The Transformer is an extremely powerful and prominent deep learning architecture. In this work, we challenge the commonly held belief in deep learning that going deeper is better, and show an alternative design approach that is building wider attention Transformers. We demonstrate that wide single layer Transformer models can compete with or outperform deeper ones in a variety of Natural Language Processing (NLP) tasks when both are trained from scratch. The impact of changing the model aspect ratio on Transformers is then studied systematically. This ratio balances the number of layers and the number of attention heads per layer while keeping the total number of attention heads and all other hyperparameters constant. On average, across 4 NLP tasks and 10 attention types, single layer wide models perform 0.3% better than their deep counterparts. We show an in-depth evaluation and demonstrate how wide models require a far smaller memory footprint and can run faster on commodity hardware, in addition, these wider models are also more interpretable. For example, a single layer Transformer on the IMDb byte level text classification has 3.1x faster inference latency on a CPU than its equally accurate deeper counterpart, and is half the size. We therefore put forward wider and shallower models as a viable and desirable alternative for small models on NLP tasks, and as an important area of research for domains beyond this.
Lifelong language learning seeks to have models continuously learn multiple tasks in a sequential order without suffering from catastrophic forgetting. State-of-the-art approaches rely on sparse experience replay as the primary approach to prevent forgetting. Experience replay usually adopts sampling methods for the memory population; however, the effect of the chosen sampling strategy on model performance has not yet been studied. In this paper, we investigate how relevant the selective memory population is in the lifelong learning process of text classification and question-answering tasks. We found that methods that randomly store a uniform number of samples from the entire data stream lead to high performances, especially for low memory size, which is consistent with computer vision studies.
Current methods for prompt learning in zeroshot scenarios widely rely on a development set with sufficient human-annotated data to select the best-performing prompt template a posteriori. This is not ideal because in a realworld zero-shot scenario of practical relevance, no labelled data is available. Thus, we propose a simple yet effective method for screening reasonable prompt templates in zero-shot text classification: Perplexity Selection (Perplection). We hypothesize that language discrepancy can be used to measure the efficacy of prompt templates, and thereby develop a substantiated perplexity-based scheme allowing for forecasting the performance of prompt templates in advance. Experiments show that our method leads to improved prediction performance in a realistic zero-shot setting, eliminating the need for any labelled examples.
In this paper, we explore how to utilize pre-trained language model to perform few-shot text classification where only a few annotated examples are given for each class. Since using traditional cross-entropy loss to fine-tune language model under this scenario causes serious overfitting and leads to sub-optimal generalization of model, we adopt supervised contrastive learning on few labeled data and consistency-regularization on vast unlabeled data. Moreover, we propose a novel contrastive consistency to further boost model performance and refine sentence representation. After conducting extensive experiments on four datasets, we demonstrate that our model (FTCC) can outperform state-of-the-art methods and has better robustness.
As free online encyclopedias with massive volumes of content, Wikipedia and Wikidata are key to many Natural Language Processing (NLP) tasks, such as information retrieval, knowledge base building, machine translation, text classification, and text summarization. In this paper, we introduce WikiDes, a novel dataset to generate short descriptions of Wikipedia articles for the problem of text summarization. The dataset consists of over 80k English samples on 6987 topics. We set up a two-phase summarization method - description generation (Phase I) and candidate ranking (Phase II) - as a strong approach that relies on transfer and contrastive learning. For description generation, T5 and BART show their superiority compared to other small-scale pre-trained models. By applying contrastive learning with the diverse input from beam search, the metric fusion-based ranking models outperform the direct description generation models significantly up to 22 ROUGE in topic-exclusive split and topic-independent split. Furthermore, the outcome descriptions in Phase II are supported by human evaluation in over 45.33% chosen compared to 23.66% in Phase I against the gold descriptions. In the aspect of sentiment analysis, the generated descriptions cannot effectively capture all sentiment polarities from paragraphs while doing this task better from the gold descriptions. The automatic generation of new descriptions reduces the human efforts in creating them and enriches Wikidata-based knowledge graphs. Our paper shows a practical impact on Wikipedia and Wikidata since there are thousands of missing descriptions. Finally, we expect WikiDes to be a useful dataset for related works in capturing salient information from short paragraphs. The curated dataset is publicly available at: https://github.com/declare-lab/WikiDes.
Unsupervised text classification, with its most common form being sentiment analysis, used to be performed by counting words in a text that were stored in a lexicon, which assigns each word to one class or as a neutral word. In recent years, these lexicon-based methods fell out of favor and were replaced by computationally demanding fine-tuning techniques for encoder-only models such as BERT and zero-shot classification using decoder-only models such as GPT-4. In this paper, we propose an alternative approach: Lex2Sent, which provides improvement over classic lexicon methods but does not require any GPU or external hardware. To classify texts, we train embedding models to determine the distances between document embeddings and the embeddings of the parts of a suitable lexicon. We employ resampling, which results in a bagging effect, boosting the performance of the classification. We show that our model outperforms lexica and provides a basis for a high performing few-shot fine-tuning approach in the task of binary sentiment analysis.
Language models are pre-trained using large corpora of generic data like book corpus, common crawl and Wikipedia, which is essential for the model to understand the linguistic characteristics of the language. New studies suggest using Domain Adaptive Pre-training (DAPT) and Task-Adaptive Pre-training (TAPT) as an intermediate step before the final finetuning task. This step helps cover the target domain vocabulary and improves the model performance on the downstream task. In this work, we study the impact of training only the embedding layer on the model's performance during TAPT and task-specific finetuning. Based on our study, we propose a simple approach to make the intermediate step of TAPT for BERT-based models more efficient by performing selective pre-training of BERT layers. We show that training only the BERT embedding layer during TAPT is sufficient to adapt to the vocabulary of the target domain and achieve comparable performance. Our approach is computationally efficient, with 78\% fewer parameters trained during TAPT. The proposed embedding layer finetuning approach can also be an efficient domain adaptation technique.
When dealing with clinical text classification on a small dataset recent studies have confirmed that a well-tuned multilayer perceptron outperforms other generative classifiers, including deep learning ones. To increase the performance of the neural network classifier, feature selection for the learning representation can effectively be used. However, most feature selection methods only estimate the degree of linear dependency between variables and select the best features based on univariate statistical tests. Furthermore, the sparsity of the feature space involved in the learning representation is ignored. Goal: Our aim is therefore to access an alternative approach to tackle the sparsity by compressing the clinical representation feature space, where limited French clinical notes can also be dealt with effectively. Methods: This study proposed an autoencoder learning algorithm to take advantage of sparsity reduction in clinical note representation. The motivation was to determine how to compress sparse, high-dimensional data by reducing the dimension of the clinical note representation feature space. The classification performance of the classifiers was then evaluated in the trained and compressed feature space. Results: The proposed approach provided overall performance gains of up to 3% for each evaluation. Finally, the classifier achieved a 92% accuracy, 91% recall, 91% precision, and 91% f1-score in detecting the patient's condition. Furthermore, the compression working mechanism and the autoencoder prediction process were demonstrated by applying the theoretic information bottleneck framework.
Recent large language models (LLMs) have demonstrated remarkable prediction performance for a growing array of tasks. However, their proliferation into high-stakes domains (e.g. medicine) and compute-limited settings has created a burgeoning need for interpretability and efficiency. We address this need by proposing Augmented Interpretable Models (Aug-imodels), a framework for leveraging the knowledge learned by LLMs to build extremely efficient and interpretable models. Aug-imodels use LLMs during fitting but not during inference, allowing complete transparency and often a speed/memory improvement of greater than 1,000x for inference compared to LLMs. We explore two instantiations of Aug-imodels in natural-language processing: (i) Aug-GAM, which augments a generalized additive model with decoupled embeddings from an LLM and (ii) Aug-Tree, which augments a decision tree with LLM feature expansions. Across a variety of text-classification datasets, both outperform their non-augmented counterparts. Aug-GAM can even outperform much larger models (e.g. a 6-billion parameter GPT-J model), despite having 10,000x fewer parameters and being fully transparent. We further explore Aug-imodels in a natural-language fMRI study, where they generate interesting interpretations from scientific data. All code for using Aug-imodels and reproducing results is made available on Github.
Current text classification methods typically encode the text merely into embedding before a naive or complicated classifier, which ignores the suggestive information contained in the label text. As a matter of fact, humans classify documents primarily based on the semantic meaning of the subcategories. We propose a novel model structure via siamese BERT and interactive double attentions named IDEA ( Interactive DoublE Attentions) to capture the information exchange of text and label names. Interactive double attentions enable the model to exploit the inter-class and intra-class information from coarse to fine, which involves distinguishing among all labels and matching the semantical subclasses of ground truth labels. Our proposed method outperforms the state-of-the-art methods using label texts significantly with more stable results.
Transformers are widely used in natural language processing, where they consistently achieve state-of-the-art performance. This is mainly due to their attention-based architecture, which allows them to model rich linguistic relations between (sub)words. However, transformers are difficult to interpret. Being able to provide reasoning for its decisions is an important property for a model in domains where human lives are affected. With transformers finding wide use in such fields, the need for interpretability techniques tailored to them arises. We propose a new technique that selects the most faithful attention-based interpretation among the several ones that can be obtained by combining different head, layer and matrix operations. In addition, two variations are introduced towards (i) reducing the computational complexity, thus being faster and friendlier to the environment, and (ii) enhancing the performance in multi-label data. We further propose a new faithfulness metric that is more suitable for transformer models and exhibits high correlation with the area under the precision-recall curve based on ground truth rationales. We validate the utility of our contributions with a series of quantitative and qualitative experiments on seven datasets.
Text classification has become widely used in various natural language processing applications like sentiment analysis. Current applications often use large transformer-based language models to classify input texts. However, there is a lack of systematic study on how much private information can be inverted when publishing models. In this paper, we formulate \emph{Text Revealer} -- the first model inversion attack for text reconstruction against text classification with transformers. Our attacks faithfully reconstruct private texts included in training data with access to the target model. We leverage an external dataset and GPT-2 to generate the target domain-like fluent text, and then perturb its hidden state optimally with the feedback from the target model. Our extensive experiments demonstrate that our attacks are effective for datasets with different text lengths and can reconstruct private texts with accuracy.
Text classification is a typical natural language processing or computational linguistics task with various interesting applications. As the number of users on social media platforms increases, data acceleration promotes emerging studies on Social Media Text Classification (SMTC) or social media text mining on these valuable resources. In contrast to English, Vietnamese, one of the low-resource languages, is still not concentrated on and exploited thoroughly. Inspired by the success of the GLUE, we introduce the Social Media Text Classification Evaluation (SMTCE) benchmark, as a collection of datasets and models across a diverse set of SMTC tasks. With the proposed benchmark, we implement and analyze the effectiveness of a variety of multilingual BERT-based models (mBERT, XLM-R, and DistilmBERT) and monolingual BERT-based models (PhoBERT, viBERT, vELECTRA, and viBERT4news) for tasks in the SMTCE benchmark. Monolingual models outperform multilingual models and achieve state-of-the-art results on all text classification tasks. It provides an objective assessment of multilingual and monolingual BERT-based models on the benchmark, which will benefit future studies about BERTology in the Vietnamese language.
Prompt learning's fine-tune performance on text classification tasks has attracted the NLP community. This paper applies it to resume information extraction, improving existing methods for this task. We created manual templates and verbalizers tailored to resume texts and compared the performance of Masked Language Model (MLM) and Seq2Seq PLMs. Also, we enhanced the verbalizer design for Knowledgeable Prompt-tuning, contributing to prompt template design across NLP tasks. We present the Manual Knowledgeable Verbalizer (MKV), a rule for constructing verbalizers for specific applications. Our tests show that MKV rules yield more effective, robust templates and verbalizers than existing methods. Our MKV approach resolved sample imbalance, surpassing current automatic prompt methods. This study underscores the value of tailored prompt learning for resume extraction, stressing the importance of custom-designed templates and verbalizers.
Current graph representation learning techniques use Graph Neural Networks (GNNs) to extract features from dataset embeddings. In this work, we examine the quality of these embeddings and assess how changing them can affect the accuracy of GNNs. We explore different embedding extraction techniques for both images and texts; and find that the performance of different GNN architectures is dependent on the embedding style used. We see a prevalence of bag of words (BoW) embeddings and text classification tasks in available graph datasets. Given the impact embeddings has on GNN performance. this leads to a phenomenon that GNNs being optimised for BoW vectors.
The key to the text classification task is language representation and important information extraction, and there are many related studies. In recent years, the research on graph neural network (GNN) in text classification has gradually emerged and shown its advantages, but the existing models mainly focus on directly inputting words as graph nodes into the GNN models ignoring the different levels of semantic structure information in the samples. To address the issue, we propose a new hierarchical graph neural network (HieGNN) which extracts corresponding information from word-level, sentence-level and document-level respectively. Experimental results on several benchmark datasets achieve better or similar results compared to several baseline methods, which demonstrate that our model is able to obtain more useful information for classification from samples.
Large Language Models have demonstrated remarkable few-shot performance, but the performance can be sensitive to the selection of few-shot instances. We propose PATRON, a new method that uses prompt-based uncertainty estimation for data selection for pre-trained language model fine-tuning under cold-start scenarios, i.e., no initial labeled data are available. In PATRON, we design (1) a prompt-based uncertainty propagation approach to estimate the importance of data points and (2) a partition-then-rewrite (PTR) strategy to promote sample diversity when querying for annotations. Experiments on six text classification datasets show that PATRON outperforms the strongest cold-start data selection baselines by up to 6.9%. Besides, with 128 labels only, PATRON achieves 91.0% and 92.1% of the fully supervised performance based on vanilla fine-tuning and prompt-based learning respectively. Our implementation of PATRON is available at \url{https://github.com/yueyu1030/Patron}.
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
BERT is a widely used pre-trained model in natural language processing. However, since BERT is quadratic to the text length, the BERT model is difficult to be used directly on the long-text corpus. In some fields, the collected text data may be quite long, such as in the health care field. Therefore, to apply the pre-trained language knowledge of BERT to long text, in this paper, imitating the skimming-intensive reading method used by humans when reading a long paragraph, the Skimming-Intensive Model (SkIn) is proposed. It can dynamically select the critical information in the text so that the sentence input into the BERT-Base model is significantly shortened, which can effectively save the cost of the classification algorithm. Experiments show that the SkIn method has achieved superior accuracy than the baselines on long-text classification datasets in the medical field, while its time and space requirements increase linearly with the text length, alleviating the time and space overflow problem of basic BERT on long-text data.
This paper analyses the degree to which dialect classifiers based on syntactic representations remain stable over space and time. While previous work has shown that the combination of grammar induction and geospatial text classification produces robust dialect models, we do not know what influence both changing grammars and changing populations have on dialect models. This paper constructs a test set for 12 dialects of English that spans three years at monthly intervals with a fixed spatial distribution across 1,120 cities. Syntactic representations are formulated within the usage-based Construction Grammar paradigm (CxG). The decay rate of classification performance for each dialect over time allows us to identify regions undergoing syntactic change. And the distribution of classification accuracy within dialect regions allows us to identify the degree to which the grammar of a dialect is internally heterogeneous. The main contribution of this paper is to show that a rigorous evaluation of dialect classification models can be used to find both variation over space and change over time.