Black-box deep neural networks excel in text classification, yet their application in high-stakes domains is hindered by their lack of interpretability. To address this, we propose Text Bottleneck Models (TBM), an intrinsically interpretable text classification framework that offers both global and local explanations. Rather than directly predicting the output label, TBM predicts categorical values for a sparse set of salient concepts and uses a linear layer over those concept values to produce the final prediction. These concepts can be automatically discovered and measured by a Large Language Model (LLM) without the need for human curation. Experiments on 12 diverse text understanding datasets demonstrate that TBM can rival the performance of black-box baselines such as few-shot GPT-4 and finetuned DeBERTa while falling short against finetuned GPT-3.5. Comprehensive human evaluation validates that TBM can generate high-quality concepts relevant to the task, and the concept measurement aligns well with human judgments, suggesting that the predictions made by TBMs are interpretable. Overall, our findings suggest that TBM is a promising new framework that enhances interpretability with minimal performance tradeoffs.
Recursion is a prominent feature of human language, and fundamentally challenging for self-attention due to the lack of an explicit recursive-state tracking mechanism. Consequently, Transformer language models poorly capture long-tail recursive structure and exhibit sample-inefficient syntactic generalization. This work introduces Pushdown Layers, a new self-attention layer that models recursive state via a stack tape that tracks estimated depths of every token in an incremental parse of the observed prefix. Transformer LMs with Pushdown Layers are syntactic language models that autoregressively and synchronously update this stack tape as they predict new tokens, in turn using the stack tape to softly modulate attention over tokens -- for instance, learning to "skip" over closed constituents. When trained on a corpus of strings annotated with silver constituency parses, Transformers equipped with Pushdown Layers achieve dramatically better and 3-5x more sample-efficient syntactic generalization, while maintaining similar perplexities. Pushdown Layers are a drop-in replacement for standard self-attention. We illustrate this by finetuning GPT2-medium with Pushdown Layers on an automatically parsed WikiText-103, leading to improvements on several GLUE text classification tasks.
Language Representation Models (LRMs) trained with real-world data may capture and exacerbate undesired bias and cause unfair treatment of people in various demographic groups. Several techniques have been investigated for applying interventions to LRMs to remove bias in benchmark evaluations on, for example, word embeddings. However, the negative side effects of debiasing interventions are usually not revealed in the downstream tasks. We propose xGAP-DEBIAS, a set of evaluations on assessing the fairness of debiasing. In this work, We examine four debiasing techniques on a real-world text classification task and show that reducing biasing is at the cost of degrading performance for all demographic groups, including those the debiasing techniques aim to protect. We advocate that a debiasing technique should have good downstream performance with the constraint of ensuring no harm to the protected group.
We propose a general class of sample based explanations of machine learning models, which we term generalized representers. To measure the effect of a training sample on a model's test prediction, generalized representers use two components: a global sample importance that quantifies the importance of the training point to the model and is invariant to test samples, and a local sample importance that measures similarity between the training sample and the test point with a kernel. A key contribution of the paper is to show that generalized representers are the only class of sample based explanations satisfying a natural set of axiomatic properties. We discuss approaches to extract global importances given a kernel, and also natural choices of kernels given modern non-linear models. As we show, many popular existing sample based explanations could be cast as generalized representers with particular choices of kernels and approaches to extract global importances. Additionally, we conduct empirical comparisons of different generalized representers on two image and two text classification datasets.
As one of the most popular neural network modules, Transformer plays a central role in many fundamental deep learning models, e.g., the ViT in computer vision and the BERT and GPT in natural language processing. The effectiveness of the Transformer is often attributed to its multi-head attention (MHA) mechanism. In this study, we discuss the limitations of MHA, including the high computational complexity due to its ``query-key-value'' architecture and the numerical issue caused by its softmax operation. Considering the above problems and the recent development tendency of the attention layer, we propose an effective and efficient surrogate of the Transformer, called Sliceformer. Our Sliceformer replaces the classic MHA mechanism with an extremely simple ``slicing-sorting'' operation, i.e., projecting inputs linearly to a latent space and sorting them along different feature dimensions (or equivalently, called channels). For each feature dimension, the sorting operation implicitly generates an implicit attention map with sparse, full-rank, and doubly-stochastic structures. We consider different implementations of the slicing-sorting operation and analyze their impacts on the Sliceformer. We test the Sliceformer in the Long-Range Arena benchmark, image classification, text classification, and molecular property prediction, demonstrating its advantage in computational complexity and universal effectiveness in discriminative tasks. Our Sliceformer achieves comparable or better performance with lower memory cost and faster speed than the Transformer and its variants. Moreover, the experimental results reveal that applying our Sliceformer can empirically suppress the risk of mode collapse when representing data. The code is available at \url{https://github.com/SDS-Lab/sliceformer}.
Reproducibility in scientific work has been becoming increasingly important in research communities such as machine learning, natural language processing, and computer vision communities due to the rapid development of the research domains supported by recent advances in deep learning. In this work, we present a significantly upgraded version of torchdistill, a modular-driven coding-free deep learning framework significantly upgraded from the initial release, which supports only image classification and object detection tasks for reproducible knowledge distillation experiments. To demonstrate that the upgraded framework can support more tasks with third-party libraries, we reproduce the GLUE benchmark results of BERT models using a script based on the upgraded torchdistill, harmonizing with various Hugging Face libraries. All the 27 fine-tuned BERT models and configurations to reproduce the results are published at Hugging Face, and the model weights have already been widely used in research communities. We also reimplement popular small-sized models and new knowledge distillation methods and perform additional experiments for computer vision tasks.
In order to build reliable and trustworthy NLP applications, models need to be both fair across different demographics and explainable. Usually these two objectives, fairness and explainability, are optimized and/or examined independently of each other. Instead, we argue that forthcoming, trustworthy NLP systems should consider both. In this work, we perform a first study to understand how they influence each other: do fair(er) models rely on more plausible rationales? and vice versa. To this end, we conduct experiments on two English multi-class text classification datasets, BIOS and ECtHR, that provide information on gender and nationality, respectively, as well as human-annotated rationales. We fine-tune pre-trained language models with several methods for (i) bias mitigation, which aims to improve fairness; (ii) rationale extraction, which aims to produce plausible explanations. We find that bias mitigation algorithms do not always lead to fairer models. Moreover, we discover that empirical fairness and explainability are orthogonal.
Choquet integral is a widely used aggregation oper- ator on one-dimensional and interval-valued information, since it is able to take into account the possible interaction among data. However, there are many cases where the information taken into account is vectorial, such as Long Short-Term Memories (LSTM). LSTM units are a kind of Recurrent Neural Networks that have become one of the most powerful tools to deal with sequential information since they have the power of controlling the information flow. In this paper, we first generalize the standard Choquet integral to admit an input composed by n- dimensional vectors, which produces an n-dimensional vector output. We study several properties and construction methods of vector Choquet integrals. Then, we use this integral in the place of the summation operator, introducing in this way the new VCI- LSTM architecture. Finally, we use the proposed VCI-LSTM to deal with two problems: sequential image classification and text classification.
Similar case retrieval (SCR) is a representative legal AI application that plays a pivotal role in promoting judicial fairness. However, existing SCR datasets only focus on the fact description section when judging the similarity between cases, ignoring other valuable sections (e.g., the court's opinion) that can provide insightful reasoning process behind. Furthermore, the case similarities are typically measured solely by the textual semantics of the fact descriptions, which may fail to capture the full complexity of legal cases from the perspective of legal knowledge. In this work, we present MUSER, a similar case retrieval dataset based on multi-view similarity measurement and comprehensive legal element with sentence-level legal element annotations. Specifically, we select three perspectives (legal fact, dispute focus, and law statutory) and build a comprehensive and structured label schema of legal elements for each of them, to enable accurate and knowledgeable evaluation of case similarities. The constructed dataset originates from Chinese civil cases and contains 100 query cases and 4,024 candidate cases. We implement several text classification algorithms for legal element prediction and various retrieval methods for retrieving similar cases on MUSER. The experimental results indicate that incorporating legal elements can benefit the performance of SCR models, but further efforts are still required to address the remaining challenges posed by MUSER. The source code and dataset are released at https://github.com/THUlawtech/MUSER.
Although large language models (LLMs) have apparently acquired a certain level of grammatical knowledge and the ability to make generalizations, they fail to interpret negation, a crucial step in Natural Language Processing. We try to clarify the reasons for the sub-optimal performance of LLMs understanding negation. We introduce a large semi-automatically generated dataset of circa 400,000 descriptive sentences about commonsense knowledge that can be true or false in which negation is present in about 2/3 of the corpus in different forms. We have used our dataset with the largest available open LLMs in a zero-shot approach to grasp their generalization and inference capability and we have also fine-tuned some of the models to assess whether the understanding of negation can be trained. Our findings show that, while LLMs are proficient at classifying affirmative sentences, they struggle with negative sentences and lack a deep understanding of negation, often relying on superficial cues. Although fine-tuning the models on negative sentences improves their performance, the lack of generalization in handling negation is persistent, highlighting the ongoing challenges of LLMs regarding negation understanding and generalization. The dataset and code are publicly available.
Security classifiers, designed to detect malicious content in computer systems and communications, can underperform when provided with insufficient training data. In the security domain, it is often easy to find samples of the negative (benign) class, and challenging to find enough samples of the positive (malicious) class to train an effective classifier. This study evaluates the application of natural language text generators to fill this data gap in multiple security-related text classification tasks. We describe a variety of previously-unexamined language-model fine-tuning approaches for this purpose and consider in particular the impact of disproportionate class-imbalances in the training set. Across our evaluation using three state-of-the-art classifiers designed for offensive language detection, review fraud detection, and SMS spam detection, we find that models trained with GPT-3 data augmentation strategies outperform both models trained without augmentation and models trained using basic data augmentation strategies already in common usage. In particular, we find substantial benefits for GPT-3 data augmentation strategies in situations with severe limitations on known positive-class samples.
Biomedical language understanding benchmarks are the driving forces for artificial intelligence applications with large language model (LLM) back-ends. However, most current benchmarks: (a) are limited to English which makes it challenging to replicate many of the successes in English for other languages, or (b) focus on knowledge probing of LLMs and neglect to evaluate how LLMs apply these knowledge to perform on a wide range of bio-medical tasks, or (c) have become a publicly available corpus and are leaked to LLMs during pre-training. To facilitate the research in medical LLMs, we re-build the Chinese Biomedical Language Understanding Evaluation (CBLUE) benchmark into a large scale prompt-tuning benchmark, PromptCBLUE. Our benchmark is a suitable test-bed and an online platform for evaluating Chinese LLMs' multi-task capabilities on a wide range bio-medical tasks including medical entity recognition, medical text classification, medical natural language inference, medical dialogue understanding and medical content/dialogue generation. To establish evaluation on these tasks, we have experimented and report the results with the current 9 Chinese LLMs fine-tuned with differtent fine-tuning techniques.
The popularity of transformer-based text embeddings calls for better statistical tools for measuring distributions of such embeddings. One such tool would be a method for ranking texts within a corpus by centrality, i.e. assigning each text a number signifying how representative that text is of the corpus as a whole. However, an intrinsic center-outward ordering of high-dimensional text representations is not trivial. A statistical depth is a function for ranking k-dimensional objects by measuring centrality with respect to some observed k-dimensional distribution. We adopt a statistical depth to measure distributions of transformer-based text embeddings, transformer-based text embedding (TTE) depth, and introduce the practical use of this depth for both modeling and distributional inference in NLP pipelines. We first define TTE depth and an associated rank sum test for determining whether two corpora differ significantly in embedding space. We then use TTE depth for the task of in-context learning prompt selection, showing that this approach reliably improves performance over statistical baseline approaches across six text classification tasks. Finally, we use TTE depth and the associated rank sum test to characterize the distributions of synthesized and human-generated corpora, showing that five recent synthetic data augmentation processes cause a measurable distributional shift away from associated human-generated text.
Multi-label text classification is a critical task in the industry. It helps to extract structured information from large amount of textual data. We propose Text to Topic (Text2Topic), which achieves high multi-label classification performance by employing a Bi-Encoder Transformer architecture that utilizes concatenation, subtraction, and multiplication of embeddings on both text and topic. Text2Topic also supports zero-shot predictions, produces domain-specific text embeddings, and enables production-scale batch-inference with high throughput. The final model achieves accurate and comprehensive results compared to state-of-the-art baselines, including large language models (LLMs). In this study, a total of 239 topics are defined, and around 1.6 million text-topic pairs annotations (in which 200K are positive) are collected on approximately 120K texts from 3 main data sources on Booking.com. The data is collected with optimized smart sampling and partial labeling. The final Text2Topic model is deployed on a real-world stream processing platform, and it outperforms other models with 92.9% micro mAP, as well as a 75.8% macro mAP score. We summarize the modeling choices which are extensively tested through ablation studies, and share detailed in-production decision-making steps.
Data augmentation is a widely used technique to address the problem of text classification when there is a limited amount of training data. Recent work often tackles this problem using large language models (LLMs) like GPT3 that can generate new examples given already available ones. In this work, we propose a method to generate more helpful augmented data by utilizing the LLM's abilities to follow instructions and perform few-shot classifications. Our specific PromptMix method consists of two steps: 1) generate challenging text augmentations near class boundaries; however, generating borderline examples increases the risk of false positives in the dataset, so we 2) relabel the text augmentations using a prompting-based LLM classifier to enhance the correctness of labels in the generated data. We evaluate the proposed method in challenging 2-shot and zero-shot settings on four text classification datasets: Banking77, TREC6, Subjectivity (SUBJ), and Twitter Complaints. Our experiments show that generating and, crucially, relabeling borderline examples facilitates the transfer of knowledge of a massive LLM like GPT3.5-turbo into smaller and cheaper classifiers like DistilBERT$_{base}$ and BERT$_{base}$. Furthermore, 2-shot PromptMix outperforms multiple 5-shot data augmentation methods on the four datasets. Our code is available at https://github.com/ServiceNow/PromptMix-EMNLP-2023.
Semi-supervised text classification (SSTC) has gained increasing attention due to its ability to leverage unlabeled data. However, existing approaches based on pseudo-labeling suffer from the issues of pseudo-label bias and error accumulation. In this paper, we propose JointMatch, a holistic approach for SSTC that addresses these challenges by unifying ideas from recent semi-supervised learning and the task of learning with noise. JointMatch adaptively adjusts classwise thresholds based on the learning status of different classes to mitigate model bias towards current easy classes. Additionally, JointMatch alleviates error accumulation by utilizing two differently initialized networks to teach each other in a cross-labeling manner. To maintain divergence between the two networks for mutual learning, we introduce a strategy that weighs more disagreement data while also allowing the utilization of high-quality agreement data for training. Experimental results on benchmark datasets demonstrate the superior performance of JointMatch, achieving a significant 5.13% improvement on average. Notably, JointMatch delivers impressive results even in the extremely-scarce-label setting, obtaining 86% accuracy on AG News with only 5 labels per class. We make our code available at https://github.com/HenryPengZou/JointMatch.
We propose the application of Transformer-based language models for classifying entity legal forms from raw legal entity names. Specifically, we employ various BERT variants and compare their performance against multiple traditional baselines. Our evaluation encompasses a substantial subset of freely available Legal Entity Identifier (LEI) data, comprising over 1.1 million legal entities from 30 different legal jurisdictions. The ground truth labels for classification per jurisdiction are taken from the Entity Legal Form (ELF) code standard (ISO 20275). Our findings demonstrate that pre-trained BERT variants outperform traditional text classification approaches in terms of F1 score, while also performing comparably well in the Macro F1 Score. Moreover, the validity of our proposal is supported by the outcome of third-party expert reviews conducted in ten selected jurisdictions. This study highlights the significant potential of Transformer-based models in advancing data standardization and data integration. The presented approaches can greatly benefit financial institutions, corporations, governments and other organizations in assessing business relationships, understanding risk exposure, and promoting effective governance.
Zero-shot classification enables text to be classified into classes not seen during training. In this study, we examine the efficacy of zero-shot learning models in classifying healthcare consultation responses from Doctors and AI systems. The models evaluated include BART, BERT, XLM, XLM-R and DistilBERT. The models were tested on three different datasets based on a binary and multi-label analysis to identify the origins of text in health consultations without any prior corpus training. According to our findings, the zero-shot language models show a good understanding of language generally, but has limitations when trying to classify doctor and AI responses to healthcare consultations. This research provides a foundation for future research in the field of medical text classification by informing the development of more accurate methods of classifying text written by Doctors and AI systems in health consultations.
Prompt-based learning has shown its effectiveness in few-shot text classification. One important factor in its success is a verbalizer, which translates output from a language model into a predicted class. Notably, the simplest and widely acknowledged verbalizer employs manual labels to represent the classes. However, manual selection does not guarantee the optimality of the selected words when conditioned on the chosen language model. Therefore, we propose Label-Aware Automatic Verbalizer (LAAV), effectively augmenting the manual labels to achieve better few-shot classification results. Specifically, we use the manual labels along with the conjunction "and" to induce the model to generate more effective words for the verbalizer. The experimental results on five datasets across five languages demonstrate that LAAV significantly outperforms existing verbalizers. Furthermore, our analysis reveals that LAAV suggests more relevant words compared to similar approaches, especially in mid-to-low resource languages.
Contrastive explanations, where one decision is explained in contrast to another, are supposed to be closer to how humans explain a decision than non-contrastive explanations, where the decision is not necessarily referenced to an alternative. This claim has never been empirically validated. We analyze four English text-classification datasets (SST2, DynaSent, BIOS and DBpedia-Animals). We fine-tune and extract explanations from three different models (RoBERTa, GTP-2, and T5), each in three different sizes and apply three post-hoc explainability methods (LRP, GradientxInput, GradNorm). We furthermore collect and release human rationale annotations for a subset of 100 samples from the BIOS dataset for contrastive and non-contrastive settings. A cross-comparison between model-based rationales and human annotations, both in contrastive and non-contrastive settings, yields a high agreement between the two settings for models as well as for humans. Moreover, model-based explanations computed in both settings align equally well with human rationales. Thus, we empirically find that humans do not necessarily explain in a contrastive manner.9 pages, long paper at ACL 2022 proceedings.