This paper addresses the quality issues in existing Twitter-based paraphrase datasets, and discusses the necessity of using two separate definitions of paraphrase for identification and generation tasks. We present a new Multi-Topic Paraphrase in Twitter (MultiPIT) corpus that consists of a total of 130k sentence pairs with crowdsoursing (MultiPIT_crowd) and expert (MultiPIT_expert) annotations using two different paraphrase definitions for paraphrase identification, in addition to a multi-reference test set (MultiPIT_NMR) and a large automatically constructed training set (MultiPIT_Auto) for paraphrase generation. With improved data annotation quality and task-specific paraphrase definition, the best pre-trained language model fine-tuned on our dataset achieves the state-of-the-art performance of 84.2 F1 for automatic paraphrase identification. Furthermore, our empirical results also demonstrate that the paraphrase generation models trained on MultiPIT_Auto generate more diverse and high-quality paraphrases compared to their counterparts fine-tuned on other corpora such as Quora, MSCOCO, and ParaNMT.
Paraphrase Identification is a fundamental task in Natural Language Processing. While much progress has been made in the field, the performance of many state-of-the-art models often suffer from distribution shift during inference time. We verify that a major source of this performance drop comes from biases introduced by negative examples. To overcome these biases, we propose in this paper to train two separate models, one that only utilizes the positive pairs and the other the negative pairs. This enables us the option of deciding how much to utilize the negative model, for which we introduce a perplexity based out-of-distribution metric that we show can effectively and automatically determine how much weight it should be given during inference. We support our findings with strong empirical results.
This dissertation explores the linguistic and computational aspects of the meaning relations that can hold between two or more complex linguistic expressions (phrases, clauses, sentences, paragraphs). In particular, it focuses on Paraphrasing, Textual Entailment, Contradiction, and Semantic Similarity. In Part I: "Similarity at the Level of Words and Phrases", I study the Distributional Hypothesis (DH) and explore several different methodologies for quantifying semantic similarity at the levels of words and short phrases. In Part II: "Paraphrase Typology and Paraphrase Identification", I focus on the meaning relation of paraphrasing and the empirical task of automated Paraphrase Identification (PI). In Part III: "Paraphrasing, Textual Entailment, and Semantic Similarity", I present a novel direction in the research on textual meaning relations, resulting from joint research carried out on on paraphrasing, textual entailment, contradiction, and semantic similarity.
Deep neural models (e.g. Transformer) naturally learn spurious features, which create a ``shortcut'' between the labels and inputs, thus impairing the generalization and robustness. This paper advances the self-attention mechanism to its robust variant for Transformer-based pre-trained language models (e.g. BERT). We propose \textit{Adversarial Self-Attention} mechanism (ASA), which adversarially biases the attentions to effectively suppress the model reliance on features (e.g. specific keywords) and encourage its exploration of broader semantics. We conduct a comprehensive evaluation across a wide range of tasks for both pre-training and fine-tuning stages. For pre-training, ASA unfolds remarkable performance gains compared to naive training for longer steps. For fine-tuning, ASA-empowered models outweigh naive models by a large margin considering both generalization and robustness.
Being able to rank the similarity of short text segments is an interesting bonus feature of neural machine translation. Translation-based similarity measures include direct and pivot translation probability, as well as translation cross-likelihood, which has not been studied so far. We analyze these measures in the common framework of multilingual NMT, releasing the NMTScore library (available at https://github.com/ZurichNLP/nmtscore). Compared to baselines such as sentence embeddings, translation-based measures prove competitive in paraphrase identification and are more robust against adversarial or multilingual input, especially if proper normalization is applied. When used for reference-based evaluation of data-to-text generation in 2 tasks and 17 languages, translation-based measures show a relatively high correlation to human judgments.
Text matching is a fundamental technique in both information retrieval and natural language processing. Text matching tasks share the same paradigm that determines the relationship between two given texts. The relationships vary from task to task, e.g.~relevance in document retrieval, semantic alignment in paraphrase identification and answerable judgment in question answering. However, the essential signals for text matching remain in a finite scope, i.e.~exact matching, semantic matching, and inference matching. Ideally, a good text matching model can learn to capture and aggregate these signals for different matching tasks to achieve competitive performance, while recent state-of-the-art text matching models, e.g.~Pre-trained Language Models (PLMs), are hard to generalize. It is because the end-to-end supervised learning on task-specific dataset makes model overemphasize the data sample bias and task-specific signals instead of the essential matching signals. To overcome this problem, we adopt a specialization-generalization training strategy and refer to it as Match-Prompt. In specialization stage, descriptions of different matching tasks are mapped to a few prompt tokens. In generalization stage, matching model explores the essential matching signals by being trained on diverse matching tasks. High diverse matching tasks avoid model fitting the data bias on a specific task, so that model can focus on learning the essential matching signals. Meanwhile, the prompt tokens obtained in the first step help the model distinguish different task-specific matching signals. Experimental results on public datasets show that Match-Prompt can improve multi-task generalization capability of PLMs in text matching and yield better in-domain multi-task, out-of-domain multi-task and new task adaptation performance than multi-task and task-specific models trained by previous fine-tuning paradigm.
While the general idea of self-supervised learning is identical across modalities, the actual algorithms and objectives differ widely because they were developed with a single modality in mind. To get us closer to general self-supervised learning, we present data2vec, a framework that uses the same learning method for either speech, NLP or computer vision. The core idea is to predict latent representations of the full input data based on a masked view of the input in a self-distillation setup using a standard Transformer architecture. Instead of predicting modality-specific targets such as words, visual tokens or units of human speech which are local in nature, data2vec predicts contextualized latent representations that contain information from the entire input. Experiments on the major benchmarks of speech recognition, image classification, and natural language understanding demonstrate a new state of the art or competitive performance to predominant approaches.
In this paper we introduce PerPaDa, a Persian paraphrase dataset that is collected from users' input in a plagiarism detection system. As an implicit crowdsourcing experience, we have gathered a large collection of original and paraphrased sentences from Hamtajoo; a Persian plagiarism detection system, in which users try to conceal cases of text re-use in their documents by paraphrasing and re-submitting manuscripts for analysis. The compiled dataset contains 2446 instances of paraphrasing. In order to improve the overall quality of the collected data, some heuristics have been used to exclude sentences that don't meet the proposed criteria. The introduced corpus is much larger than the available datasets for the task of paraphrase identification in Persian. Moreover, there is less bias in the data compared to the similar datasets, since the users did not try some fixed predefined rules in order to generate similar texts to their original inputs.
Predictive uncertainty estimation of pre-trained language models is an important measure of how likely people can trust their predictions. However, little is known about what makes a model prediction uncertain. Explaining predictive uncertainty is an important complement to explaining prediction labels in helping users understand model decision making and gaining their trust on model predictions, while has been largely ignored in prior works. In this work, we propose to explain the predictive uncertainty of pre-trained language models by extracting uncertain words from existing model explanations. We find the uncertain words are those identified as making negative contributions to prediction labels, while actually explaining the predictive uncertainty. Experiments show that uncertainty explanations are indispensable to explaining models and helping humans understand model prediction behavior.
In NLP, a large volume of tasks involve pairwise comparison between two sequences (e.g. sentence similarity and paraphrase identification). Predominantly, two formulations are used for sentence-pair tasks: bi-encoders and cross-encoders. Bi-encoders produce fixed-dimensional sentence representations and are computationally efficient, however, they usually underperform cross-encoders. Cross-encoders can leverage their attention heads to exploit inter-sentence interactions for better performance but they require task fine-tuning and are computationally more expensive. In this paper, we present a completely unsupervised sentence representation model termed as Trans-Encoder that combines the two learning paradigms into an iterative joint framework to simultaneously learn enhanced bi- and cross-encoders. Specifically, on top of a pre-trained Language Model (PLM), we start with converting it to an unsupervised bi-encoder, and then alternate between the bi- and cross-encoder task formulations. In each alternation, one task formulation will produce pseudo-labels which are used as learning signals for the other task formulation. We then propose an extension to conduct such self-distillation approach on multiple PLMs in parallel and use the average of their pseudo-labels for mutual-distillation. Trans-Encoder creates, to the best of our knowledge, the first completely unsupervised cross-encoder and also a state-of-the-art unsupervised bi-encoder for sentence similarity. Both the bi-encoder and cross-encoder formulations of Trans-Encoder outperform recently proposed state-of-the-art unsupervised sentence encoders such as Mirror-BERT and SimCSE by up to 5% on the sentence similarity benchmarks.
Task-adaptive pre-training (TAPT) and Self-training (ST) have emerged as the major semi-supervised approaches to improve natural language understanding (NLU) tasks with massive amount of unlabeled data. However, it's unclear whether they learn similar representations or they can be effectively combined. In this paper, we show that TAPT and ST can be complementary with simple TFS protocol by following TAPT -> Finetuning -> Self-training (TFS) process. Experimental results show that TFS protocol can effectively utilize unlabeled data to achieve strong combined gains consistently across six datasets covering sentiment classification, paraphrase identification, natural language inference, named entity recognition and dialogue slot classification. We investigate various semi-supervised settings and consistently show that gains from TAPT and ST can be strongly additive by following TFS procedure. We hope that TFS could serve as an important semi-supervised baseline for future NLP studies.
Transformers-based pretrained language models achieve outstanding results in many well-known NLU benchmarks. However, while pretraining methods are very convenient, they are expensive in terms of time and resources. This calls for a study of the impact of pretraining data size on the knowledge of the models. We explore this impact on the syntactic capabilities of RoBERTa, using models trained on incremental sizes of raw text data. First, we use syntactic structural probes to determine whether models pretrained on more data encode a higher amount of syntactic information. Second, we perform a targeted syntactic evaluation to analyze the impact of pretraining data size on the syntactic generalization performance of the models. Third, we compare the performance of the different models on three downstream applications: part-of-speech tagging, dependency parsing and paraphrase identification. We complement our study with an analysis of the cost-benefit trade-off of training such models. Our experiments show that while models pretrained on more data encode more syntactic knowledge and perform better on downstream applications, they do not always offer a better performance across the different syntactic phenomena and come at a higher financial and environmental cost.
Question Paraphrase Identification (QPI) is a critical task for large-scale Question-Answering forums. The purpose of QPI is to determine whether a given pair of questions are semantically identical or not. Previous approaches for this task have yielded promising results, but have often relied on complex recurrence mechanisms that are expensive and time-consuming in nature. In this paper, we propose a novel architecture combining a Bidirectional Transformer Encoder with Convolutional Neural Networks for the QPI task. We produce the predictions from the proposed architecture using two different inference setups: Siamese and Matched Aggregation. Experimental results demonstrate that our model achieves state-of-the-art performance on the Quora Question Pairs dataset. We empirically prove that the addition of convolution layers to the model architecture improves the results in both inference setups. We also investigate the impact of partial and complete fine-tuning and analyze the trade-off between computational power and accuracy in the process. Based on the obtained results, we conclude that the Matched-Aggregation setup consistently outperforms the Siamese setup. Our work provides insights into what architecture combinations and setups are likely to produce better results for the QPI task.
Consistency, which refers to the capability of generating the same predictions for semantically similar contexts, is a highly desirable property for a sound language understanding model. Although recent pretrained language models (PLMs) deliver outstanding performance in various downstream tasks, they should exhibit consistent behaviour provided the models truly understand language. In this paper, we propose a simple framework named consistency analysis on language understanding models (CALUM)} to evaluate the model's lower-bound consistency ability. Through experiments, we confirmed that current PLMs are prone to generate inconsistent predictions even for semantically identical inputs. We also observed that multi-task training with paraphrase identification tasks is of benefit to improve consistency, increasing the consistency by 13% on average.
Sentence semantic matching requires an agent to determine the semantic relation between two sentences, which is widely used in various natural language tasks, such as Natural Language Inference (NLI), Paraphrase Identification (PI), and so on. Much recent progress has been made in this area, especially attention-based methods and pre-trained language model based methods. However, most of these methods focus on all the important parts in sentences in a static way and only emphasize how important the words are to the query, inhibiting the ability of attention mechanism. In order to overcome this problem and boost the performance of attention mechanism, we propose a novel dynamic re-read attention, which can pay close attention to one small region of sentences at each step and re-read the important parts for better sentence representations. Based on this attention variation, we develop a novel Dynamic Re-read Network (DRr-Net) for sentence semantic matching. Moreover, selecting one small region in dynamic re-read attention seems insufficient for sentence semantics, and employing pre-trained language models as input encoders will introduce incomplete and fragile representation problems. To this end, we extend DRrNet to Locally-Aware Dynamic Re-read Attention Net (LadRa-Net), in which local structure of sentences is employed to alleviate the shortcoming of Byte-Pair Encoding (BPE) in pre-trained language models and boost the performance of dynamic reread attention. Extensive experiments on two popular sentence semantic matching tasks demonstrate that DRr-Net can significantly improve the performance of sentence semantic matching. Meanwhile, LadRa-Net is able to achieve better performance by considering the local structures of sentences. In addition, it is exceedingly interesting that some discoveries in our experiments are consistent with some findings of psychological research.
While achieving state-of-the-art results in multiple tasks and languages, translation-based cross-lingual transfer is often overlooked in favour of massively multilingual pre-trained encoders. Arguably, this is due to its main limitations: 1) translation errors percolating to the classification phase and 2) the insufficient expressiveness of the maximum-likelihood translation. To remedy this, we propose a new technique that integrates both steps of the traditional pipeline (translation and classification) into a single model, by treating the intermediate translations as a latent random variable. As a result, 1) the neural machine translation system can be fine-tuned with a variant of Minimum Risk Training where the reward is the accuracy of the downstream task classifier. Moreover, 2) multiple samples can be drawn to approximate the expected loss across all possible translations during inference. We evaluate our novel latent translation-based model on a series of multilingual NLU tasks, including commonsense reasoning, paraphrase identification, and natural language inference. We report gains for both zero-shot and few-shot learning setups, up to 2.7 accuracy points on average, which are even more prominent for low-resource languages (e.g., Haitian Creole). Finally, we carry out in-depth analyses comparing different underlying NMT models and assessing the impact of alternative translations on the downstream performance.
State-of-the-art models in natural language processing rely on separate rigid subword tokenization algorithms, which limit their generalization ability and adaptation to new settings. In this paper, we propose a new model inductive bias that learns a subword tokenization end-to-end as part of the model. To this end, we introduce a soft gradient-based subword tokenization module (GBST) that automatically learns latent subword representations from characters in a data-driven fashion. Concretely, GBST enumerates candidate subword blocks and learns to score them in a position-wise fashion using a block scoring network. We additionally introduce Charformer, a deep Transformer model that integrates GBST and operates on the byte level. Via extensive experiments on English GLUE, multilingual, and noisy text datasets, we show that Charformer outperforms a series of competitive byte-level baselines while generally performing on par and sometimes outperforming subword-based models. Additionally, Charformer is fast, improving the speed of both vanilla byte-level and subword-level Transformers by 28%-100% while maintaining competitive quality. We believe this work paves the way for highly performant token-free models that are trained completely end-to-end.
If two sentences have the same meaning, it should follow that they are equivalent in their inferential properties, i.e., each sentence should textually entail the other. However, many paraphrase datasets currently in widespread use rely on a sense of paraphrase based on word overlap and syntax. Can we teach them instead to identify paraphrases in a way that draws on the inferential properties of the sentences, and is not over-reliant on lexical and syntactic similarities of a sentence pair? We apply the adversarial paradigm to this question, and introduce a new adversarial method of dataset creation for paraphrase identification: the Adversarial Paraphrasing Task (APT), which asks participants to generate semantically equivalent (in the sense of mutually implicative) but lexically and syntactically disparate paraphrases. These sentence pairs can then be used both to test paraphrase identification models (which get barely random accuracy) and then improve their performance. To accelerate dataset generation, we explore automation of APT using T5, and show that the resulting dataset also improves accuracy. We discuss implications for paraphrase detection and release our dataset in the hope of making paraphrase detection models better able to detect sentence-level meaning equivalence.
We propose a method for generating paraphrases of English questions that retain the original intent but use a different surface form. Our model combines a careful choice of training objective with a principled information bottleneck, to induce a latent encoding space that disentangles meaning and form. We train an encoder-decoder model to reconstruct a question from a paraphrase with the same meaning and an exemplar with the same surface form, leading to separated encoding spaces. We use a Vector-Quantized Variational Autoencoder to represent the surface form as a set of discrete latent variables, allowing us to use a classifier to select a different surface form at test time. Crucially, our method does not require access to an external source of target exemplars. Extensive experiments and a human evaluation show that we are able to generate paraphrases with a better tradeoff between semantic preservation and syntactic novelty compared to previous methods.
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.