BERTs are incapable of processing long texts due to its quadratically increasing memory and time consumption. The straightforward thoughts to address this problem, such as slicing the text by a sliding window or simplifying transformers, suffer from insufficient long-range attentions or need customized CUDA kernels. The limited text length of BERT reminds us the limited capacity (5∼ 9 chunks) of the working memory of humans – then how do human beings Cognize Long TeXts? Founded on the cognitive theory stemming from Baddeley, our CogLTX framework identifies key sentences by training a judge model, concatenates them for reasoning and enables multi-step reasoning via rehearsal and decay. Since relevance annotations are usually unavailable, we propose to use treatment experiments to create supervision. As a general algorithm, CogLTX outperforms or gets comparable results to SOTA models on NewsQA, HotpotQA, multi-class and multi-label long-text classification tasks with memory overheads independent of the text length.
Interpreting the inner workings of neural models is a key step in ensuring the robustness and trustworthiness of the models, but work on neural network interpretability typically faces a trade-off: either the models are too constrained to be very useful, or the solutions found by the models are too complex to interpret. We propose a novel strategy for achieving interpretability that -- in our experiments -- avoids this trade-off. Our approach builds on the success of using probability as the central quantity, such as for instance within the attention mechanism. In our architecture, DoLFIn (Distributions over Latent Features for Interpretability), we do no determine beforehand what each feature represents, and features go altogether into an unordered set. Each feature has an associated probability ranging from 0 to 1, weighing its importance for further processing. We show that, unlike attention and saliency map approaches, this set-up makes it straight-forward to compute the probability with which an input component supports the decision the neural model makes. To demonstrate the usefulness of the approach, we apply DoLFIn to text classification, and show that DoLFIn not only provides interpretable solutions, but even slightly outperforms the classical CNN and BiLSTM text classifiers on the SST2 and AG-news datasets.
In this paper we propose a multi-task sequence prediction system, based on recurrent neural networks and used to annotate on multiple levels an Arabizi Tunisian corpus. The annotation performed are text classification, tokenization, PoS tagging and encoding of Tunisian Arabizi into CODA* Arabic orthography. The system is learned to predict all the annotation levels in cascade, starting from Arabizi input. We evaluate the system on the TIGER German corpus, suitably converting data to have a multi-task problem, in order to show the effectiveness of our neural architecture. We show also how we used the system in order to annotate a Tunisian Arabizi corpus, which has been afterwards manually corrected and used to further evaluate sequence models on Tunisian data. Our system is developed for the Fairseq framework, which allows for a fast and easy use for any other sequence prediction problem.
This paper presents EstBERT, a large pretrained transformer-based language-specific BERT model for Estonian. Recent work has evaluated multilingual BERT models on Estonian tasks and found them to outperform the baselines. Still, based on existing studies on other languages, a language-specific BERT model is expected to improve over the multilingual ones. We first describe the EstBERT pretraining process and then present the results of the models based on finetuned EstBERT for multiple NLP tasks, including POS and morphological tagging, named entity recognition and text classification. The evaluation results show that the models based on EstBERT outperform multilingual BERT models on five tasks out of six, providing further evidence towards a view that training language-specific BERT models are still useful, even when multilingual models are available.
We propose a new character-based text classification framework for non-alphabetic languages, such as Chinese and Japanese. Our framework consists of a variational character encoder (VCE) and character-level text classifier. The VCE is composed of a $\beta$-variational auto-encoder ($\beta$-VAE) that learns the proposed glyph-aware disentangled character embedding (GDCE). Since our GDCE provides zero-mean unit-variance character embeddings that are dimensionally independent, it is applicable for our interpretable data augmentation, namely, semantic sub-character augmentation (SSA). In this paper, we evaluated our framework using Japanese text classification tasks at the document- and sentence-level. We confirmed that our GDCE and SSA not only provided embedding interpretability but also improved the classification performance. Our proposal achieved a competitive result to the state-of-the-art model while also providing model interpretability. Our code is available on https://github.com/IyatomiLab/GDCE-SSA
Text classification has been one of the earliest problems in NLP. Over time the scope of application areas has broadened and the difficulty of dealing with new areas (e.g., noisy social media content) has increased. The problem-solving strategy switched from classical machine learning to deep learning algorithms. One of the recent deep neural network architecture is the Transformer. Models designed with this type of network and its variants recently showed their success in many downstream natural language processing tasks, especially for resource-rich languages, e.g., English. However, these models have not been explored fully for Bangla text classification tasks. In this work, we fine-tune multilingual transformer models for Bangla text classification tasks in different domains, including sentiment analysis, emotion detection, news categorization, and authorship attribution. We obtain the state of the art results on six benchmark datasets, improving upon the previous results by 5-29% accuracy across different tasks.
Digitized archives contain and preserve the knowledge of generations of scholars in millions of documents. The size of these archives calls for automatic analysis since a manual analysis by specialists is often too expensive. In this paper, we focus on the analysis of handwriting in scanned documents from the art-historic archive of the WPI. Since the archive consists of documents written in several languages and lacks annotated training data for the creation of recognition models, we propose the task of handwriting classification as a new step for a handwriting OCR pipeline. We propose a handwriting classification model that labels extracted text fragments, eg, numbers, dates, or words, based on their visual structure. Such a classification supports historians by highlighting documents that contain a specific class of text without the need to read the entire content. To this end, we develop and compare several deep learning-based models for text classification. In extensive experiments, we show the advantages and disadvantages of our proposed approach and discuss possible usage scenarios on a real-world dataset.
Language models based on the Transformer architecture have achieved state-of-the-art performance on a wide range of NLP tasks such as text classification, question-answering, and token classification. However, this performance is usually tested and reported on high-resource languages, like English, French, Spanish, and German. Indian languages, on the other hand, are underrepresented in such benchmarks. Despite some Indian languages being included in training multilingual Transformer models, they have not been the primary focus of such work. In order to evaluate the performance on Indian languages specifically, we analyze these language models through extensive experiments on multiple downstream tasks in Hindi, Bengali, and Telugu language. Here, we compare the efficacy of fine-tuning model parameters of pre-trained models against that of training a language model from scratch. Moreover, we empirically argue against the strict dependency between the dataset size and model performance, but rather encourage task-specific model and method selection. We achieve state-of-the-art performance on Hindi and Bengali languages for text classification task. Finally, we present effective strategies for handling the modeling of Indian languages and we release our model checkpoints for the community : https://huggingface.co/neuralspace-reverie.
Emotion recognition is predominantly formulated as text classification in which textual units are assigned to an emotion from a predefined inventory (e.g., fear, joy, anger, disgust, sadness, surprise, trust, anticipation). More recently, semantic role labeling approaches have been developed to extract structures from the text to answer questions like: "who is described to feel the emotion?" (experiencer), "what causes this emotion?" (stimulus), and at which entity is it directed?" (target). Though it has been shown that jointly modeling stimulus and emotion category prediction is beneficial for both subtasks, it remains unclear which of these semantic roles enables a classifier to infer the emotion. Is it the experiencer, because the identity of a person is biased towards a particular emotion (X is always happy)? Is it a particular target (everybody loves X) or a stimulus (doing X makes everybody sad)? We answer these questions by training emotion classification models on five available datasets annotated with at least one semantic role by masking the fillers of these roles in the text in a controlled manner and find that across multiple corpora, stimuli and targets carry emotion information, while the experiencer might be considered a confounder. Further, we analyze if informing the model about the position of the role improves the classification decision. Particularly on literature corpora we find that the role information improves the emotion classification.
Most pre-trained language models (PLMs) construct word representations at subword level with Byte-Pair Encoding (BPE) or its variations, by which OOV (out-of-vocab) words are almost avoidable. However, those methods split a word into subword units and make the representation incomplete and fragile. In this paper, we propose a character-aware pre-trained language model named CharBERT improving on the previous methods (such as BERT, RoBERTa) to tackle these problems. We first construct the contextual word embedding for each token from the sequential character representations, then fuse the representations of characters and the subword representations by a novel heterogeneous interaction module. We also propose a new pre-training task named NLM (Noisy LM) for unsupervised character representation learning. We evaluate our method on question answering, sequence labeling, and text classification tasks, both on the original datasets and adversarial misspelling test sets. The experimental results show that our method can significantly improve the performance and robustness of PLMs simultaneously. Pretrained models, evaluation sets, and code are available at https://github.com/wtma/CharBERT
Producing or sharing Child Sexual Exploitation Material (CSEM) is a serious crime fought vigorously by Law Enforcement Agencies (LEAs). When an LEA seizes a computer from a potential producer or consumer of CSEM, they need to analyze the suspect's hard disk's files looking for pieces of evidence. However, a manual inspection of the file content looking for CSEM is a time-consuming task. In most cases, it is unfeasible in the amount of time available for the Spanish police using a search warrant. Instead of analyzing its content, another approach that can be used to speed up the process is to identify CSEM by analyzing the file names and their absolute paths. The main challenge for this task lies behind dealing with short text distorted deliberately by the owners of this material using obfuscated words and user-defined naming patterns. This paper presents and compares two approaches based on short text classification to identify CSEM files. The first one employs two independent supervised classifiers, one for the file name and the other for the path, and their outputs are later on fused into a single score. Conversely, the second approach uses only the file name classifier to iterate over the file's absolute path. Both approaches operate at the character n-grams level, while binary and orthographic features enrich the file name representation, and a binary Logistic Regression model is used for classification. The presented file classifier achieved an average class recall of 0.98. This solution could be integrated into forensic tools and services to support Law Enforcement Agencies to identify CSEM without tackling every file's visual content, which is computationally much more highly demanding.
One critical issue of zero anaphora resolution (ZAR) is the scarcity of labeled data. This study explores how effectively this problem can be alleviated by data augmentation. We adopt a state-of-the-art data augmentation method, called the contextual data augmentation (CDA), that generates labeled training instances using a pretrained language model. The CDA has been reported to work well for several other natural language processing tasks, including text classification and machine translation. This study addresses two underexplored issues on CDA, that is, how to reduce the computational cost of data augmentation and how to ensure the quality of the generated data. We also propose two methods to adapt CDA to ZAR: [MASK]-based augmentation and linguistically-controlled masking. Consequently, the experimental results on Japanese ZAR show that our methods contribute to both the accuracy gain and the computation cost reduction. Our closer analysis reveals that the proposed method can improve the quality of the augmented training data when compared to the conventional CDA.
In natural languages, words are used in association to construct sentences. It is not words in isolation, but the appropriate combination of hierarchical structures that conveys the meaning of the whole sentence. Neural networks can capture expressive language features; however, insights into the link between words and sentences are difficult to acquire automatically. In this work, we design a deep neural network architecture that explicitly wires lower and higher linguistic components; we then evaluate its ability to perform the same task at different hierarchical levels. Settling on broad text classification tasks, we show that our model, MHAL, learns to simultaneously solve them at different levels of granularity by fluidly transferring knowledge between hierarchies. Using a multi-head attention mechanism to tie the representations between single words and full sentences, MHAL systematically outperforms equivalent models that are not incentivized towards developing compositional representations. Moreover, we demonstrate that, with the proposed architecture, the sentence information flows naturally to individual words, allowing the model to behave like a sequence labeller (which is a lower, word-level task) even without any word supervision, in a zero-shot fashion.
Text classification is a critical research topic with broad applications in natural language processing. Recently, graph neural networks (GNNs) have received increasing attention in the research community and demonstrated their promising results on this canonical task. Despite the success, their performance could be largely jeopardized in practice since they are: (1) unable to capture high-order interaction between words; (2) inefficient to handle large datasets and new documents. To address those issues, in this paper, we propose a principled model -- hypergraph attention networks (HyperGAT), which can obtain more expressive power with less computational consumption for text representation learning. Extensive experiments on various benchmark datasets demonstrate the efficacy of the proposed approach on the text classification task.
Natural language data exhibit tree-like hierarchical structures such as the hypernym-hyponym relations in WordNet. FastText, as the state-of-the-art text classifier based on shallow neural network in Euclidean space, may not model such hierarchies precisely with limited representation capacity. Considering that hyperbolic space is naturally suitable for modeling tree-like hierarchical data, we propose a new model named HyperText for efficient text classification by endowing FastText with hyperbolic geometry. Empirically, we show that HyperText outperforms FastText on a range of text classification tasks with much reduced parameters.
Existing work in multilingual pretraining has demonstrated the potential of cross-lingual transferability by training a unified Transformer encoder for multiple languages. However, much of this work only relies on the shared vocabulary and bilingual contexts to encourage the correlation across languages, which is loose and implicit for aligning the contextual representations between languages. In this paper, we plug a cross-attention module into the Transformer encoder to explicitly build the interdependence between languages. It can effectively avoid the degeneration of predicting masked words only conditioned on the context in its own language. More importantly, when fine-tuning on downstream tasks, the cross-attention module can be plugged in or out on-demand, thus naturally benefiting a wider range of cross-lingual tasks, from language understanding to generation. As a result, the proposed cross-lingual model delivers new state-of-the-art results on various cross-lingual understanding tasks of the XTREME benchmark, covering text classification, sequence labeling, question answering, and sentence retrieval. For cross-lingual generation tasks, it also outperforms all existing cross-lingual models and state-of-the-art Transformer variants on WMT14 English-to-German and English-to-French translation datasets, with gains of up to 1~2 BLEU.
Despite the widespread application of recurrent neural networks (RNNs) across a variety of tasks, a unified understanding of how RNNs solve these tasks remains elusive. In particular, it is unclear what dynamical patterns arise in trained RNNs, and how those patterns depend on the training dataset or task. This work addresses these questions in the context of a specific natural language processing task: text classification. Using tools from dynamical systems analysis, we study recurrent networks trained on a battery of both natural and synthetic text classification tasks. We find the dynamics of these trained RNNs to be both interpretable and low-dimensional. Specifically, across architectures and datasets, RNNs accumulate evidence for each class as they process the text, using a low-dimensional attractor manifold as the underlying mechanism. Moreover, the dimensionality and geometry of the attractor manifold are determined by the structure of the training dataset; in particular, we describe how simple word-count statistics computed on the training dataset can be used to predict these properties. Our observations span multiple architectures and datasets, reflecting a common mechanism RNNs employ to perform text classification. To the degree that integration of evidence towards a decision is a common computational primitive, this work lays the foundation for using dynamical systems techniques to study the inner workings of RNNs.
In order to improve the accuracy performance of Chinese text classification models with low hardware requirements, an improved concatenation-based model is designed in this paper, which is a concatenation of 5 different sub-models, including TextCNN, LSTM, and Bi-LSTM. Compared with the existing ensemble learning method, for a text classification mission, this model's accuracy is 2% higher. Meanwhile, the hardware requirements of this model are much lower than the BERT-based model.
Large multi-label text classification is a challenging Natural Language Processing (NLP) problem that is concerned with text classification for datasets with thousands of labels. We tackle this problem in the legal domain, where datasets, such as JRC-Acquis and EURLEX57K labeled with the EuroVoc vocabulary were created within the legal information systems of the European Union. The EuroVoc taxonomy includes around 7000 concepts. In this work, we study the performance of various recent transformer-based models in combination with strategies such as generative pretraining, gradual unfreezing and discriminative learning rates in order to reach competitive classification performance, and present new state-of-the-art results of 0.661 (F1) for JRC-Acquis and 0.754 for EURLEX57K. Furthermore, we quantify the impact of individual steps, such as language model fine-tuning or gradual unfreezing in an ablation study, and provide reference dataset splits created with an iterative stratification algorithm.
Fine-tuned language models have been shown to exhibit biases against protected groups in a host of modeling tasks such as text classification and coreference resolution. Previous works focus on detecting these biases, reducing bias in data representations, and using auxiliary training objectives to mitigate bias during fine-tuning. Although these techniques achieve bias reduction for the task and domain at hand, the effects of bias mitigation may not directly transfer to new tasks, requiring additional data collection and customized annotation of sensitive attributes, and re-evaluation of appropriate fairness metrics. We explore the feasibility and benefits of upstream bias mitigation (UBM) for reducing bias on downstream tasks, by first applying bias mitigation to an upstream model through fine-tuning and subsequently using it for downstream fine-tuning. We find, in extensive experiments across hate speech detection, toxicity detection, occupation prediction, and coreference resolution tasks over various bias factors, that the effects of UBM are indeed transferable to new downstream tasks or domains via fine-tuning, creating less biased downstream models than directly fine-tuning on the downstream task or transferring from a vanilla upstream model. Though challenges remain, we show that UBM promises more efficient and accessible bias mitigation in LM fine-tuning.