Active learning promises to alleviate the massive data needs of supervised machine learning: it has successfully improved sample efficiency by an order of magnitude on traditional tasks like topic classification and object recognition. However, we uncover a striking contrast to this promise: across 5 models and 4 datasets on the task of visual question answering, a wide variety of active learning approaches fail to outperform random selection. To understand this discrepancy, we profile 8 active learning methods on a per-example basis, and identify the problem as collective outliers -- groups of examples that active learning methods prefer to acquire but models fail to learn (e.g., questions that ask about text in images or require external knowledge). Through systematic ablation experiments and qualitative visualizations, we verify that collective outliers are a general phenomenon responsible for degrading pool-based active learning. Notably, we show that active learning sample efficiency increases significantly as the number of collective outliers in the active learning pool decreases. We conclude with a discussion and prescriptive recommendations for mitigating the effects of these outliers in future work.
Topic classification systems on spoken documents usually consist of two modules: an automatic speech recognition (ASR) module to convert speech into text and a text topic classification (TTC) module to predict the topic class from the decoded text. In this paper, instead of using the ASR transcripts, the fusion of deep acoustic and linguistic features is used for topic classification on spoken documents. More specifically, a conventional CTC-based acoustic model (AM) using phonemes as output units is first trained, and the outputs of the layer before the linear phoneme classifier in the trained AM are used as the deep acoustic features of spoken documents. Furthermore, these deep acoustic features are fed to a phoneme-to-word (P2W) module to obtain deep linguistic features. Finally, a local multi-head attention module is proposed to fuse these two types of deep features for topic classification. Experiments conducted on a subset selected from Switchboard corpus show that our proposed framework outperforms the conventional ASR+TTC systems and achieves a 3.13% improvement in ACC.
We introduce Korean Language Understanding Evaluation (KLUE) benchmark. KLUE is a collection of 8 Korean natural language understanding (NLU) tasks, including Topic Classification, SemanticTextual Similarity, Natural Language Inference, Named Entity Recognition, Relation Extraction, Dependency Parsing, Machine Reading Comprehension, and Dialogue State Tracking. We build all of the tasks from scratch from diverse source corpora while respecting copyrights, to ensure accessibility for anyone without any restrictions. With ethical considerations in mind, we carefully design annotation protocols. Along with the benchmark tasks and data, we provide suitable evaluation metrics and fine-tuning recipes for pretrained language models for each task. We furthermore release the pretrained language models (PLM), KLUE-BERT and KLUE-RoBERTa, to help reproducing baseline models on KLUE and thereby facilitate future research. We make a few interesting observations from the preliminary experiments using the proposed KLUE benchmark suite, already demonstrating the usefulness of this new benchmark suite. First, we find KLUE-RoBERTa-large outperforms other baselines, including multilingual PLMs and existing open-source Korean PLMs. Second, we see minimal degradation in performance even when we replace personally identifiable information from the pretraining corpus, suggesting that privacy and NLU capability are not at odds with each other. Lastly, we find that using BPE tokenization in combination with morpheme-level pre-tokenization is effective in tasks involving morpheme-level tagging, detection and generation. In addition to accelerating Korean NLP research, our comprehensive documentation on creating KLUE will facilitate creating similar resources for other languages in the future. KLUE is available at https://klue-benchmark.com.
We introduce DocSCAN, a completely unsupervised text classification approach using Semantic Clustering by Adopting Nearest-Neighbors (SCAN). For each document, we obtain semantically informative vectors from a large pre-trained language model. Similar documents have proximate vectors, so neighbors in the representation space tend to share topic labels. Our learnable clustering approach uses pairs of neighboring datapoints as a weak learning signal. The proposed approach learns to assign classes to the whole dataset without provided ground-truth labels. On five topic classification benchmarks, we improve on various unsupervised baselines by a large margin. In datasets with relatively few and balanced outcome classes, DocSCAN approaches the performance of supervised classification. The method fails for other types of classification, such as sentiment analysis, pointing to important conceptual and practical differences between classifying images and texts.
Large-scale pre-trained models like BERT, have obtained a great success in various Natural Language Processing (NLP) tasks, while it is still a challenge to adapt them to the math-related tasks. Current pre-trained models neglect the structural features and the semantic correspondence between formula and its context. To address these issues, we propose a novel pre-trained model, namely \textbf{MathBERT}, which is jointly trained with mathematical formulas and their corresponding contexts. In addition, in order to further capture the semantic-level structural features of formulas, a new pre-training task is designed to predict the masked formula substructures extracted from the Operator Tree (OPT), which is the semantic structural representation of formulas. We conduct various experiments on three downstream tasks to evaluate the performance of MathBERT, including mathematical information retrieval, formula topic classification and formula headline generation. Experimental results demonstrate that MathBERT significantly outperforms existing methods on all those three tasks. Moreover, we qualitatively show that this pre-trained model effectively captures the semantic-level structural information of formulas. To the best of our knowledge, MathBERT is the first pre-trained model for mathematical formula understanding.
Large pre-trained language models (LMs) have demonstrated remarkable ability as few-shot learners. However, their success hinges largely on scaling model parameters to a degree that makes it challenging to train and serve. In this paper, we propose a new approach, named as EFL, that can turn small LMs into better few-shot learners. The key idea of this approach is to reformulate potential NLP task into an entailment one, and then fine-tune the model with as little as 8 examples. We further demonstrate our proposed method can be: (i) naturally combined with an unsupervised contrastive learning-based data augmentation method; (ii) easily extended to multilingual few-shot learning. A systematic evaluation on 18 standard NLP tasks demonstrates that this approach improves the various existing SOTA few-shot learning methods by 12\%, and yields competitive few-shot performance with 500 times larger models, such as GPT-3.
Sentence embedding models aim to provide general purpose embeddings for sentences. Most of the models studied in this paper claim to perform well on STS tasks - but they do not report on their suitability for clustering. This paper looks at four recent sentence embedding models (Universal Sentence Encoder (Cer et al., 2018), Sentence-BERT (Reimers and Gurevych, 2019), LASER (Artetxe and Schwenk, 2019), and DeCLUTR (Giorgi et al., 2020)). It gives a brief overview of the ideas behind their implementations. It then investigates how well topic classes in two text classification datasets (Amazon Reviews (Ni et al., 2019) and News Category Dataset (Misra, 2018)) map to clusters in their corresponding sentence embedding space. While the performance of the resulting classification model is far from perfect, it is better than random. This is interesting because the classification model has been constructed in an unsupervised way. The topic classes in these real life topic classification datasets can be partly reconstructed by clustering the corresponding sentence embeddings.
Neural topic models can augment or replace bag-of-words inputs with the learned representations of deep pre-trained transformer-based word prediction models. One added benefit when using representations from multilingual models is that they facilitate zero-shot polylingual topic modeling. However, while it has been widely observed that pre-trained embeddings should be fine-tuned to a given task, it is not immediately clear what supervision should look like for an unsupervised task such as topic modeling. Thus, we propose several methods for fine-tuning encoders to improve both monolingual and zero-shot polylingual neural topic modeling. We consider fine-tuning on auxiliary tasks, constructing a new topic classification task, integrating the topic classification objective directly into topic model training, and continued pre-training. We find that fine-tuning encoder representations on topic classification and integrating the topic classification task directly into topic modeling improves topic quality, and that fine-tuning encoder representations on any task is the most important factor for facilitating cross-lingual transfer.
Text classification is the most basic natural language processing task. It has a wide range of applications ranging from sentiment analysis to topic classification. Recently, deep learning approaches based on CNN, LSTM, and Transformers have been the de facto approach for text classification. In this work, we highlight a common issue associated with these approaches. We show that these systems are over-reliant on the important words present in the text that are useful for classification. With limited training data and discriminative training strategy, these approaches tend to ignore the semantic meaning of the sentence and rather just focus on keywords or important n-grams. We propose a simple black box technique ShutText to present the shortcomings of the model and identify the over-reliance of the model on keywords. This involves randomly shuffling the words in a sentence and evaluating the classification accuracy. We see that on common text classification datasets there is very little effect of shuffling and with high probability these models predict the original class. We also evaluate the effect of language model pretraining on these models and try to answer questions around model robustness to out of domain sentences. We show that simple models based on CNN or LSTM as well as complex models like BERT are questionable in terms of their syntactic and semantic understanding.
Humans have been shown to give contrastive explanations, which explain why an observed event happened rather than some other counterfactual event (the contrast case). Despite the influential role that contrastivity plays in how humans explain, this property is largely missing from current methods for explaining NLP models. We present Minimal Contrastive Editing (MiCE), a method for producing contrastive explanations of model predictions in the form of edits to inputs that change model outputs to the contrast case. Our experiments across three tasks--binary sentiment classification, topic classification, and multiple-choice question answering--show that MiCE is able to produce edits that are not only contrastive, but also minimal and fluent, consistent with human contrastive edits. We demonstrate how MiCE edits can be used for two use cases in NLP system development--debugging incorrect model outputs and uncovering dataset artifacts--and thereby illustrate that producing contrastive explanations is a promising research direction for model interpretability.
We tackle the challenge of topic classification of tweets in the context of analyzing a large collection of curated streams by news outlets and other organizations to deliver relevant content to users. Our approach is novel in applying distant supervision based on semi-automatically identifying curated streams that are topically focused (for example, on politics, entertainment, or sports). These streams provide a source of labeled data to train topic classifiers that can then be applied to categorize tweets from more topically-diffuse streams. Experiments on both noisy labels and human ground-truth judgments demonstrate that our approach yields good topic classifiers essentially "for free", and that topic classifiers trained in this manner are able to dynamically adjust for topic drift as news on Twitter evolves.
Language exhibits structure at different scales, ranging from subwords to words, sentences, paragraphs, and documents. To what extent do deep models capture information at these scales, and can we force them to better capture structure across this hierarchy? We approach this question by focusing on individual neurons, analyzing the behavior of their activations at different timescales. We show that signal processing provides a natural framework for separating structure across scales, enabling us to 1) disentangle scale-specific information in existing embeddings and 2) train models to learn more about particular scales. Concretely, we apply spectral filters to the activations of a neuron across an input, producing filtered embeddings that perform well on part of speech tagging (word-level), dialog speech acts classification (utterance-level), or topic classification (document-level), while performing poorly on the other tasks. We also present a prism layer for training models, which uses spectral filters to constrain different neurons to model structure at different scales. Our proposed BERT + Prism model can better predict masked tokens using long-range context and produces multiscale representations that perform better at utterance- and document-level tasks. Our methods are general and readily applicable to other domains besides language, such as images, audio, and video.
Even though Variational Autoencoders (VAEs) are widely used for semi-supervised learning, the reason why they work remains unclear. In fact, the addition of the unsupervised objective is most often vaguely described as a regularization. The strength of this regularization is controlled by down-weighting the objective on the unlabeled part of the training set. Through an analysis of the objective of semi-supervised VAEs, we observe that they use the posterior of the learned generative model to guide the inference model in learning the partially observed latent variable. We show that given this observation, it is possible to gain finer control on the effect of the unsupervised objective on the training procedure. Using importance weighting, we derive two novel objectives that prioritize either one of the partially observed latent variable, or the unobserved latent variable. Experiments on the IMDB english sentiment analysis dataset and on the AG News topic classification dataset show the improvements brought by our prioritization mechanism and exhibit a behavior that is inline with our description of the inner working of Semi-Supervised VAEs.
Multilingual transformer models like mBERT and XLM-RoBERTa have obtained great improvements for many NLP tasks on a variety of languages. However, recent works also showed that results from high-resource languages could not be easily transferred to realistic, low-resource scenarios. In this work, we study trends in performance for different amounts of available resources for the three African languages Hausa, isiXhosa and Yor\`ub\'a on both NER and topic classification. We show that in combination with transfer learning or distant supervision, these models can achieve with as little as 10 or 100 labeled sentences the same performance as baselines with much more supervised training data. However, we also find settings where this does not hold. Our discussions and additional experiments on assumptions such as time and hardware restrictions highlight challenges and opportunities in low-resource learning.
In this paper, we study bidirectional LSTM network for the task of text classification using both supervised and semi-supervised approaches. Several prior works have suggested that either complex pretraining schemes using unsupervised methods such as language modeling (Dai and Le 2015; Miyato, Dai, and Goodfellow 2016) or complicated models (Johnson and Zhang 2017) are necessary to achieve a high classification accuracy. However, we develop a training strategy that allows even a simple BiLSTM model, when trained with cross-entropy loss, to achieve competitive results compared with more complex approaches. Furthermore, in addition to cross-entropy loss, by using a combination of entropy minimization, adversarial, and virtual adversarial losses for both labeled and unlabeled data, we report state-of-the-art results for text classification task on several benchmark datasets. In particular, on the ACL-IMDB sentiment analysis and AG-News topic classification datasets, our method outperforms current approaches by a substantial margin. We also show the generality of the mixed objective function by improving the performance on relation extraction task.
The authors compared oversampling methods for the problem of multi-class topic classification. The SMOTE algorithm underlies one of the most popular oversampling methods. It consists in choosing two examples of a minority class and generating a new example based on them. In the paper, the authors compared the basic SMOTE method with its two modifications (Borderline SMOTE and ADASYN) and random oversampling technique on the example of one of text classification tasks. The paper discusses the k-nearest neighbor algorithm, the support vector machine algorithm and three types of neural networks (feedforward network, long short-term memory (LSTM) and bidirectional LSTM). The authors combine these machine learning algorithms with different text representations and compared synthetic oversampling methods. In most cases, the use of oversampling techniques can significantly improve the quality of classification. The authors conclude that for this task, the quality of the KNN and SVM algorithms is more influenced by class imbalance than neural networks.
Emotion recognition has become a major problem in computer vision in recent years that made a lot of effort by researchers to overcome the difficulties in this task. In the field of affective computing, emotion recognition has a wide range of applications, such as healthcare, robotics, human-computer interaction. Due to its practical importance for other tasks, many techniques and approaches have been investigated for different problems and various data sources. Nevertheless, comprehensive fusion of the audio-visual and language modalities to get the benefits from them is still a problem to solve. In this paper, we present and discuss our classification methodology for MuSe-Topic Sub-challenge, as well as the data and results. For the topic classification, we ensemble two language models which are ALBERT and RoBERTa to predict 10 classes of topics. Moreover, for the classification of valence and arousal, SVM and Random forests are employed in conjunction with feature selection to enhance the performance.
Coronavirus disease (COVID-19) has been declared as a pandemic by WHO with thousands of cases being reported each day. Numerous scientific articles are being published on the disease raising the need for a service which can organize, and query them in a reliable fashion. To support this cause we present AWS CORD-19 Search (ACS), a public, COVID-19 specific, neural search engine that is powered by several machine learning systems to support natural language based searches. ACS with capabilities such as document ranking, passage ranking, question answering and topic classification provides a scalable solution to COVID-19 researchers and policy makers in their search and discovery for answers to high priority scientific questions. We present a quantitative evaluation and qualitative analysis of the system against other leading COVID-19 search platforms. ACS is top performing across these systems yielding quality results which we detail with relevant examples in this work.
Identifying the topic (domain) of each user's utterance in open-domain conversational systems is a crucial step for all subsequent language understanding and response tasks. In particular, for complex domains, an utterance is often routed to a single component responsible for that domain. Thus, correctly mapping a user utterance to the right domain is critical. To address this problem, we introduce ConCET: a Concurrent Entity-aware conversational Topic classifier, which incorporates entity-type information together with the utterance content features. Specifically, ConCET utilizes entity information to enrich the utterance representation, combining character, word, and entity-type embeddings into a single representation. However, for rich domains with millions of available entities, unrealistic amounts of labeled training data would be required. To complement our model, we propose a simple and effective method for generating synthetic training data, to augment the typically limited amounts of labeled training data, using commonly available knowledge bases to generate additional labeled utterances. We extensively evaluate ConCET and our proposed training method first on an openly available human-human conversational dataset called Self-Dialogue, to calibrate our approach against previous state-of-the-art methods; second, we evaluate ConCET on a large dataset of human-machine conversations with real users, collected as part of the Amazon Alexa Prize. Our results show that ConCET significantly improves topic classification performance on both datasets, including 8-10% improvements over state-of-the-art deep learning methods. We complement our quantitative results with detailed analysis of system performance, which could be used for further improvements of conversational agents.
One of the keys to enable chatbots to communicate with human in a more natural way is the ability to handle long and complex user's utterances. In order to achieve this goal, we propose to integrate the Sentence Boundary Detection (SBD) module into the chatbot architecture, whose role is to take as input a user's utterance from an automatic speech recognition device, in which sentence boundaries are not available, and output the corresponding list of punctuated sentences for downstream modules such as Intent Detection, Topic Classification, Sentiment Analysis, Named Entity Recognition, as well as Coreference Recognition. To address the SBD task, we reformulate it as a sequence labeling task. In this way, both deep neural network models (e.g., Bi-directional Long Short-Term Memory, Convolutional Neural Network) and structured prediction models (e.g., Hidden Markov Model, Maximum Entropy Model, Conditional Random Field) can be leveraged. After reformulating the SBD task, we built a hybrid deep neural network model and achieved good performance on both CornellMovie-Dialog and DailyDialog datasets.