Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code.
The uncertainty measurement of classifiers' predictions is especially important in applications such as medical diagnoses that need to ensure limited human resources can focus on the most uncertain predictions returned by machine learning models. However, few existing uncertainty models attempt to improve overall prediction accuracy where human resources are involved in the text classification task. In this paper, we propose a novel neural-network-based model that applies a new dropout-entropy method for uncertainty measurement. We also design a metric learning method on feature representations, which can boost the performance of dropout-based uncertainty methods with smaller prediction variance in accurate prediction trials. Extensive experiments on real-world data sets demonstrate that our method can achieve a considerable improvement in overall prediction accuracy compared to existing approaches. In particular, our model improved the accuracy from 0.78 to 0.92 when 30\% of the most uncertain predictions were handed over to human experts in "20NewsGroup" data.
Despite the recent success of deep transfer learning approaches in NLP, there is a lack of quantitative studies demonstrating the gains these models offer in low-shot text classification tasks over existing paradigms. Deep transfer learning approaches such as BERT and ULMFiT demonstrate that they can beat state-of-the-art results on larger datasets, however when one has only 100-1000 labelled examples per class, the choice of approach is less clear, with classical machine learning and deep transfer learning representing valid options. This paper compares the current best transfer learning approach with top classical machine learning approaches on a trinary sentiment classification task to assess the best paradigm. We find that BERT, representing the best of deep transfer learning, is the best performing approach, outperforming top classical machine learning algorithms by 9.7% on average when trained with 100 examples per class, narrowing to 1.8% at 1000 labels per class. We also show the robustness of deep transfer learning in moving across domains, where the maximum loss in accuracy is only 0.7% in similar domain tasks and 3.2% cross domain, compared to classical machine learning which loses up to 20.6%.
The introduction of pre-trained language models has revolutionized natural language research communities. However, researchers still know relatively little regarding their theoretical and empirical properties. In this regard, Peters et al. perform several experiments which demonstrate that it is better to adapt BERT with a light-weight task-specific head, rather than building a complex one on top of the pre-trained language model, and freeze the parameters in the said language model. However, there is another option to adopt. In this paper, we propose a new adaptation method which we first train the task model with the BERT parameters frozen and then fine-tune the entire model together. Our experimental results show that our model adaptation method can achieve 4.7% accuracy improvement in semantic similarity task, 0.99% accuracy improvement in sequence labeling task and 0.72% accuracy improvement in the text classification task.
Neural network-based generative language models like ELMo and BERT can work effectively as general purpose sentence encoders in text classification without further fine-tuning. Is it possible to adapt them in a similar way for use as general-purpose decoders? For this to be possible, it would need to be the case that for any target sentence of interest, there is some continuous representation that can be passed to the language model to cause it to reproduce that sentence. We set aside the difficult problem of designing an encoder that can produce such representations and, instead, ask directly whether such representations exist at all. To do this, we introduce a pair of effective, complementary methods for feeding representations into pretrained unconditional language models and a corresponding set of methods to map sentences into and out of this representation space, the reparametrized sentence space. We then investigate the conditions under which a language model can be made to generate a sentence through the identification of a point in such a space and find that it is possible to recover arbitrary sentences nearly perfectly with language models and representations of moderate size without modifying any model parameters.
Automatic email categorization is an important application of text classification. We study the automatic reply of email business messages in Brazilian Portuguese. We present a novel corpus containing messages from a real application, and baseline categorization experiments using Naive Bayes and support Vector Machines. We then discuss the effect of lemmatization and the role of part-of-speech tagging filtering on precision and recall. Support Vector Machines classification coupled with nonlemmatized selection of verbs, nouns and adjectives was the best approach, with 87.3% maximum accuracy. Straightforward lemmatization in Portuguese led to the lowest classification results in the group, with 85.3% and 81.7% precision in SVM and Naive Bayes respectively. Thus, while lemmatization reduced precision and recall, part-of-speech filtering improved overall results.
We study the effect of different approaches to text augmentation. To do this we use 3 datasets that include social media and formal text in the form of news articles. Our goal is to provide insights for practitioners and researchers on making choices for augmentation for classification use cases. We observe that Word2vec-based augmentation is a viable option when one does not have access to a formal synonym model (like WordNet-based augmentation). The use of \emph{mixup} further improves performance of all text based augmentations and reduces the effects of overfitting on a tested deep learning model. Round-trip translation with a translation service proves to be harder to use due to cost and as such is less accessible for both normal and low resource use-cases.
Auto dealerships receive thousands of calls daily from customers who are interested in sales, service, vendors and jobseekers. With so many calls, it is very important for auto dealers to understand the intent of these calls to provide positive customer experiences that ensure customer satisfaction, deep customer engagement to boost sales and revenue, and optimum allocation of agents or customer service representatives across the business. In this paper, we define the problem of customer phone call intent as a multi-class classification problem stemming from the large database of recorded phone call transcripts. To solve this problem, we develop a convolutional neural network (CNN)-based supervised learning model to classify the customer calls into four intent categories: sales, service, vendor and jobseeker. Experimental results show that with the thrust of our scalable data labeling method to provide sufficient training data, the CNN-based predictive model performs very well on long text classification according to the quantitative metrics of F1-Score, precision, recall, and accuracy.
Intent Detection and Slot Filling are two pillar tasks in Spoken Natural Language Understanding. Common approaches adopt joint Deep Learning architectures in attention-based recurrent frameworks. In this work, we aim at exploiting the success of "recurrence-less" models for these tasks. We introduce Bert-Joint, i.e., a multi-lingual joint text classification and sequence labeling framework. The experimental evaluation over two well-known English benchmarks demonstrates the strong performances that can be obtained with this model, even when few annotated data is available. Moreover, we annotated a new dataset for the Italian language, and we observed similar performances without the need for changing the model.
While very deep neural networks have shown effectiveness for computer vision and text classification applications, how to increase the network depth of neural machine translation (NMT) models for better translation quality remains a challenging problem. Directly stacking more blocks to the NMT model results in no improvement and even reduces performance. In this work, we propose an effective two-stage approach with three specially designed components to construct deeper NMT models, which result in significant improvements over the strong Transformer baselines on WMT$14$ English$\to$German and English$\to$French translation tasks\footnote{Our code is available at \url{https://github.com/apeterswu/Depth_Growing_NMT}}.
A major obstacle to the development of Natural Language Processing (NLP) methods in the biomedical domain is data accessibility. This problem can be addressed by generating medical data artificially. Most previous studies have focused on the generation of short clinical text, and evaluation of the data utility has been limited. We propose a generic methodology to guide the generation of clinical text with key phrases. We use the artificial data as additional training data in two key biomedical NLP tasks: text classification and temporal relation extraction. We show that artificially generated training data used in conjunction with real training data can lead to performance boosts for data-greedy neural network algorithms. We also demonstrate the usefulness of the generated data for NLP setups where it fully replaces real training data.
In this work, we present graph star net (GraphStar), a novel and unified graph neural net architecture which utilizes message-passing relay and attention mechanism for multiple prediction tasks - node classification, graph classification and link prediction. GraphStar addresses many earlier challenges facing graph neural nets and achieves non-local representation without increasing the model depth or bearing heavy computational costs. We also propose a new method to tackle topic-specific sentiment analysis based on node classification and text classification as graph classification. Our work shows that 'star nodes' can learn effective graph-data representation and improve on current methods for the three tasks. Specifically, for graph classification and link prediction, GraphStar outperforms the current state-of-the-art models by 2-5% on several key benchmarks.
This paper deals with automatic systems for image recipe recognition. For this purpose, we compare and evaluate leading vision-based and text-based technologies on a new very large multimodal dataset (UPMC Food-101) containing about 100,000 recipes for a total of 101 food categories. Each item in this dataset is represented by one image plus textual information. We present deep experiments of recipe recognition on our dataset using visual, textual information and fusion. Additionally, we present experiments with text-based embedding technology to represent any food word in a semantical continuous space. We also compare our dataset features with a twin dataset provided by ETHZ university: we revisit their data collection protocols and carry out transfer learning schemes to highlight similarities and differences between both datasets. Finally, we propose a real application for daily users to identify recipes. This application is a web search engine that allows any mobile device to send a query image and retrieve the most relevant recipes in our dataset.
Automatic text generation has received much attention owing to rapid development of deep neural networks. In general, text generation systems based on statistical language model will not consider anthropomorphic characteristics, which results in machine-like generated texts. To fill the gap, we propose a conditional language generation model with Big Five Personality (BFP) feature vectors as input context, which writes human-like short texts. The short text generator consists of a layer of long short memory network (LSTM), where a BFP feature vector is concatenated as one part of input for each cell. To enable supervised training generation model, a text classification model based convolution neural network (CNN) has been used to prepare BFP-tagged Chinese micro-blog corpora. Validated by a BFP linguistic computational model, our generated Chinese short texts exhibit discriminative personality styles, which are also syntactically correct and semantically smooth with appropriate emoticons. With combination of natural language generation with psychological linguistics, our proposed BFP-dependent text generation model can be widely used for individualization in machine translation, image caption, dialogue generation and so on.
Text classification is a fundamental task for text data mining. In order to train a generalizable model, a large volume of text must be collected. To address data insufficiency, cross-lingual data may occasionally be necessary. Cross-lingual data sources may however suffer from data incompatibility, as text written in different languages can hold distinct word sequences and semantic patterns. Machine translation and word embedding alignment provide an effective way to transform and combine data for cross-lingual data training. To the best of our knowledge, there has been little work done on evaluating how the methodology used to conduct semantic space transformation and data combination affects the performance of classification models trained from cross-lingual resources. In this paper, we systematically evaluated the performance of two commonly used CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) text classifiers with differing data transformation and combination strategies. Monolingual models were trained from English and French alongside their translated and aligned embeddings. Our results suggested that semantic space transformation may conditionally promote the performance of monolingual models. Bilingual models were trained from a combination of both English and French. Our results indicate that a cross-lingual classification model can significantly benefit from cross-lingual data by learning from translated or aligned embedding spaces.
Recent progress in AutoML has lead to state-of-the-art methods (e.g., AutoSKLearn) that can be readily used by non-experts to approach any supervised learning problem. Whereas these methods are quite effective, they are still limited in the sense that they work for tabular (matrix formatted) data only. This paper describes one step forward in trying to automate the design of supervised learning methods in the context of text mining. We introduce a meta learning methodology for automatically obtaining a representation for text mining tasks starting from raw text. We report experiments considering 60 different textual representations and more than 80 text mining datasets associated to a wide variety of tasks. Experimental results show the proposed methodology is a promising solution to obtain highly effective off the shell text classification pipelines.
Feature attribution methods, proposed recently, help users interpret the predictions of complex models. Our approach integrates feature attributions into the objective function to allow machine learning practitioners to incorporate priors in model building. To demonstrate the effectiveness our technique, we apply it to two tasks: (1) mitigating unintended bias in text classifiers by neutralizing identity terms; (2) improving classifier performance in a scarce data setting by forcing the model to focus on toxic terms. Our approach adds an L2 distance loss between feature attributions and task-specific prior values to the objective. Our experiments show that i) a classifier trained with our technique reduces undesired model biases without a trade off on the original task; ii) incorporating priors helps model performance in scarce data settings.
With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, under comparable experiment settings, XLNet outperforms BERT on 20 tasks, often by a large margin, including question answering, natural language inference, sentiment analysis, and document ranking.
This paper addresses the task of readability assessment for the texts aimed at second language (L2) learners. One of the major challenges in this task is the lack of significantly sized level-annotated data. For the present work, we collected a dataset of CEFR-graded texts tailored for learners of English as an L2 and investigated text readability assessment for both native and L2 learners. We applied a generalization method to adapt models trained on larger native corpora to estimate text readability for learners, and explored domain adaptation and self-learning techniques to make use of the native data to improve system performance on the limited L2 data. In our experiments, the best performing model for readability on learner texts achieves an accuracy of 0.797 and PCC of $0.938$.
We report our ongoing work about a new deep architecture working in tandem with a statistical test procedure for jointly training texts and their label descriptions for multi-label and multi-class classification tasks. A statistical hypothesis testing method is used to extract the most informative words for each given class. These words are used as a class description for more label-aware text classification. Intuition is to help the model to concentrate on more informative words rather than more frequent ones. The model leverages the use of label descriptions in addition to the input text to enhance text classification performance. Our method is entirely data-driven, has no dependency on other sources of information than the training data, and is adaptable to different classification problems by providing appropriate training data without major hyper-parameter tuning. We trained and tested our system on several publicly available datasets, where we managed to improve the state-of-the-art on one set with a high margin, and to obtain competitive results on all other ones.