Learning representations that accurately model semantics is an important goal of natural language processing research. Many semantic phenomena depend on syntactic structure. Recent work examines the extent to which state-of-the-art models for pre-training representations, such as BERT, capture such structure-dependent phenomena, but is largely restricted to one phenomenon in English: number agreement between subjects and verbs. We evaluate BERT's sensitivity to four types of structure-dependent agreement relations in a new semi-automatically curated dataset across 26 languages. We show that both the single-language and multilingual BERT models capture syntax-sensitive agreement patterns well in general, but we also highlight the specific linguistic contexts in which their performance degrades.
FinNuE: Exposing the Risks of Using BERTScore for Numerical Semantic Evaluation in Finance
Data Science Degree Program, National Taiwan University（国立台湾大学数据科学学位学程）；Department of Computer Science, National Yang Ming Chiao Tung University（阳明交通大学计算机科学系）；Department of Information Management, National Taiwan University（国立台湾大学信息管理系）；Department of Electrical Engineering, National Taiwan University（国立台湾大学电机系）；Research Center for Information Technology Innovation, Academia Sinica（学术院资讯科技创新研究中心）
University of Illinois at Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）
Automatic evaluation metrics are indispensable for evaluating generated text. To date, these metrics have focused almost exclusively on the content selection aspect of the system output, ignoring the linguistic quality aspect altogether. We bridge this gap by proposing GRUEN for evaluating Grammaticality, non-Redundancy, focUs, structure and coherENce of generated text. GRUEN utilizes a BERT-based model and a class of syntactic, semantic, and contextual features to examine the system output. Unlike most existing evaluation metrics which require human references as an input, GRUEN is reference-…
Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing in English, as a case study for our experiments. NPIs like "any" are grammatical only if they appear in a licensing environment like negation ("Sue doesn't have any cats" vs. "Sue has any cats"). This phenomenon is cha…
Massachusetts Institute of Technology（麻省理工学院）；Stanford University（斯坦福大学）；Princeton University（普林斯顿大学）；Tel Aviv University（特拉维夫大学）；University of Washington（华盛顿大学）；NAVER；Google AI（谷歌人工智能）；Allen Institute for Artificial Intelligence（艾伦人工智能研究所）
We present the results of the Machine Reading for Question Answering (MRQA) 2019 shared task on evaluating the generalization capabilities of reading comprehension systems. In this task, we adapted and unified 18 distinct question answering datasets into the same format. Among them, six datasets were made available for training, six datasets were made available for development, and the final six were hidden for final evaluation. Ten teams submitted systems, which explored various ideas including data sampling, multi-task learning, adversarial training and ensembling. The best system achieved a…
We propose BERTScore, an automatic evaluation metric for text generation. Analogously to common metrics, BERTScore computes a similarity score for each token in the candidate sentence with each token in the reference sentence. However, instead of exact matches, we compute token similarity using contextual embeddings. We evaluate using the outputs of 363 machine translation and image captioning systems. BERTScore correlates better with human judgments and provides stronger model selection performance than existing metrics. Finally, we use an adversarial paraphrase detection task to show that BE…
Institute of Computer Science, University of Tartu（塔尔图大学计算机科学研究所）
Recently, large pre-trained language models, such as BERT, have reached state-of-the-art performance in many natural language processing tasks, but for many languages, including Estonian, BERT models are not yet available. However, there exist several multilingual BERT models that can handle multiple languages simultaneously and that have been trained also on Estonian data. In this paper, we evaluate four multilingual models -- multilingual BERT, multilingual distilled BERT, XLM and XLM-RoBERTa -- on several NLP tasks including POS and morphological tagging, NER and text classification. Our ai…
Fine-tuning language models, such as BERT, on domain specific corpora has proven to be valuable in domains like scientific papers and biomedical text. In this paper, we show that fine-tuning BERT on legal documents similarly provides valuable improvements on NLP tasks in the legal domain. Demonstrating this outcome is significant for analyzing commercial agreements, because obtaining large legal corpora is challenging due to their confidential nature. As such, we show that having access to large legal corpora is a competitive advantage for commercial applications, and academic research on anal…
University of Colorado Boulder（科罗拉多大学博尔德分校）；New York University（纽约大学）
Linguistically informed analyses of language models (LMs) contribute to the understanding and improvement of these models. Here, we introduce the corpus of Chinese linguistic minimal pairs (CLiMP), which can be used to investigate what knowledge Chinese LMs acquire. CLiMP consists of sets of 1,000 minimal pairs (MPs) for 16 syntactic contrasts in Mandarin, covering 9 major Mandarin linguistic phenomena. The MPs are semi-automatically generated, and human agreement with the labels in CLiMP is 95.8%. We evaluated 11 different LMs on CLiMP, covering n-grams, LSTMs, and Chinese BERT. We find that …
Transformer-based language models such as BERT have outperformed previous models on a large number of English benchmarks, but their evaluation is often limited to English or a small number of well-resourced languages. In this work, we evaluate monolingual, multilingual, and randomly initialized language models from the BERT family on a variety of Uralic languages including Estonian, Finnish, Hungarian, Erzya, Moksha, Karelian, Livvi, Komi Permyak, Komi Zyrian, Northern Sámi, and Skolt Sámi. When monolingual models are available (currently only et, fi, hu), these perform better on their native …
The Ohio State University（俄亥俄州立大学）；National Library of Medicine（美国国家医学图书馆）；AI Institute, University of South Carolina（南卡罗来纳大学人工智能研究所）；George Washington University（乔治华盛顿大学）
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 existin…
Indian Institute of Science（印度科学理工学院）；Google Brain（谷歌大脑）
Recent research has achieved impressive results on understanding and improving source code by building up on machine-learning techniques developed for natural languages. A significant advancement in natural-language understanding has come with the development of pre-trained contextual embeddings, such as BERT, which can be fine-tuned for downstream tasks with less labeled data and training budget, while achieving better accuracies. However, there is no attempt yet to obtain a high-quality contextual embedding of source code, and to evaluate it on multiple program-understanding tasks simultaneo…
Heidelberg Institute for Theoretical Studies（海德堡理论研究所）；Bielefeld University（比勒费尔德大学）；Technische Universität Darmstadt（达姆施塔特工业大学）
Recently, there has been a growing interest in designing text generation systems from a discourse coherence perspective, e.g., modeling the interdependence between sentences. Still, recent BERT-based evaluation metrics are weak in recognizing coherence, and thus are not reliable in a way to spot the discourse-level improvements of those text generation systems. In this work, we introduce DiscoScore, a parametrized discourse metric, which uses BERT to model discourse coherence from different perspectives, driven by Centering theory. Our experiments encompass 16 non-discourse and discourse metri…
While automatic summarization evaluation methods developed for English are routinely applied to other languages, this is the first attempt to systematically quantify their panlinguistic efficacy. We take a summarization corpus for eight different languages, and manually annotate generated summaries for focus (precision) and coverage (recall). Based on this, we evaluate 19 summarization evaluation metrics, and find that using multilingual BERT within BERTScore performs well across all languages, at a level above that for English.
University of Illinois at Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；Zhejiang University（浙江大学）；Tencent（腾讯）
Recent advances in large-scale language representation models such as BERT have improved the state-of-the-art performances in many NLP tasks. Meanwhile, character-level Chinese NLP models, including BERT for Chinese, have also demonstrated that they can outperform the existing models. In this paper, we show that, however, such BERT-based models are vulnerable under character-level adversarial attacks. We propose a novel Chinese char-level attack method against BERT-based classifiers. Essentially, we generate "small" perturbation on the character level in the embedding space and guide the chara…
Transformers allow attention between all pairs of tokens, but there is reason to believe that most of these connections - and their quadratic time and memory - may not be necessary. But which ones? We evaluate the impact of sparsification patterns with a series of ablation experiments. First, we compare masks based on syntax, lexical similarity, and token position to random connections, and measure which patterns reduce performance the least. We find that on three common finetuning tasks even using attention that is at least 78% sparse can have little effect on performance if applied at later …
Computer Science Department, Bar Ilan University（巴伊兰大学计算机科学系）；Allen Institute for Artificial Intelligence（艾伦人工智能研究所）
I assess the extent to which the recently introduced BERT model captures English syntactic phenomena, using (1) naturally-occurring subject-verb agreement stimuli; (2) "coloreless green ideas" subject-verb agreement stimuli, in which content words in natural sentences are randomly replaced with words sharing the same part-of-speech and inflection; and (3) manually crafted stimuli for subject-verb agreement and reflexive anaphora phenomena. The BERT model performs remarkably well on all cases.
Peer assessment has been widely applied across diverse academic fields over the last few decades and has demonstrated its effectiveness. However, the advantages of peer assessment can only be achieved with high-quality peer reviews. Previous studies have found that high-quality review comments usually comprise several features (e.g., contain suggestions, mention problems, use a positive tone). Thus, researchers have attempted to evaluate peer-review comments by detecting different features using various machine learning and deep learning models. However, there is no single study that investiga…
Brazilian Network Information Center (NIC.br)（巴西网络信息中心 (NIC.br)）；Pontifical Catholic University of São Paulo (PUC-SP)（圣保罗天主教大学 (PUC-SP)）
This paper investigates how Natural Language Understanding (NLU) could be applied in Emotion Recognition, a specific task in affective computing. We finetuned different transformers language models (BERT, DistilBERT, RoBERTa, XLNet, and ELECTRA) using a fine-grained emotion dataset and evaluating them in terms of performance (f1-score) and time to complete.
École Normale Supérieure-PSL（巴黎高等师范学院-PSL大学）；CNRS（法国国家科学研究中心）；U. Sorbonne Nouvelle（新索邦大学）；University of Pisa（比萨大学）
Although transformer-based Neural Language Models demonstrate impressive performance on a variety of tasks, their generalization abilities are not well understood. They have been shown to perform strongly on subject-verb number agreement in a wide array of settings, suggesting that they learned to track syntactic dependencies during their training even without explicit supervision. In this paper, we examine the extent to which BERT is able to perform lexically-independent subject-verb number agreement (NA) on targeted syntactic templates. To do so, we disrupt the lexical patterns found in natu…
École Normale Supérieure-PSL（巴黎高等师范学院（PSL大学））；CNRS（法国国家科学研究中心）；U. Sorbonne Nouvelle（新索邦大学）；University of Pisa（比萨大学）
Both humans and neural language models are able to perform subject-verb number agreement (SVA). In principle, semantics shouldn't interfere with this task, which only requires syntactic knowledge. In this work we test whether meaning interferes with this type of agreement in English in syntactic structures of various complexities. To do so, we generate both semantically well-formed and nonsensical items. We compare the performance of BERT-base to that of humans, obtained with a psycholinguistic online crowdsourcing experiment. We find that BERT and humans are both sensitive to our semantic man…
Microsoft Research Asia（微软亚洲研究院）；University of California, San Diego（加利福尼亚大学圣迭戈分校）；Stanford University（斯坦福大学）；Beihang University（北京航空航天大学）
Recent studies on compression of pretrained language models (e.g., BERT) usually use preserved accuracy as the metric for evaluation. In this paper, we propose two new metrics, label loyalty and probability loyalty that measure how closely a compressed model (i.e., student) mimics the original model (i.e., teacher). We also explore the effect of compression with regard to robustness under adversarial attacks. We benchmark quantization, pruning, knowledge distillation and progressive module replacing with loyalty and robustness. By combining multiple compression techniques, we provide a practic…
Toxic comments in online platforms are an unavoidable social issue under the cloak of anonymity. Hate speech detection has been actively done for languages such as English, German, or Italian, where manually labeled corpus has been released. In this work, we first present 9.4K manually labeled entertainment news comments for identifying Korean toxic speech, collected from a widely used online news platform in Korea. The comments are annotated regarding social bias and hate speech since both aspects are correlated. The inter-annotator agreement Krippendorff's alpha score is 0.492 and 0.496, res…
The lack of a commonly used benchmark data set (collection) such as (Super-)GLUE (Wang et al., 2018, 2019) for the evaluation of non-English pre-trained language models is a severe shortcoming of current English-centric NLP-research. It concentrates a large part of the research on English, neglecting the uncertainty when transferring conclusions found for the English language to other languages. We evaluate the performance of the German and multilingual BERT-based models currently available via the huggingface transformers library on the four tasks of the GermEval17 workshop. We compare them t…
University of Illinois at Urbana Champaign（伊利诺伊大学香槟分校）；TIB Leibniz Information Centre for Science and Technology（莱布尼茨科学与技术信息中心）；L3S Research Center at Leibniz University of Hannover（汉诺威莱布尼茨大学L3S研究中心）
With the rapid growth of research publications, there is a vast amount of scholarly knowledge that needs to be organized in digital libraries. To deal with this challenge, techniques relying on knowledge-graph structures are being advocated. Within such graph-based pipelines, inferring relation types between related scientific concepts is a crucial step. Recently, advanced techniques relying on language models pre-trained on the large corpus have been popularly explored for automatic relation classification. Despite remarkable contributions that have been made, many of these methods were evalu…
Automated classification of chief complaints from patient-generated text is a critical first step in developing scalable platforms to triage patients without human intervention. In this work, we evaluate several approaches to chief complaint classification using a novel Chief Complaint (CC) Dataset that contains ~200,000 patient-generated reasons-for-visit entries mapped to a set of 795 discrete chief complaints. We examine the use of several fine-tuned bidirectional transformer (BERT) models trained on both unrelated texts as well as on the CC dataset. We contrast this performance with a TF-I…
Sebelas Maret University（塞贝拉马特大学）；University of Western Australia（西澳大学）
It is challenging to control the quality of online information due to the lack of supervision over all the information posted online. Manual checking is almost impossible given the vast number of posts made on online media and how quickly they spread. Therefore, there is a need for automated rumour detection techniques to limit the adverse effects of spreading misinformation. Previous studies mainly focused on finding and extracting the significant features of text data. However, extracting features is time-consuming and not a highly effective process. This study proposes the BERT- based pre-t…
Objectives: To adapt and evaluate a deep learning language model for answering why-questions based on patient-specific clinical text. Materials and Methods: Bidirectional encoder representations from transformers (BERT) models were trained with varying data sources to perform SQuAD 2.0 style why-question answering (why-QA) on clinical notes. The evaluation focused on: 1) comparing the merits from different training data, 2) error analysis. Results: The best model achieved an accuracy of 0.707 (or 0.760 by partial match). Training toward customization for the clinical language helped increase 6…
Automatic readability assessment (ARA) is the task of evaluating the level of ease or difficulty of text documents for a target audience. For researchers, one of the many open problems in the field is to make such models trained for the task show efficacy even for low-resource languages. In this study, we propose an alternative way of utilizing the information-rich embeddings of BERT models with handcrafted linguistic features through a combined method for readability assessment. Results show that the proposed method outperforms classical approaches in readability assessment using English and …
In attempts to "explain" predictions of machine learning models, researchers have proposed hundreds of techniques for attributing predictions to features that are deemed important. While these attributions are often claimed to hold the potential to improve human "understanding" of the models, surprisingly little work explicitly evaluates progress towards this aspiration. In this paper, we conduct a crowdsourcing study, where participants interact with deception detection models that have been trained to distinguish between genuine and fake hotel reviews. They are challenged both to simulate th…
Evaluating BERT-based Scientific Relation Classifiers for Scholarly Knowledge Graph Construction on Digital Library Collections
University of Illinois at Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；TIB Leibniz Information Centre for Science and Technology（莱布尼茨科学技术信息中心（TIB））；L3S Research Center at Leibniz University of Hannover（汉诺威莱布尼茨大学L3S研究中心）
This work studies the semantic representations learned by BERT for compounds, that is, expressions such as sunlight or bodyguard. We build on recent studies that explore semantic information in Transformers at the word level and test whether BERT aligns with human semantic intuitions when dealing with expressions (e.g., sunlight) whose overall meaning depends -- to a various extent -- on the semantics of the constituent words (sun, light). We leverage a dataset that includes human judgments on two psycholinguistic measures of compound semantic analysis: lexeme meaning dominance (LMD; quantifyi…
Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS)（智能技术单一研究中心（CiTIUS））；Universidade de Santiago de Compostela（圣地亚哥德孔波斯特拉大学）
This paper explores the ability of Transformer models to capture subject-verb and noun-adjective agreement dependencies in Galician. We conduct a series of word prediction experiments in which we manipulate dependency length together with the presence of an attractor noun that acts as a lure. First, we evaluate the overall performance of the existing monolingual and multilingual models for Galician. Secondly, to observe the effects of the training process, we compare the different degrees of achievement of two monolingual BERT models at different training points. We also release their checkpoi…
Inria（法国国家信息与自动化研究所）；Sorbonne Université（索邦大学）；Bar Ilan University（巴伊兰大学）；Allen Institute for Artificial Intelligence（艾伦人工智能研究所）
Multilingual pretrained language models have demonstrated remarkable zero-shot cross-lingual transfer capabilities. Such transfer emerges by fine-tuning on a task of interest in one language and evaluating on a distinct language, not seen during the fine-tuning. Despite promising results, we still lack a proper understanding of the source of this transfer. Using a novel layer ablation technique and analyses of the model's internal representations, we show that multilingual BERT, a popular multilingual language model, can be viewed as the stacking of two sub-networks: a multilingual encoder fol…
How and to what extent does BERT encode syntactically-sensitive hierarchical information or positionally-sensitive linear information? Recent work has shown that contextual representations like BERT perform well on tasks that require sensitivity to linguistic structure. We present here two studies which aim to provide a better understanding of the nature of BERT's representations. The first of these focuses on the identification of structurally-defined elements using diagnostic classifiers, while the second explores BERT's representation of subject-verb agreement and anaphor-antecedent depende…
This work describes experiments which probe the hidden representations of several BERT-style models for morphological content. The goal is to examine the extent to which discrete linguistic structure, in the form of morphological features and feature values, presents itself in the vector representations and attention distributions of pre-trained language models for five European languages. The experiments contained herein show that (i) Transformer architectures largely partition their embedding space into convex sub-regions highly correlated with morphological feature value, (ii) the contextua…
Budapest University of Technology and Economics（布达佩斯技术与经济大学）；Institute for Computer Science and Control（计算机科学与控制研究所）；Alfréd Rényi Institute of Mathematics（阿尔弗雷德·莱尼数学研究所）
We present an extended comparison of contextualized language models for Hungarian. We compare huBERT, a Hungarian model against 4 multilingual models including the multilingual BERT model. We evaluate these models through three tasks, morphological probing, POS tagging and NER. We find that huBERT works better than the other models, often by a large margin, particularly near the global optimum (typically at the middle layers). We also find that huBERT tends to generate fewer subwords for one word and that using the last subword for token-level tasks is generally a better choice than using the …
Bar Ilan University（巴伊兰大学）；Allen Institute for Artificial Intelligence（艾伦人工智能研究所）；Johns Hopkins University（约翰斯·霍普金斯大学）；New York University（纽约大学）
When language models process syntactically complex sentences, do they use their representations of syntax in a manner that is consistent with the grammar of the language? We propose AlterRep, an intervention-based method to address this question. For any linguistic feature of a given sentence, AlterRep generates counterfactual representations by altering how the feature is encoded, while leaving intact all other aspects of the original representation. By measuring the change in a model's word prediction behavior when these counterfactual representations are substituted for the original ones, w…
The recently proposed massively multilingual neural machine translation (NMT) system has been shown to be capable of translating over 100 languages to and from English within a single model. Its improved translation performance on low resource languages hints at potential cross-lingual transfer capability for downstream tasks. In this paper, we evaluate the cross-lingual effectiveness of representations from the encoder of a massively multilingual NMT model on 5 downstream classification and sequence labeling tasks covering a diverse set of over 50 languages. We compare against a strong baseli…
Pre-trained models have demonstrated their effectiveness in many downstream natural language processing (NLP) tasks. The availability of multilingual pre-trained models enables zero-shot transfer of NLP tasks from high resource languages to low resource ones. However, recent research in improving pre-trained models focuses heavily on English. While it is possible to train the latest neural architectures for other languages from scratch, it is undesirable due to the required amount of compute. In this work, we tackle the problem of transferring an existing pre-trained model from English to othe…
Johns Hopkins University（约翰斯·霍普金斯大学）；New York University（纽约大学）
Structural probing work has found evidence for latent syntactic information in pre-trained language models. However, much of this analysis has focused on monolingual models, and analyses of multilingual models have employed correlational methods that are confounded by the choice of probing tasks. In this study, we causally probe multilingual language models (XGLM and multilingual BERT) as well as monolingual BERT-based models across various languages; we do this by performing counterfactual perturbations on neuron activations and observing the effect on models' subject-verb agreement probabili…
