SVM). Unlike SVM, MMR is continuously scaled by the radius of the bounding sphere (i.e., the maximal norm of the feature vector in the data), which is constantly changing during training. We empirically demonstrate that by a simple supplement to the loss function, our method achieves better results on various classification tasks across domains. Using the same concept, we also derive a selective sampling scheme and demonstrate accelerated training of DNNs by selecting samples according to a minimal margin score (MMS). This score measures the minimal amount of displacement an input should undergo until its predicted classification is switched. We evaluate our proposed methods on three image classification tasks and six language text classification tasks. Specifically, we show improved empirical results on CIFAR10, CIFAR100 and ImageNet using state-of-the-art convolutional neural networks (CNNs) and BERT-BASE architecture for the MNLI, QQP, QNLI, MRPC, SST-2 and RTE benchmarks.
This paper presents our models for WNUT 2020 shared task2. The shared task2 involves identification of COVID-19 related informative tweets. We treat this as binary text classification problem and experiment with pre-trained language models. Our first model which is based on CT-BERT achieves F1-score of 88.7% and second model which is an ensemble of CT-BERT, RoBERTa and SVM achieves F1-score of 88.52%.
Compared to English, the amount of labeled data for Indonesian text classification tasks is very small. Recently developed multilingual language models have shown its ability to create multilingual representations effectively. This paper investigates the effect of combining English and Indonesian data on building Indonesian text classification (e.g., sentiment analysis and hate speech) using multilingual language models. Using the feature-based approach, we observe its performance on various data sizes and total added English data. The experiment showed that the addition of English data, especially if the amount of Indonesian data is small, improves performance. Using the fine-tuning approach, we further showed its effectiveness in utilizing the English language to build Indonesian text classification models.
Text classification has long been a staple within Natural Language Processing (NLP) with applications spanning across diverse areas such as sentiment analysis, recommender systems and spam detection. With such a powerful solution, it is often tempting to use it as the go-to tool for all NLP problems since when you are holding a hammer, everything looks like a nail. However, we argue here that many tasks which are currently addressed using classification are in fact being shoehorned into a classification mould and that if we instead address them as a ranking problem, we not only improve the model, but we achieve better performance. We propose a novel end-to-end ranking approach consisting of a Transformer network responsible for producing representations for a pair of text sequences, which are in turn passed into a context aggregating network outputting ranking scores used to determine an ordering to the sequences based on some notion of relevance. We perform numerous experiments on publicly-available datasets and investigate the applications of ranking in problems often solved using classification. In an experiment on a heavily-skewed sentiment analysis dataset, converting ranking results to classification labels yields an approximately 22% improvement over state-of-the-art text classification, demonstrating the efficacy of text ranking over text classification in certain scenarios.
Recent advancements in the NLP field showed that transfer learning helps with achieving state-of-the-art results for new tasks by tuning pre-trained models instead of starting from scratch. Transformers have made a significant improvement in creating new state-of-the-art results for many NLP tasks including but not limited to text classification, text generation, and sequence labeling. Most of these success stories were based on large datasets. In this paper we focus on a real-life scenario that scientists in academia and industry face frequently: given a small dataset, can we use a large pre-trained model like BERT and get better results than simple models? To answer this question, we use a small dataset for intent classification collected for building chatbots and compare the performance of a simple bidirectional LSTM model with a pre-trained BERT model. Our experimental results show that bidirectional LSTM models can achieve significantly higher results than a BERT model for a small dataset and these simple models get trained in much less time than tuning the pre-trained counterparts. We conclude that the performance of a model is dependent on the task and the data, and therefore before making a model choice, these factors should be taken into consideration instead of directly choosing the most popular model.
Contextual language models (CLMs) have pushed the NLP benchmarks to a new height. It has become a new norm to utilize CLM provided word embeddings in downstream tasks such as text classification. However, unless addressed, CLMs are prone to learn intrinsic gender-bias in the dataset. As a result, predictions of downstream NLP models can vary noticeably by varying gender words, such as replacing "he" to "she", or even gender-neutral words. In this paper, we focus our analysis on a popular CLM, i.e., BERT. We analyse the gender-bias it induces in five downstream tasks related to emotion and sentiment intensity prediction. For each task, we train a simple regressor utilizing BERT's word embeddings. We then evaluate the gender-bias in regressors using an equity evaluation corpus. Ideally and from the specific design, the models should discard gender informative features from the input. However, the results show a significant dependence of the system's predictions on gender-particular words and phrases. We claim that such biases can be reduced by removing genderspecific features from word embedding. Hence, for each layer in BERT, we identify directions that primarily encode gender information. The space formed by such directions is referred to as the gender subspace in the semantic space of word embeddings. We propose an algorithm that finds fine-grained gender directions, i.e., one primary direction for each BERT layer. This obviates the need of realizing gender subspace in multiple dimensions and prevents other crucial information from being omitted. Experiments show that removing embedding components in such directions achieves great success in reducing BERT-induced bias in the downstream tasks.