This paper summarizes our submission to Task 2 of the second track of the 10th Dialog System Technology Challenge (DSTC10) "Knowledge-grounded Task-oriented Dialogue Modeling on Spoken Conversations". Similar to the previous year's iteration, the task consists of three subtasks: detecting whether a turn is knowledge seeking, selecting the relevant knowledge document and finally generating a grounded response. This year, the focus lies on adapting the system to noisy ASR transcripts. We explore different approaches to make the models more robust to this type of input and to adapt the generated responses to the style of spoken conversations. For the latter, we get the best results with a noisy channel model that additionally reduces the number of short and generic responses. Our best system achieved the 1st rank in the automatic and the 3rd rank in the human evaluation of the challenge.
Enhancing the robustness of vision algorithms in real-world scenarios is challenging. One reason is that existing robustness benchmarks are limited, as they either rely on synthetic data or ignore the effects of individual nuisance factors. We introduce OOD-CV, a benchmark dataset that includes out-of-distribution examples of 10 object categories in terms of pose, shape, texture, context and the weather conditions, and enables benchmarking models for image classification, object detection, and 3D pose estimation. In addition to this novel dataset, we contribute extensive experiments using popular baseline methods, which reveal that: 1. Some nuisance factors have a much stronger negative effect on the performance compared to others, also depending on the vision task. 2. Current approaches to enhance robustness have only marginal effects, and can even reduce robustness. 3. We do not observe significant differences between convolutional and transformer architectures. We believe our dataset provides a rich testbed to study robustness and will help push forward research in this area.
This paper presents SimMIM, a simple framework for masked image modeling. We simplify recently proposed related approaches without special designs such as block-wise masking and tokenization via discrete VAE or clustering. To study what let the masked image modeling task learn good representations, we systematically study the major components in our framework, and find that simple designs of each component have revealed very strong representation learning performance: 1) random masking of the input image with a moderately large masked patch size (e.g., 32) makes a strong pre-text task; 2) predicting raw pixels of RGB values by direct regression performs no worse than the patch classification approaches with complex designs; 3) the prediction head can be as light as a linear layer, with no worse performance than heavier ones. Using ViT-B, our approach achieves 83.8% top-1 fine-tuning accuracy on ImageNet-1K by pre-training also on this dataset, surpassing previous best approach by +0.6%. When applied on a larger model of about 650 million parameters, SwinV2-H, it achieves 87.1% top-1 accuracy on ImageNet-1K using only ImageNet-1K data. We also leverage this approach to facilitate the training of a 3B model (SwinV2-G), that by $40\times$ less data than that in previous practice, we achieve the state-of-the-art on four representative vision benchmarks. The code and models will be publicly available at https://github.com/microsoft/SimMIM.
This paper describes our participation in the DEtection of TOXicity in comments In Spanish (DETOXIS) shared task 2021 at the 3rd Workshop on Iberian Languages Evaluation Forum. The shared task is divided into two related classification tasks: (i) Task 1: toxicity detection and; (ii) Task 2: toxicity level detection. They focus on the xenophobic problem exacerbated by the spread of toxic comments posted in different online news articles related to immigration. One of the necessary efforts towards mitigating this problem is to detect toxicity in the comments. Our main objective was to implement an accurate model to detect xenophobia in comments about web news articles within the DETOXIS shared task 2021, based on the competition's official metrics: the F1-score for Task 1 and the Closeness Evaluation Metric (CEM) for Task 2. To solve the tasks, we worked with two types of machine learning models: (i) statistical models and (ii) Deep Bidirectional Transformers for Language Understanding (BERT) models. We obtained our best results in both tasks using BETO, an BERT model trained on a big Spanish corpus. We obtained the 3rd place in Task 1 official ranking with the F1-score of 0.5996, and we achieved the 6th place in Task 2 official ranking with the CEM of 0.7142. Our results suggest: (i) BERT models obtain better results than statistical models for toxicity detection in text comments; (ii) Monolingual BERT models have an advantage over multilingual BERT models in toxicity detection in text comments in their pre-trained language.
The popularity of social media has created problems such as hate speech and sexism. The identification and classification of sexism in social media are very relevant tasks, as they would allow building a healthier social environment. Nevertheless, these tasks are considerably challenging. This work proposes a system to use multilingual and monolingual BERT and data points translation and ensemble strategies for sexism identification and classification in English and Spanish. It was conducted in the context of the sEXism Identification in Social neTworks shared 2021 (EXIST 2021) task, proposed by the Iberian Languages Evaluation Forum (IberLEF). The proposed system and its main components are described, and an in-depth hyperparameters analysis is conducted. The main results observed were: (i) the system obtained better results than the baseline model (multilingual BERT); (ii) ensemble models obtained better results than monolingual models; and (iii) an ensemble model considering all individual models and the best standardized values obtained the best accuracies and F1-scores for both tasks. This work obtained first place in both tasks at EXIST, with the highest accuracies (0.780 for task 1 and 0.658 for task 2) and F1-scores (F1-binary of 0.780 for task 1 and F1-macro of 0.579 for task 2).
Machine reading comprehension (MRC) is a challenging natural language processing (NLP) task. Recently, the emergence of pre-trained models (PTM) has brought this research field into a new era, in which the training objective plays a key role. The masked language model (MLM) is a self-supervised training objective that widely used in various PTMs. With the development of training objectives, many variants of MLM have been proposed, such as whole word masking, entity masking, phrase masking, span masking, and so on. In different MLM, the length of the masked tokens is different. Similarly, in different machine reading comprehension tasks, the length of the answer is also different, and the answer is often a word, phrase, or sentence. Thus, in MRC tasks with different answer lengths, whether the length of MLM is related to performance is a question worth studying. If this hypothesis is true, it can guide us how to pre-train the MLM model with a relatively suitable mask length distribution for MRC task. In this paper, we try to uncover how much of MLM's success in the machine reading comprehension tasks comes from the correlation between masking length distribution and answer length in MRC dataset. In order to address this issue, herein, (1) we propose four MRC tasks with different answer length distributions, namely short span extraction task, long span extraction task, short multiple-choice cloze task, long multiple-choice cloze task; (2) four Chinese MRC datasets are created for these tasks; (3) we also have pre-trained four masked language models according to the answer length distributions of these datasets; (4) ablation experiments are conducted on the datasets to verify our hypothesis. The experimental results demonstrate that our hypothesis is true.
Graph Neural Networks (GNNs) have achieved promising performance in various real-world applications. Building a powerful GNN model is not a trivial task, as it requires a large amount of training data, powerful computing resources, and human expertise in fine-tuning the model. Moreover, with the development of adversarial attacks, e.g., model stealing attacks, GNNs raise challenges to model authentication. To avoid copyright infringement on GNNs, verifying the ownership of the GNN models is necessary. This paper presents a watermarking framework for GNNs for both graph and node classification tasks. We 1) design two strategies to generate watermarked data for the graph classification task and one for the node classification task, 2) embed the watermark into the host model through training to obtain the watermarked GNN model, and 3) verify the ownership of the suspicious model in a black-box setting. The experiments show that our framework can verify the ownership of GNN models with a very high probability (up to $99\%$) for both tasks. Finally, we experimentally show that our watermarking approach is robust against a state-of-the-art model extraction technique and four state-of-the-art defenses against backdoor attacks.
We present an overview of the second edition of the CheckThat! Lab at CLEF 2019. The lab featured two tasks in two different languages: English and Arabic. Task 1 (English) challenged the participating systems to predict which claims in a political debate or speech should be prioritized for fact-checking. Task 2 (Arabic) asked to (A) rank a given set of Web pages with respect to a check-worthy claim based on their usefulness for fact-checking that claim, (B) classify these same Web pages according to their degree of usefulness for fact-checking the target claim, (C) identify useful passages from these pages, and (D) use the useful pages to predict the claim's factuality. CheckThat! provided a full evaluation framework, consisting of data in English (derived from fact-checking sources) and Arabic (gathered and annotated from scratch) and evaluation based on mean average precision (MAP) and normalized discounted cumulative gain (nDCG) for ranking, and F1 for classification. A total of 47 teams registered to participate in this lab, and fourteen of them actually submitted runs (compared to nine last year). The evaluation results show that the most successful approaches to Task 1 used various neural networks and logistic regression. As for Task 2, learning-to-rank was used by the highest scoring runs for subtask A, while different classifiers were used in the other subtasks. We release to the research community all datasets from the lab as well as the evaluation scripts, which should enable further research in the important tasks of check-worthiness estimation and automatic claim verification.
The current UMLS (Unified Medical Language System) Metathesaurus construction process for integrating over 200 biomedical source vocabularies is expensive and error-prone as it relies on the lexical algorithms and human editors for deciding if the two biomedical terms are synonymous. Recent advances in Natural Language Processing such as Transformer models like BERT and its biomedical variants with contextualized word embeddings have achieved state-of-the-art (SOTA) performance on downstream tasks. We aim to validate if these approaches using the BERT models can actually outperform the existing approaches for predicting synonymy in the UMLS Metathesaurus. In the existing Siamese Networks with LSTM and BioWordVec embeddings, we replace the BioWordVec embeddings with the biomedical BERT embeddings extracted from each BERT model using different ways of extraction. In the Transformer architecture, we evaluate the use of the different biomedical BERT models that have been pre-trained using different datasets and tasks. Given the SOTA performance of these BERT models for other downstream tasks, our experiments yield surprisingly interesting results: (1) in both model architectures, the approaches employing these biomedical BERT-based models do not outperform the existing approaches using Siamese Network with BioWordVec embeddings for the UMLS synonymy prediction task, (2) the original BioBERT large model that has not been pre-trained with the UMLS outperforms the SapBERT models that have been pre-trained with the UMLS, and (3) using the Siamese Networks yields better performance for synonymy prediction when compared to using the biomedical BERT models.
We describe the fourth edition of the CheckThat! Lab, part of the 2021 Conference and Labs of the Evaluation Forum (CLEF). The lab evaluates technology supporting tasks related to factuality, and covers Arabic, Bulgarian, English, Spanish, and Turkish. Task 1 asks to predict which posts in a Twitter stream are worth fact-checking, focusing on COVID-19 and politics (in all five languages). Task 2 asks to determine whether a claim in a tweet can be verified using a set of previously fact-checked claims (in Arabic and English). Task 3 asks to predict the veracity of a news article and its topical domain (in English). The evaluation is based on mean average precision or precision at rank k for the ranking tasks, and macro-F1 for the classification tasks. This was the most popular CLEF-2021 lab in terms of team registrations: 132 teams. Nearly one-third of them participated: 15, 5, and 25 teams submitted official runs for tasks 1, 2, and 3, respectively.
We present the results and the main findings of the NLP4IF-2021 shared tasks. Task 1 focused on fighting the COVID-19 infodemic in social media, and it was offered in Arabic, Bulgarian, and English. Given a tweet, it asked to predict whether that tweet contains a verifiable claim, and if so, whether it is likely to be false, is of general interest, is likely to be harmful, and is worthy of manual fact-checking; also, whether it is harmful to society, and whether it requires the attention of policy makers. Task~2 focused on censorship detection, and was offered in Chinese. A total of ten teams submitted systems for task 1, and one team participated in task 2; nine teams also submitted a system description paper. Here, we present the tasks, analyze the results, and discuss the system submissions and the methods they used. Most submissions achieved sizable improvements over several baselines, and the best systems used pre-trained Transformers and ensembles. The data, the scorers and the leaderboards for the tasks are available at http://gitlab.com/NLP4IF/nlp4if-2021.
This paper presents the JHU-Microsoft joint submission for WMT 2021 quality estimation shared task. We only participate in Task 2 (post-editing effort estimation) of the shared task, focusing on the target-side word-level quality estimation. The techniques we experimented with include Levenshtein Transformer training and data augmentation with a combination of forward, backward, round-trip translation, and pseudo post-editing of the MT output. We demonstrate the competitiveness of our system compared to the widely adopted OpenKiwi-XLM baseline. Our system is also the top-ranking system on the MT MCC metric for the English-German language pair.
Reducing air pollution, such as CO2 and PM2.5 emissions, is one of the most important issues for many countries worldwide. Selecting an environmentally friendly transport mode can be an effective approach of individuals to reduce air pollution in daily life. In this study, we propose a method to simulate the effectiveness of an eco-friendly transport mode selection for reducing air pollution by using map search logs. We formulate the transport mode selection as a combinatorial optimization problem with the constraints regarding the total amount of CO2 emissions as an example of air pollution and the average travel time. The optimization results show that the total amount of CO2 emissions can be reduced by 9.23%, whereas the average travel time can in fact be reduced by 9.96%. Our research proposal won first prize in Regular Machine Learning Competition Track Task 2 at KDD Cup 2019.
Quality Estimation (QE) is an important component of the machine translation workflow as it assesses the quality of the translated output without consulting reference translations. In this paper, we discuss our submission to the WMT 2021 QE Shared Task. We participate in Task 2 sentence-level sub-task that challenge participants to predict the HTER score for sentence-level post-editing effort. Our proposed system is an ensemble of multilingual BERT (mBERT)-based regression models, which are generated by fine-tuning on different input settings. It demonstrates comparable performance with respect to the Pearson's correlation and beats the baseline system in MAE/ RMSE for several language pairs. In addition, we adapt our system for the zero-shot setting by exploiting target language-relevant language pairs and pseudo-reference translations.
This article presents SVC-onGoing, an on-going competition for on-line signature verification where researchers can easily benchmark their systems against the state of the art in an open common platform using large-scale public databases, such as DeepSignDB and SVC2021_EvalDB, and standard experimental protocols. SVC-onGoing is based on the ICDAR 2021 Competition on On-Line Signature Verification (SVC 2021), which has been extended to allow participants anytime. The goal of SVC-onGoing is to evaluate the limits of on-line signature verification systems on popular scenarios (office/mobile) and writing inputs (stylus/finger) through large-scale public databases. Three different tasks are considered in the competition, simulating realistic scenarios as both random and skilled forgeries are simultaneously considered on each task. The results obtained in SVC-onGoing prove the high potential of deep learning methods in comparison with traditional methods. In particular, the best signature verification system has obtained Equal Error Rate (EER) values of 3.33% (Task 1), 7.41% (Task 2), and 6.04% (Task 3). Future studies in the field should be oriented to improve the performance of signature verification systems on the challenging mobile scenarios of SVC-onGoing in which several mobile devices and the finger are used during the signature acquisition.
Scene video text spotting (SVTS) is a very important research topic because of many real-life applications. However, only a little effort has put to spotting scene video text, in contrast to massive studies of scene text spotting in static images. Due to various environmental interferences like motion blur, spotting scene video text becomes very challenging. To promote this research area, this competition introduces a new challenge dataset containing 129 video clips from 21 natural scenarios in full annotations. The competition containts three tasks, that is, video text detection (Task 1), video text tracking (Task 2) and end-to-end video text spotting (Task3). During the competition period (opened on 1st March, 2021 and closed on 11th April, 2021), a total of 24 teams participated in the three proposed tasks with 46 valid submissions, respectively. This paper includes dataset descriptions, task definitions, evaluation protocols and results summaries of the ICDAR 2021 on SVTS competition. Thanks to the healthy number of teams as well as submissions, we consider that the SVTS competition has been successfully held, drawing much attention from the community and promoting the field research and its development.
Automatic classification of disordered speech can provide an objective tool for identifying the presence and severity of speech impairment. Classification approaches can also help identify hard-to-recognize speech samples to teach ASR systems about the variable manifestations of impaired speech. Here, we develop and compare different deep learning techniques to classify the intelligibility of disordered speech on selected phrases. We collected samples from a diverse set of 661 speakers with a variety of self-reported disorders speaking 29 words or phrases, which were rated by speech-language pathologists for their overall intelligibility using a five-point Likert scale. We then evaluated classifiers developed using 3 approaches: (1) a convolutional neural network (CNN) trained for the task, (2) classifiers trained on non-semantic speech representations from CNNs that used an unsupervised objective [1], and (3) classifiers trained on the acoustic (encoder) embeddings from an ASR system trained on typical speech [2]. We found that the ASR encoder's embeddings considerably outperform the other two on detecting and classifying disordered speech. Further analysis shows that the ASR embeddings cluster speech by the spoken phrase, while the non-semantic embeddings cluster speech by speaker. Also, longer phrases are more indicative of intelligibility deficits than single words.
The BraTS 2021 challenge celebrates its 10th anniversary and is jointly organized by the Radiological Society of North America (RSNA), the American Society of Neuroradiology (ASNR), and the Medical Image Computing and Computer Assisted Interventions (MICCAI) society. Since its inception, BraTS has been focusing on being a common benchmarking venue for brain glioma segmentation algorithms, with well-curated multi-institutional multi-parametric magnetic resonance imaging (mpMRI) data. Gliomas are the most common primary malignancies of the central nervous system, with varying degrees of aggressiveness and prognosis. The RSNA-ASNR-MICCAI BraTS 2021 challenge targets the evaluation of computational algorithms assessing the same tumor compartmentalization, as well as the underlying tumor's molecular characterization, in pre-operative baseline mpMRI data from 2,040 patients. Specifically, the two tasks that BraTS 2021 focuses on are: a) the segmentation of the histologically distinct brain tumor sub-regions, and b) the classification of the tumor's O[6]-methylguanine-DNA methyltransferase (MGMT) promoter methylation status. The performance evaluation of all participating algorithms in BraTS 2021 will be conducted through the Sage Bionetworks Synapse platform (Task 1) and Kaggle (Task 2), concluding in distributing to the top ranked participants monetary awards of $60,000 collectively.
In this paper, we propose a solution to allow speaker conditioned speech models, such as VoiceFilter-Lite, to support an arbitrary number of enrolled users in a single pass. This is achieved by using an attention mechanism on multiple speaker embeddings to compute a single attentive embedding, which is then used as a side input to the model. We implemented multi-user VoiceFilter-Lite and evaluated it for three tasks: (1) a streaming automatic speech recognition (ASR) task; (2) a text-independent speaker verification task; and (3) a personalized keyphrase detection task, where ASR has to detect keyphrases from multiple enrolled users in a noisy environment. Our experiments show that, with up to four enrolled users, multi-user VoiceFilter-Lite is able to significantly reduce speech recognition and speaker verification errors when there is overlapping speech, without affecting performance under other acoustic conditions. This attentive speaker embedding approach can also be easily applied to other speaker-conditioned models such as personal VAD and personalized ASR.
With the expressed goal of improving system transparency and visual grounding in the reasoning process in VQA, we present a modular system for the task of compositional VQA based on scene graphs. Our system is called "Adventurer's Treasure Hunt" (or ATH), named after an analogy we draw between our model's search procedure for an answer and an adventurer's search for treasure. We developed ATH with three characteristic features in mind: 1. By design, ATH allows us to explicitly quantify the impact of each of the sub-components on overall VQA performance, as well as their performance on their individual sub-task. 2. By modeling the search task after a treasure hunt, ATH inherently produces an explicit, visually grounded inference path for the processed question. 3. ATH is the first GQA-trained VQA system that dynamically extracts answers by querying the visual knowledge base directly, instead of selecting one from a specially learned classifier's output distribution over a pre-fixed answer vocabulary. We report detailed results on all components and their contributions to overall VQA performance on the GQA dataset and show that ATH achieves the highest visual grounding score among all examined systems.