We report results of the CASE 2022 Shared Task 1 on Multilingual Protest Event Detection. This task is a continuation of CASE 2021 that consists of four subtasks that are i) document classification, ii) sentence classification, iii) event sentence coreference identification, and iv) event extraction. The CASE 2022 extension consists of expanding the test data with more data in previously available languages, namely, English, Hindi, Portuguese, and Spanish, and adding new test data in Mandarin, Turkish, and Urdu for Sub-task 1, document classification. The training data from CASE 2021 in English, Portuguese and Spanish were utilized. Therefore, predicting document labels in Hindi, Mandarin, Turkish, and Urdu occurs in a zero-shot setting. The CASE 2022 workshop accepts reports on systems developed for predicting test data of CASE 2021 as well. We observe that the best systems submitted by CASE 2022 participants achieve between 79.71 and 84.06 F1-macro for new languages in a zero-shot setting. The winning approaches are mainly ensembling models and merging data in multiple languages. The best two submissions on CASE 2021 data outperform submissions from last year for Subtask 1 and Subtask 2 in all languages. Only the following scenarios were not outperformed by new submissions on CASE 2021: Subtask 3 Portuguese \& Subtask 4 English.
Pre-trained Transformers currently dominate most NLP tasks. They impose, however, limits on the maximum input length (512 sub-words in BERT), which are too restrictive in the legal domain. Even sparse-attention models, such as Longformer and BigBird, which increase the maximum input length to 4,096 sub-words, severely truncate texts in three of the six datasets of LexGLUE. Simpler linear classifiers with TF-IDF features can handle texts of any length, require far less resources to train and deploy, but are usually outperformed by pre-trained Transformers. We explore two directions to cope with long legal texts: (i) modifying a Longformer warm-started from LegalBERT to handle even longer texts (up to 8,192 sub-words), and (ii) modifying LegalBERT to use TF-IDF representations. The first approach is the best in terms of performance, surpassing a hierarchical version of LegalBERT, which was the previous state of the art in LexGLUE. The second approach leads to computationally more efficient models at the expense of lower performance, but the resulting models still outperform overall a linear SVM with TF-IDF features in long legal document classification.
Pre-trained language models have attracted increasing attention in the biomedical domain, inspired by their great success in the general natural language domain. Among the two main branches of pre-trained language models in the general language domain, i.e., BERT (and its variants) and GPT (and its variants), the first one has been extensively studied in the biomedical domain, such as BioBERT and PubMedBERT. While they have achieved great success on a variety of discriminative downstream biomedical tasks, the lack of generation ability constrains their application scope. In this paper, we propose BioGPT, a domain-specific generative Transformer language model pre-trained on large scale biomedical literature. We evaluate BioGPT on six biomedical NLP tasks and demonstrate that our model outperforms previous models on most tasks. Especially, we get 44.98%, 38.42% and 40.76% F1 score on BC5CDR, KD-DTI and DDI end-to-end relation extraction tasks respectively, and 78.2% accuracy on PubMedQA, creating a new record. Our case study on text generation further demonstrates the advantage of BioGPT on biomedical literature to generate fluent descriptions for biomedical terms. Code is available at https://github.com/microsoft/BioGPT.
The ability of a document classifier to handle inputs that are drawn from a distribution different from the training distribution is crucial for robust deployment and generalizability. The RVL-CDIP corpus is the de facto standard benchmark for document classification, yet to our knowledge all studies that use this corpus do not include evaluation on out-of-distribution documents. In this paper, we curate and release a new out-of-distribution benchmark for evaluating out-of-distribution performance for document classifiers. Our new out-of-distribution benchmark consists of two types of documents: those that are not part of any of the 16 in-domain RVL-CDIP categories (RVL-CDIP-O), and those that are one of the 16 in-domain categories yet are drawn from a distribution different from that of the original RVL-CDIP dataset (RVL-CDIP-N). While prior work on document classification for in-domain RVL-CDIP documents reports high accuracy scores, we find that these models exhibit accuracy drops of between roughly 15-30% on our new out-of-domain RVL-CDIP-N benchmark, and further struggle to distinguish between in-domain RVL-CDIP-N and out-of-domain RVL-CDIP-O inputs. Our new benchmark provides researchers with a valuable new resource for analyzing out-of-distribution performance on document classifiers. Our new out-of-distribution data can be found at https://github.com/gxlarson/rvl-cdip-ood.
In this paper, we consider the task of retrieving documents with predefined topics from an unlabeled document dataset using an unsupervised approach. The proposed unsupervised approach requires only a small number of keywords describing the respective topics and no labeled document. Existing approaches either heavily relied on a large amount of additionally encoded world knowledge or on term-document frequencies. Contrariwise, we introduce a method that learns jointly embedded document and word vectors solely from the unlabeled document dataset in order to find documents that are semantically similar to the topics described by the keywords. The proposed method requires almost no text preprocessing but is simultaneously effective at retrieving relevant documents with high probability. When successively retrieving documents on different predefined topics from publicly available and commonly used datasets, we achieved an average area under the receiver operating characteristic curve value of 0.95 on one dataset and 0.92 on another. Further, our method can be used for multiclass document classification, without the need to assign labels to the dataset in advance. Compared with an unsupervised classification baseline, we increased F1 scores from 76.6 to 82.7 and from 61.0 to 75.1 on the respective datasets. For easy replication of our approach, we make the developed Lbl2Vec code publicly available as a ready-to-use tool under the 3-Clause BSD license.
Non-hierarchical sparse attention Transformer-based models, such as Longformer and Big Bird, are popular approaches to working with long documents. There are clear benefits to these approaches compared to the original Transformer in terms of efficiency, but Hierarchical Attention Transformer (HAT) models are a vastly understudied alternative. We develop and release fully pre-trained HAT models that use segment-wise followed by cross-segment encoders and compare them with Longformer models and partially pre-trained HATs. In several long document downstream classification tasks, our best HAT model outperforms equally-sized Longformer models while using 10-20% less GPU memory and processing documents 40-45% faster. In a series of ablation studies, we find that HATs perform best with cross-segment contextualization throughout the model than alternative configurations that implement either early or late cross-segment contextualization. Our code is on GitHub: https://github.com/coastalcph/hierarchical-transformers.
We investigate semi-structured document classification in a zero-shot setting. Classification of semi-structured documents is more challenging than that of standard unstructured documents, as positional, layout, and style information play a vital role in interpreting such documents. The standard classification setting where categories are fixed during both training and testing falls short in dynamic environments where new document categories could potentially emerge. We focus exclusively on the zero-shot setting where inference is done on new unseen classes. To address this task, we propose a matching-based approach that relies on a pairwise contrastive objective for both pretraining and fine-tuning. Our results show a significant boost in Macro F$_1$ from the proposed pretraining step in both supervised and unsupervised zero-shot settings.
Using a taxonomy to organize information requires classifying objects (documents, images, etc) with appropriate taxonomic classes. The flexible nature of zero-shot learning is appealing for this task because it allows classifiers to naturally adapt to taxonomy modifications. This work studies zero-shot multi-label document classification with fine-tuned language models under realistic taxonomy expansion scenarios in the human resource domain. Experiments show that zero-shot learning can be highly effective in this setting. When controlling for training data budget, zero-shot classifiers achieve a 12% relative increase in macro-AP when compared to a traditional multi-label classifier trained on all classes. Counterintuitively, these results suggest in some settings it would be preferable to adopt zero-shot techniques and spend resources annotating more documents with an incomplete set of classes, rather than spreading the labeling budget uniformly over all classes and using traditional classification techniques. Additional experiments demonstrate that adopting the well-known filter/re-rank decomposition from the recommender systems literature can significantly reduce the computational burden of high-performance zero-shot classifiers, empirically resulting in a 98% reduction in computational overhead for only a 2% relative decrease in performance. The evidence presented here demonstrates that zero-shot learning has the potential to significantly increase the flexibility of taxonomies and highlights directions for future research.
BERT achieves remarkable results in text classification tasks, it is yet not fully exploited, since only the last layer is used as a representation output for downstream classifiers. The most recent studies on the nature of linguistic features learned by BERT, suggest that different layers focus on different kinds of linguistic features. We propose a CNN-Enhanced Transformer-Encoder model which is trained on top of fixed BERT $[CLS]$ representations from all layers, employing Convolutional Neural Networks to generate QKV feature maps inside the Transformer-Encoder, instead of linear projections of the input into the embedding space. CNN-Trans-Enc is relatively small as a downstream classifier and doesn't require any fine-tuning of BERT, as it ensures an optimal use of the $[CLS]$ representations from all layers, leveraging different linguistic features with more meaningful, and generalizable QKV representations of the input. Using BERT with CNN-Trans-Enc keeps $98.9\%$ and $94.8\%$ of current state-of-the-art performance on the IMDB and SST-5 datasets respectably, while obtaining new state-of-the-art on YELP-5 with $82.23$ ($8.9\%$ improvement), and on Amazon-Polarity with $0.98\%$ ($0.2\%$ improvement) (K-fold Cross Validation on a 1M sample subset from both datasets). On the AG news dataset CNN-Trans-Enc achieves $99.94\%$ of the current state-of-the-art, and achieves a new top performance with an average accuracy of $99.51\%$ on DBPedia-14. Index terms: Text Classification, Natural Language Processing, Convolutional Neural Networks, Transformers, BERT
Multi-modal document pre-trained models have proven to be very effective in a variety of visually-rich document understanding (VrDU) tasks. Though existing document pre-trained models have achieved excellent performance on standard benchmarks for VrDU, the way they model and exploit the interactions between vision and language on documents has hindered them from better generalization ability and higher accuracy. In this work, we investigate the problem of vision-language joint representation learning for VrDU mainly from the perspective of supervisory signals. Specifically, a pre-training paradigm called Bi-VLDoc is proposed, in which a bidirectional vision-language supervision strategy and a vision-language hybrid-attention mechanism are devised to fully explore and utilize the interactions between these two modalities, to learn stronger cross-modal document representations with richer semantics. Benefiting from the learned informative cross-modal document representations, Bi-VLDoc significantly advances the state-of-the-art performance on three widely-used document understanding benchmarks, including Form Understanding (from 85.14% to 93.44%), Receipt Information Extraction (from 96.01% to 97.84%), and Document Classification (from 96.08% to 97.12%). On Document Visual QA, Bi-VLDoc achieves the state-of-the-art performance compared to previous single model methods.
Supervised dictionary learning (SDL) is a classical machine learning method that simultaneously seeks feature extraction and classification tasks, which are not necessarily a priori aligned objectives. The goal of SDL is to learn a class-discriminative dictionary, which is a set of latent feature vectors that can well-explain both the features as well as labels of observed data. In this paper, we provide a systematic study of SDL, including the theory, algorithm, and applications of SDL. First, we provide a novel framework that `lifts' SDL as a convex problem in a combined factor space and propose a low-rank projected gradient descent algorithm that converges exponentially to the global minimizer of the objective. We also formulate generative models of SDL and provide global estimation guarantees of the true parameters depending on the hyperparameter regime. Second, viewed as a nonconvex constrained optimization problem, we provided an efficient block coordinate descent algorithm for SDL that is guaranteed to find an $\varepsilon$-stationary point of the objective in $O(\varepsilon^{-1}(\log \varepsilon^{-1})^{2})$ iterations. For the corresponding generative model, we establish a novel non-asymptotic local consistency result for constrained and regularized maximum likelihood estimation problems, which may be of independent interest. Third, we apply SDL for imbalanced document classification by supervised topic modeling and also for pneumonia detection from chest X-ray images. We also provide simulation studies to demonstrate that SDL becomes more effective when there is a discrepancy between the best reconstructive and the best discriminative dictionaries.
Sequential data naturally have different lengths in many domains, with some very long sequences. As an important modeling tool, neural attention should capture long-range interaction in such sequences. However, most existing neural attention models admit only short sequences, or they have to employ chunking or padding to enforce a constant input length. Here we propose a simple neural network building block called ChordMixer which can model the attention for long sequences with variable lengths. Each ChordMixer block consists of a position-wise rotation layer without learnable parameters and an element-wise MLP layer. Repeatedly applying such blocks forms an effective network backbone that mixes the input signals towards the learning targets. We have tested ChordMixer on the synthetic adding problem, long document classification, and DNA sequence-based taxonomy classification. The experiment results show that our method substantially outperforms other neural attention models.
Convolutional Neural Networks (ConvNets) have been thoroughly researched for document image classification and are known for their exceptional performance in unimodal image-based document classification. Recently, however, there has been a sudden shift in the field towards multimodal approaches that simultaneously learn from the visual and textual features of the documents. While this has led to significant advances in the field, it has also led to a waning interest in improving pure ConvNets-based approaches. This is not desirable, as many of the multimodal approaches still use ConvNets as their visual backbone, and thus improving ConvNets is essential to improving these approaches. In this paper, we present DocXClassifier, a ConvNet-based approach that, using state-of-the-art model design patterns together with modern data augmentation and training strategies, not only achieves significant performance improvements in image-based document classification, but also outperforms some of the recently proposed multimodal approaches. Moreover, DocXClassifier is capable of generating transformer-like attention maps, which makes it inherently interpretable, a property not found in previous image-based classification models. Our approach achieves a new peak performance in image-based classification on two popular document datasets, namely RVL-CDIP and Tobacco3482, with a top-1 classification accuracy of 94.17% and 95.57% on the two datasets, respectively. Moreover, it sets a new record for the highest image-based classification accuracy of 90.14% on Tobacco3482 without transfer learning from RVL-CDIP. Finally, our proposed model may serve as a powerful visual backbone for future multimodal approaches, by providing much richer visual features than existing counterparts.
Document classification is the detection specific content of interest in text documents. In contrast to the data-driven machine learning classifiers, knowledge-based classifiers can be constructed based on domain specific knowledge, which usually takes the form of a collection of subject related keywords. While typical knowledge-based classifiers compute a prediction score based on the keyword abundance, it generally suffers from noisy detections due to the lack of guiding principle in gauging the keyword matches. In this paper, we propose a novel knowledge-based model equipped with Shannon Entropy, which measures the richness of information and favors uniform and diverse keyword matches. Without invoking any positive sample, such method provides a simple and explainable solution for document classification. We show that the Shannon Entropy significantly improves the recall at fixed level of false positive rate. Also, we show that the model is more robust against change of data distribution at inference while compared with traditional machine learning, particularly when the positive training samples are very limited.
While Identity Document Verification (IDV) technology on mobile devices becomes ubiquitous in modern business operations, the risk of identity theft and fraud is increasing. The identity document holder is normally required to participate in an online video interview to circumvent impostors. However, the current IDV process depends on an additional human workforce to support online step-by-step guidance which is inefficient and expensive. The performance of existing AI-based approaches cannot meet the real-time and lightweight demands of mobile devices. In this paper, we address those challenges by designing an edge intelligence-assisted approach for real-time IDV. Aiming at improving the responsiveness of the IDV process, we propose a new document localization model for mobile devices, LDRNet, to Localize the identity Document in Real-time. On the basis of a lightweight backbone network, we build three prediction branches for LDRNet, the corner points prediction, the line borders prediction and the document classification. We design novel supplementary targets, the equal-division points, and use a new loss function named Line Loss, to improve the speed and accuracy of our approach. In addition to the IDV process, LDRNet is an efficient and reliable document localization alternative for all kinds of mobile applications. As a matter of proof, we compare the performance of LDRNet with other popular approaches on localizing general document datasets. The experimental results show that LDRNet runs at a speed up to 790 FPS which is 47x faster, while still achieving comparable Jaccard Index(JI) in single-model and single-scale tests.
A typical desideratum for quantifying the uncertainty from a classification model as a prediction set is class-conditional singleton set calibration. That is, such sets should map to the output of well-calibrated selective classifiers, matching the observed frequencies of similar instances. Recent works proposing adaptive and localized conformal p-values for deep networks do not guarantee this behavior, nor do they achieve it empirically. Instead, we use the strong signals for prediction reliability from KNN-based approximations of Transformer networks to construct data-driven partitions for Mondrian Conformal Predictors, which are treated as weak selective classifiers that are then calibrated via a new Inductive Venn Predictor, the Venn-ADMIT Predictor. The resulting selective classifiers are well-calibrated, in a conservative but practically useful sense for a given threshold. They are inherently robust to changes in the proportions of the data partitions, and straightforward conservative heuristics provide additional robustness to covariate shifts. We compare and contrast to the quantities produced by recent Conformal Predictors on several representative and challenging natural language processing classification tasks, including class-imbalanced and distribution-shifted settings.
Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3$\times$ speedup on GPT-2 (seq. length 1K), and 2.4$\times$ speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).
Transformer language models provide superior accuracy over previous models but they are computationally and environmentally expensive. Borrowing the concept of model cascading from computer vision, we introduce BabyBear, a framework for cascading models for natural language processing (NLP) tasks to minimize cost. The core strategy is inference triage, exiting early when the least expensive model in the cascade achieves a sufficiently high-confidence prediction. We test BabyBear on several open source data sets related to document classification and entity recognition. We find that for common NLP tasks a high proportion of the inference load can be accomplished with cheap, fast models that have learned by observing a deep learning model. This allows us to reduce the compute cost of large-scale classification jobs by more than 50% while retaining overall accuracy. For named entity recognition, we save 33% of the deep learning compute while maintaining an F1 score higher than 95% on the CoNLL benchmark.
Multimodal learning from document data has achieved great success lately as it allows to pre-train semantically meaningful features as a prior into a learnable downstream task. In this paper, we approach the document classification problem by learning cross-modal representations through language and vision cues, considering intra- and inter-modality relationships. Instead of merging features from different modalities into a joint representation space, the proposed method exploits high-level interactions and learns relevant semantic information from effective attention flows within and across modalities. The proposed learning objective is devised between intra- and inter-modality alignment tasks, where the similarity distribution per task is computed by contracting positive sample pairs while simultaneously contrasting negative ones in the joint representation space}. Extensive experiments on public document classification datasets demonstrate the effectiveness and the generality of our model on low-scale and large-scale datasets.
Word embeddings are one of the most fundamental technologies used in natural language processing. Existing word embeddings are high-dimensional and consume considerable computational resources. In this study, we propose WordTour, unsupervised one-dimensional word embeddings. To achieve the challenging goal, we propose a decomposition of the desiderata of word embeddings into two parts, completeness and soundness, and focus on soundness in this paper. Owing to the single dimensionality, WordTour is extremely efficient and provides a minimal means to handle word embeddings. We experimentally confirmed the effectiveness of the proposed method via user study and document classification.