Feature importance (FI) estimates are a popular form of explanation, and they are commonly created and evaluated by computing the change in model confidence caused by removing certain input features at test time. For example, in the standard Sufficiency metric, only the top-k most important tokens are kept. In this paper, we study several under-explored dimensions of FI explanations, providing conceptual and empirical improvements for this form of explanation. First, we advance a new argument for why it can be problematic to remove features from an input when creating or evaluating explanations: the fact that these counterfactual inputs are out-of-distribution (OOD) to models implies that the resulting explanations are socially misaligned. The crux of the problem is that the model prior and random weight initialization influence the explanations (and explanation metrics) in unintended ways. To resolve this issue, we propose a simple alteration to the model training process, which results in more socially aligned explanations and metrics. Second, we compare among five approaches for removing features from model inputs. We find that some methods produce more OOD counterfactuals than others, and we make recommendations for selecting a feature-replacement function. Finally, we introduce four search-based methods for identifying FI explanations and compare them to strong baselines, including LIME, Anchors, and Integrated Gradients. Through experiments with six diverse text classification datasets, we find that the only method that consistently outperforms random search is a Parallel Local Search (PLS) that we introduce. Improvements over the second-best method are as large as 5.4 points for Sufficiency and 17 points for Comprehensiveness. All supporting code for experiments in this paper is publicly available at https://github.com/peterbhase/ExplanationSearch.
The task of rationalization aims to extract pieces of input text as rationales to justify neural network predictions on text classification tasks. By definition, rationales represent key text pieces used for prediction and thus should have similar classification feature distribution compared to the original input text. However, previous methods mainly focused on maximizing the mutual information between rationales and labels while neglecting the relationship between rationales and input text. To address this issue, we propose a novel rationalization method that matches the distributions of rationales and input text in both the feature space and output space. Empirically, the proposed distribution matching approach consistently outperforms previous methods by a large margin. Our data and code are available.
Accurately extracting structured content from PDFs is a critical first step for NLP over scientific papers. Recent work has improved extraction accuracy by incorporating elementary layout information, e.g., each token's 2D position on the page, into language model pretraining. We introduce new methods that explicitly model VIsual LAyout (VILA) groups, i.e., text lines or text blocks, to further improve performance. In our I-VILA approach, we show that simply inserting special tokens denoting layout group boundaries into model inputs can lead to a 1.9% Macro F1 improvement in token classification. In the H-VILA approach, we show that hierarchical encoding of layout-groups can result in up-to 47% inference time reduction with less than 0.8% Macro F1 loss. Unlike prior layout-aware approaches, our methods do not require expensive additional pretraining, only fine-tuning, which we show can reduce training cost by up to 95%. Experiments are conducted on a newly curated evaluation suite, S2-VLUE, that unifies existing automatically-labeled datasets and includes a new dataset of manual annotations covering diverse papers from 19 scientific disciplines. Pre-trained weights, benchmark datasets, and source code are available at https://github.com/allenai/VILA.
In Natural Language Processing (NLP), finding data augmentation techniques that can produce high-quality human-interpretable examples has always been challenging. Recently, leveraging kNN such that augmented examples are retrieved from large repositories of unlabelled sentences has made a step toward interpretable augmentation. Inspired by this paradigm, we introduce Minimax-kNN, a sample efficient data augmentation strategy tailored for Knowledge Distillation (KD). We exploit a semi-supervised approach based on KD to train a model on augmented data. In contrast to existing kNN augmentation techniques that blindly incorporate all samples, our method dynamically selects a subset of augmented samples that maximizes KL-divergence between the teacher and student models. This step aims to extract the most efficient samples to ensure our augmented data covers regions in the input space with maximum loss value. We evaluated our technique on several text classification tasks and demonstrated that Minimax-kNN consistently outperforms strong baselines. Our results show that Minimax-kNN requires fewer augmented examples and less computation to achieve superior performance over the state-of-the-art kNN-based augmentation techniques.
This paper presents Sparse Tensor Classifier (STC), a supervised classification algorithm for categorical data inspired by the notion of superposition of states in quantum physics. By regarding an observation as a superposition of features, we introduce the concept of wave-particle duality in machine learning and propose a generalized framework that unifies the classical and the quantum probability. We show that STC possesses a wide range of desirable properties not available in most other machine learning methods but it is at the same time exceptionally easy to comprehend and use. Empirical evaluation of STC on structured data and text classification demonstrates that our methodology achieves state-of-the-art performances compared to both standard classifiers and deep learning, at the additional benefit of requiring minimal data pre-processing and hyper-parameter tuning. Moreover, STC provides a native explanation of its predictions both for single instances and for each target label globally.
Since traditional tokenizers are isolated from a downstream task and model, they cannot output an appropriate tokenization depending on the task and model, although recent studies imply that the appropriate tokenization improves the performance. In this paper, we propose a novel method to find an appropriate tokenization to a given downstream model by jointly optimizing a tokenizer and the model. The proposed method has no restriction except for using loss values computed by the downstream model to train the tokenizer, and thus, we can apply the proposed method to any NLP task. Moreover, the proposed method can be used to explore the appropriate tokenization for an already trained model as post-processing. Therefore, the proposed method is applicable to various situations. We evaluated whether our method contributes to improving performance on text classification in three languages and machine translation in eight language pairs. Experimental results show that our proposed method improves the performance by determining appropriate tokenizations.
We propose Predict then Interpolate (PI), a simple algorithm for learning correlations that are stable across environments. The algorithm follows from the intuition that when using a classifier trained on one environment to make predictions on examples from another environment, its mistakes are informative as to which correlations are unstable. In this work, we prove that by interpolating the distributions of the correct predictions and the wrong predictions, we can uncover an oracle distribution where the unstable correlation vanishes. Since the oracle interpolation coefficients are not accessible, we use group distributionally robust optimization to minimize the worst-case risk across all such interpolations. We evaluate our method on both text classification and image classification. Empirical results demonstrate that our algorithm is able to learn robust classifiers (outperforms IRM by 23.85% on synthetic environments and 12.41% on natural environments). Our code and data are available at https://github.com/YujiaBao/Predict-then-Interpolate.
Fine-tuned pre-trained language models (PLMs) have achieved awesome performance on almost all NLP tasks. By using additional prompts to fine-tune PLMs, we can further stimulate the rich knowledge distributed in PLMs to better serve downstream tasks. Prompt tuning has achieved promising results on some few-class classification tasks such as sentiment classification and natural language inference. However, manually designing lots of language prompts is cumbersome and fallible. For those auto-generated prompts, it is also expensive and time-consuming to verify their effectiveness in non-few-shot scenarios. Hence, it is still challenging for prompt tuning to address many-class classification tasks. To this end, we propose prompt tuning with rules (PTR) for many-class text classification and apply logic rules to construct prompts with several sub-prompts. In this way, PTR is able to encode prior knowledge of each class into prompt tuning. We conduct experiments on relation classification, a typical and complicated many-class classification task, and the results show that PTR can significantly and consistently outperform existing state-of-the-art baselines. This indicates that PTR is a promising approach to take advantage of both human prior knowledge and PLMs for those complicated classification tasks.
Cross-lingual text classification aims at training a classifier on the source language and transferring the knowledge to target languages, which is very useful for low-resource languages. Recent multilingual pretrained language models (mPLM) achieve impressive results in cross-lingual classification tasks, but rarely consider factors beyond semantic similarity, causing performance degradation between some language pairs. In this paper we propose a simple yet effective method to incorporate heterogeneous information within and across languages for cross-lingual text classification using graph convolutional networks (GCN). In particular, we construct a heterogeneous graph by treating documents and words as nodes, and linking nodes with different relations, which include part-of-speech roles, semantic similarity, and document translations. Extensive experiments show that our graph-based method significantly outperforms state-of-the-art models on all tasks, and also achieves consistent performance gain over baselines in low-resource settings where external tools like translators are unavailable.
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.
Continual or lifelong learning has been a long-standing challenge in machine learning to date, especially in natural language processing (NLP). Although state-of-the-art language models such as BERT have ushered in a new era in this field due to their outstanding performance in multitask learning scenarios, they suffer from forgetting when being exposed to a continuous stream of data with shifting data distributions. In this paper, we introduce DRILL, a novel continual learning architecture for open-domain text classification. DRILL leverages a biologically inspired self-organizing neural architecture to selectively gate latent language representations from BERT in a task-incremental manner. We demonstrate in our experiments that DRILL outperforms current methods in a realistic scenario of imbalanced, non-stationary data without prior knowledge about task boundaries. To the best of our knowledge, DRILL is the first of its kind to use a self-organizing neural architecture for open-domain lifelong learning in NLP.
Recent advances in Transformer models allow for unprecedented sequence lengths, due to linear space and time complexity. In the meantime, relative positional encoding (RPE) was proposed as beneficial for classical Transformers and consists in exploiting lags instead of absolute positions for inference. Still, RPE is not available for the recent linear-variants of the Transformer, because it requires the explicit computation of the attention matrix, which is precisely what is avoided by such methods. In this paper, we bridge this gap and present Stochastic Positional Encoding as a way to generate PE that can be used as a replacement to the classical additive (sinusoidal) PE and provably behaves like RPE. The main theoretical contribution is to make a connection between positional encoding and cross-covariance structures of correlated Gaussian processes. We illustrate the performance of our approach on the Long-Range Arena benchmark and on music generation.
In this work, we propose BertGCN, a model that combines large scale pretraining and transductive learning for text classification. BertGCN constructs a heterogeneous graph over the dataset and represents documents as nodes using BERT representations. By jointly training the BERT and GCN modules within BertGCN, the proposed model is able to leverage the advantages of both worlds: large-scale pretraining which takes the advantage of the massive amount of raw data and transductive learning which jointly learns representations for both training data and unlabeled test data by propagating label influence through graph convolution. Experiments show that BertGCN achieves SOTA performances on a wide range of text classification datasets. Code is available at https://github.com/ZeroRin/BertGCN.
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.
We show that Transformer encoder architectures can be sped up, with limited accuracy costs, by replacing the self-attention sublayers with simple linear transformations that "mix" input tokens. These linear mixers, along with standard nonlinearities in feed-forward layers, prove competent at modeling semantic relationships in several text classification tasks. Most surprisingly, we find that replacing the self-attention sublayer in a Transformer encoder with a standard, unparameterized Fourier Transform achieves 92-97% of the accuracy of BERT counterparts on the GLUE benchmark, but trains 80% faster on GPUs and 70% faster on TPUs at standard 512 input lengths. At longer input lengths, our FNet model is significantly faster: when compared to the "efficient" Transformers on the Long Range Arena benchmark, FNet matches the accuracy of the most accurate models, while outpacing the fastest models across all sequence lengths on GPUs (and across relatively shorter lengths on TPUs). Finally, FNet has a light memory footprint and is particularly efficient at smaller model sizes; for a fixed speed and accuracy budget, small FNet models outperform Transformer counterparts.
The introduction of pretrained cross-lingual language models brought decisive improvements to multilingual NLP tasks. However, the lack of labelled task data necessitates a variety of methods aiming to close the gap to high-resource languages. Zero-shot methods in particular, often use translated task data as a training signal to bridge the performance gap between the source and target language(s). We introduce XeroAlign, a simple method for task-specific alignment of cross-lingual pretrained transformers such as XLM-R. XeroAlign uses translated task data to encourage the model to generate similar sentence embeddings for different languages. The XeroAligned XLM-R, called XLM-RA, shows strong improvements over the baseline models to achieve state-of-the-art zero-shot results on three multilingual natural language understanding tasks. XLM-RA's text classification accuracy exceeds that of XLM-R trained with labelled data and performs on par with state-of-the-art models on a cross-lingual adversarial paraphrasing task.
In some memory-constrained settings like IoT devices and over-the-network data pipelines, it can be advantageous to have smaller contextual embeddings. We investigate the efficacy of projecting contextual embedding data (BERT) onto a manifold, and using nonlinear dimensionality reduction techniques to compress these embeddings. In particular, we propose a novel post-processing approach, applying a combination of Isomap and PCA. We find that the geodesic distance estimations, estimates of the shortest path on a Riemannian manifold, from Isomap's k-Nearest Neighbors graph bolstered the performance of the compressed embeddings to be comparable to the original BERT embeddings. On one dataset, we find that despite a 12-fold dimensionality reduction, the compressed embeddings performed within 0.1% of the original BERT embeddings on a downstream classification task. In addition, we find that this approach works particularly well on tasks reliant on syntactic data, when compared with linear dimensionality reduction. These results show promise for a novel geometric approach to achieve lower dimensional text embeddings from existing transformers and pave the way for data-specific and application-specific embedding compressions.
We introduce a theoretical framework for understanding and predicting the complexity of sequence classification tasks, using a novel extension of the theory of Boolean function sensitivity. The sensitivity of a function, given a distribution over input sequences, quantifies the number of disjoint subsets of the input sequence that can each be individually changed to change the output. We argue that standard sequence classification methods are biased towards learning low-sensitivity functions, so that tasks requiring high sensitivity are more difficult. To that end, we show analytically that simple lexical classifiers can only express functions of bounded sensitivity, and we show empirically that low-sensitivity functions are easier to learn for LSTMs. We then estimate sensitivity on 15 NLP tasks, finding that sensitivity is higher on challenging tasks collected in GLUE than on simple text classification tasks, and that sensitivity predicts the performance both of simple lexical classifiers and of vanilla BiLSTMs without pretrained contextualized embeddings. Within a task, sensitivity predicts which inputs are hard for such simple models. Our results suggest that the success of massively pretrained contextual representations stems in part because they provide representations from which information can be extracted by low-sensitivity decoders.
Weakly-supervised text classification aims to induce text classifiers from only a few user-provided seed words. The vast majority of previous work assumes high-quality seed words are given. However, the expert-annotated seed words are sometimes non-trivial to come up with. Furthermore, in the weakly-supervised learning setting, we do not have any labeled document to measure the seed words' efficacy, making the seed word selection process "a walk in the dark". In this work, we remove the need for expert-curated seed words by first mining (noisy) candidate seed words associated with the category names. We then train interim models with individual candidate seed words. Lastly, we estimate the interim models' error rate in an unsupervised manner. The seed words that yield the lowest estimated error rates are added to the final seed word set. A comprehensive evaluation of six binary classification tasks on four popular datasets demonstrates that the proposed method outperforms a baseline using only category name seed words and obtained comparable performance as a counterpart using expert-annotated seed words.
Position encoding recently has shown effective in the transformer architecture. It enables valuable supervision for dependency modeling between elements at different positions of the sequence. In this paper, we first investigate various methods to integrate positional information into the learning process of transformer-based language models. Then, we propose a novel method named Rotary Position Embedding(RoPE) to effectively leverage the positional information. Specifically, the proposed RoPE encodes the absolute position with a rotation matrix and meanwhile incorporates the explicit relative position dependency in self-attention formulation. Notably, RoPE enables valuable properties, including the flexibility of sequence length, decaying inter-token dependency with increasing relative distances, and the capability of equipping the linear self-attention with relative position encoding. Finally, we evaluate the enhanced transformer with rotary position embedding, also called RoFormer, on various long text classification benchmark datasets. Our experiments show that it consistently overcomes its alternatives. Furthermore, we provide a theoretical analysis to explain some experimental results. RoFormer is already integrated into Huggingface: \url{https://huggingface.co/docs/transformers/model_doc/roformer}.