Abstract:Multilingual pre-training significantly improves many multilingual NLP tasks, including machine translation. Most existing methods are based on some variants of masked language modeling and text-denoising objectives on monolingual data. Multilingual pre-training on monolingual data ignores the availability of parallel data in many language pairs. Also, some other works integrate the available human-generated parallel translation data in their pre-training. This kind of parallel data is definitely helpful, but it is limited even in high-resource language pairs. This paper introduces a novel semi-supervised method, SPDG, that generates high-quality pseudo-parallel data for multilingual pre-training. First, a denoising model is pre-trained on monolingual data to reorder, add, remove, and substitute words, enhancing the pre-training documents' quality. Then, we generate different pseudo-translations for each pre-training document using dictionaries for word-by-word translation and applying the pre-trained denoising model. The resulting pseudo-parallel data is then used to pre-train our multilingual sequence-to-sequence model, PEACH. Our experiments show that PEACH outperforms existing approaches used in training mT5 and mBART on various translation tasks, including supervised, zero- and few-shot scenarios. Moreover, PEACH's ability to transfer knowledge between similar languages makes it particularly useful for low-resource languages. Our results demonstrate that with high-quality dictionaries for generating accurate pseudo-parallel, PEACH can be valuable for low-resource languages.
Abstract:While deep learning in the form of recurrent neural networks (RNNs) has caused a significant improvement in neural language modeling, the fact that they are extremely prone to overfitting is still a mainly unresolved issue. In this paper we propose a regularization method based on generative adversarial networks (GANs) and adversarial training (AT), that can prevent overfitting in neural language models. Unlike common adversarial training methods such as the fast gradient sign method (FGSM) that require a second back-propagation through time, and therefore effectively require at least twice the amount of time for regular training, the overhead of our method does not exceed more than 20% of the training of the baselines.
Abstract:Differentiable neural architecture search (DARTS) is a popular method for neural architecture search (NAS), which performs cell-search and utilizes continuous relaxation to improve the search efficiency via gradient-based optimization. The main shortcoming of DARTS is performance collapse, where the discovered architecture suffers from a pattern of declining quality during search. Performance collapse has become an important topic of research, with many methods trying to solve the issue through either regularization or fundamental changes to DARTS. However, the weight-sharing framework used for cell-search in DARTS and the convergence of architecture parameters has not been analyzed yet. In this paper, we provide a thorough and novel theoretical and empirical analysis on DARTS and its point of convergence. We show that DARTS suffers from a specific structural flaw due to its weight-sharing framework that limits the convergence of DARTS to saturation points of the softmax function. This point of convergence gives an unfair advantage to layers closer to the output in choosing the optimal architecture, causing performance collapse. We then propose two new regularization terms that aim to prevent performance collapse by harmonizing operation selection via aligning gradients of layers. Experimental results on six different search spaces and three different datasets show that our method ($\Lambda$-DARTS) does indeed prevent performance collapse, providing justification for our theoretical analysis and the proposed remedy.
Abstract:Abstractive text summarization is one of the areas influenced by the emergence of pre-trained language models. Current pre-training works in abstractive summarization give more points to the summaries with more words in common with the main text and pay less attention to the semantic similarity between generated sentences and the original document. We propose ARMAN, a Transformer-based encoder-decoder model pre-trained with three novel objectives to address this issue. In ARMAN, salient sentences from a document are selected according to a modified semantic score to be masked and form a pseudo summary. To summarize more accurately and similar to human writing patterns, we applied modified sentence reordering. We evaluated our proposed models on six downstream Persian summarization tasks. Experimental results show that our proposed model achieves state-of-the-art performance on all six summarization tasks measured by ROUGE and BERTScore. Our models also outperform prior works in textual entailment, question paraphrasing, and multiple choice question answering. Finally, we established a human evaluation and show that using the semantic score significantly improves summarization results.
Abstract:This paper presents a technical report of our submission to the 4th task of SemEval-2021, titled: Reading Comprehension of Abstract Meaning. In this task, we want to predict the correct answer based on a question given a context. Usually, contexts are very lengthy and require a large receptive field from the model. Thus, common contextualized language models like BERT miss fine representation and performance due to the limited capacity of the input tokens. To tackle this problem, we used the Longformer model to better process the sequences. Furthermore, we utilized the method proposed in the Longformer benchmark on Wikihop dataset which improved the accuracy on our task data from 23.01% and 22.95% achieved by the baselines for subtask 1 and 2, respectively, to 70.30% and 64.38%.
Abstract:Detecting which parts of a sentence contribute to that sentence's toxicity -- rather than providing a sentence-level verdict of hatefulness -- would increase the interpretability of models and allow human moderators to better understand the outputs of the system. This paper presents our team's, UTNLP, methodology and results in the SemEval-2021 shared task 5 on toxic spans detection. We test multiple models and contextual embeddings and report the best setting out of all. The experiments start with keyword-based models and are followed by attention-based, named entity-based, transformers-based, and ensemble models. Our best approach, an ensemble model, achieves an F1 of 0.684 in the competition's evaluation phase.
Abstract:Response retrieval is a subset of neural ranking in which a model selects a suitable response from a set of candidates given a conversation history. Retrieval-based chat-bots are typically employed in information seeking conversational systems such as customer support agents. In order to make pairwise comparisons between a conversation history and a candidate response, two approaches are common: cross-encoders performing full self-attention over the pair and bi-encoders encoding the pair separately. The former gives better prediction quality but is too slow for practical use. In this paper, we propose a new cross-encoder architecture and transfer knowledge from this model to a bi-encoder model using distillation. This effectively boosts bi-encoder performance at no cost during inference time. We perform a detailed analysis of this approach on three response retrieval datasets.
Abstract:Work on retrieval-based chatbots, like most sequence pair matching tasks, can be divided into Cross-encoders that perform word matching over the pair, and Bi-encoders that encode the pair separately. The latter has better performance, however since candidate responses cannot be encoded offline, it is also much slower. Lately, multi-layer transformer architectures pre-trained as language models have been used to great effect on a variety of natural language processing and information retrieval tasks. Recent work has shown that these language models can be used in text-matching scenarios to create Bi-encoders that perform almost as well as Cross-encoders while having a much faster inference speed. In this paper, we expand upon this work by developing a sequence matching architecture that %takes into account contexts in the training dataset at inference time. utilizes the entire training set as a makeshift knowledge-base during inference. We perform detailed experiments demonstrating that this architecture can be used to further improve Bi-encoders performance while still maintaining a relatively high inference speed.
Abstract:News media websites are important online resources that have drawn great attention of text mining researchers. The main aim of this study is to propose a framework for ranking online news websites from different viewpoints. The ranking of news websites is useful information, which can benefit many news-related tasks such as news retrieval and news recommendation. In the proposed framework, the ranking of news websites is obtained by calculating three measures introduced in the paper and based on user-generated content. Each proposed measure is concerned with the performance of news websites from a particular viewpoint including the completeness of news reports, the diversity of events being covered by the website and its speed. The use of user-generated content in this framework, as a partly-unbiased, real-time and low cost content on the web distinguishes the proposed news website ranking framework from the literature. The results obtained for three prominent news websites, BBC, CNN, NYTimes, show that BBC has the best performance in terms of news completeness and speed, and NYTimes has the best diversity in comparison with the other two websites.
Abstract:In a disaster situation, first responders need to quickly acquire situational awareness and prioritize response based on the need, resources available and impact. Can they do this based on digital media such as Twitter alone, or newswire alone, or some combination of the two? We examine this question in the context of the 2015 Nepal Earthquakes. Because newswire articles are longer, effective summaries can be helpful in saving time yet giving key content. We evaluate the effectiveness of several unsupervised summarization techniques in capturing key content. We propose a method to link tweets written by the public and newswire articles, so that we can compare their key characteristics: timeliness, whether tweets appear earlier than their corresponding news articles, and content. A novel idea is to view relevant tweets as a summary of the matching news article and evaluate these summaries. Whenever possible, we present both quantitative and qualitative evaluations. One of our main findings is that tweets and newswire articles provide complementary perspectives that form a holistic view of the disaster situation.