Dong Nguyen

Iman Munire Bilal, Bo Wang, Adam Tsakalidis et al.

We introduce the task of microblog opinion summarisation (MOS) and share a dataset of 3100 gold-standard opinion summaries to facilitate research in this domain. The dataset contains summaries of tweets spanning a 2-year period and covers more topics than any other public Twitter summarisation dataset. Summaries are abstractive in nature and have been created by journalists skilled in summarising news articles following a template separating factual information (main story) from author opinions. Our method differs from previous work on generating gold-standard summaries from social media, which usually involves selecting representative posts and thus favours extractive summarisation models. To showcase the dataset's utility and challenges, we benchmark a range of abstractive and extractive state-of-the-art summarisation models and achieve good performance, with the former outperforming the latter. We also show that fine-tuning is necessary to improve performance and investigate the benefits of using different sample sizes.

Anna Wegmann, Marijn Schraagen, Dong Nguyen

Linguistic style is an integral component of language. Recent advances in the development of style representations have increasingly used training objectives from authorship verification (AV): Do two texts have the same author? The assumption underlying the AV training task (same author approximates same writing style) enables self-supervised and, thus, extensive training. However, a good performance on the AV task does not ensure good "general-purpose" style representations. For example, as the same author might typically write about certain topics, representations trained on AV might also encode content information instead of style alone. We introduce a variation of the AV training task that controls for content using conversation or domain labels. We evaluate whether known style dimensions are represented and preferred over content information through an original variation to the recently proposed STEL framework. We find that representations trained by controlling for conversation are better than representations trained with domain or no content control at representing style independent from content.

Yupei Du, Dong Nguyen

Fine-tuning pre-trained language models on downstream tasks with varying random seeds has been shown to be unstable, especially on small datasets. Many previous studies have investigated this instability and proposed methods to mitigate it. However, most studies only used the standard deviation of performance scores (SD) as their measure, which is a narrow characterization of instability. In this paper, we analyze SD and six other measures quantifying instability at different levels of granularity. Moreover, we propose a systematic framework to evaluate the validity of these measures. Finally, we analyze the consistency and difference between different measures by reassessing existing instability mitigation methods. We hope our results will inform the development of better measurements of fine-tuning instability.

Christian Fang, Qixiang Fang, Dong Nguyen

We describe our experiments for SemEval-2023 Task 4 on the identification of human values behind arguments (ValueEval). Because human values are subjective concepts which require precise definitions, we hypothesize that incorporating the definitions of human values (in the form of annotation instructions and validated survey items) during model training can yield better prediction performance. We explore this idea and show that our proposed models perform better than the challenge organizers' baselines, with improvements in macro F1 scores of up to 18%.

Yupei Du, Albert Gatt, Dong Nguyen

Despite the massive success of fine-tuning Pre-trained Language Models (PLMs), they remain susceptible to out-of-distribution input. Dataset cartography is a simple yet effective dual-model approach that improves the robustness of fine-tuned PLMs. It involves fine-tuning a model on the original training set (i.e. reference model), selecting a subset of important training instances based on the training dynamics, and fine-tuning again only on these selected examples (i.e. main model). However, this approach requires fine-tuning the same model twice, which is computationally expensive for large PLMs. In this paper, we show that (1) training dynamics are highly transferable across model sizes and pre-training methods, and that (2) fine-tuning main models using these selected training instances achieves higher training efficiency than empirical risk minimization (ERM). Building on these observations, we propose a novel fine-tuning approach: Fine-Tuning by transFerring Training dynamics (FTFT). Compared with dataset cartography, FTFT uses more efficient reference models and aggressive early stopping. FTFT achieves robustness improvements over ERM while lowering the training cost by up to $\sim 50\%$.

Anna Wegmann, Dong Nguyen, David Jurgens

Variation in language is ubiquitous and often systematically linked to regional, social, and contextual factors. Tokenizers split texts into smaller units and might behave differently for less common linguistic forms. This might affect downstream LLM performance differently on two types of tasks: Tasks where the model should be robust to language variation (e.g., for semantic tasks like NLI, labels do not depend on whether a text uses British or American spelling) and tasks where the model should be sensitive to language variation (e.g., for form-based tasks like authorship verification, labels depend on whether a text uses British or American spelling). We pre-train BERT base models with the popular Byte-Pair Encoding algorithm to investigate how key tokenization design choices impact the performance of downstream models: the corpus used to train the tokenizer, the pre-tokenizer and the vocabulary size. We find that the best tokenizer varies on the two task types and that the pre-tokenizer has the biggest overall impact on performance. Further, we introduce a new approach to estimate tokenizer impact on downstream LLM performance, showing substantial improvement over metrics like Rényi efficiency. We encourage more work on language variation and its relation to tokenizers and thus LLM performance.

Goya van Boven, Yupei Du, Dong Nguyen

Gender-neutral pronouns are increasingly being introduced across Western languages. Recent evaluations have however demonstrated that English NLP systems are unable to correctly process gender-neutral pronouns, with the risk of erasing and misgendering non-binary individuals. This paper examines a Dutch coreference resolution system's performance on gender-neutral pronouns, specifically hen and die. In Dutch, these pronouns were only introduced in 2016, compared to the longstanding existence of singular they in English. We additionally compare two debiasing techniques for coreference resolution systems in non-binary contexts: Counterfactual Data Augmentation (CDA) and delexicalisation. Moreover, because pronoun performance can be hard to interpret from a general evaluation metric like LEA, we introduce an innovative evaluation metric, the pronoun score, which directly represents the portion of correctly processed pronouns. Our results reveal diminished performance on gender-neutral pronouns compared to gendered counterparts. Nevertheless, although delexicalisation fails to yield improvements, CDA substantially reduces the performance gap between gendered and gender-neutral pronouns. We further show that CDA remains effective in low-resource settings, in which a limited set of debiasing documents is used. This efficacy extends to previously unseen neopronouns, which are currently infrequently used but may gain popularity in the future, underscoring the viability of effective debiasing with minimal resources and low computational costs.

Anna Wegmann, Tijs van den Broek, Dong Nguyen

Best practices for high conflict conversations like counseling or customer support almost always include recommendations to paraphrase the previous speaker. Although paraphrase classification has received widespread attention in NLP, paraphrases are usually considered independent from context, and common models and datasets are not applicable to dialog settings. In this work, we investigate paraphrases in dialog (e.g., Speaker 1: "That book is mine." becomes Speaker 2: "That book is yours."). We provide an operationalization of context-dependent paraphrases, and develop a training for crowd-workers to classify paraphrases in dialog. We introduce a dataset with utterance pairs from NPR and CNN news interviews annotated for context-dependent paraphrases. To enable analyses on label variation, the dataset contains 5,581 annotations on 600 utterance pairs. We present promising results with in-context learning and with token classification models for automatic paraphrase detection in dialog.

Qixiang Fang, Daniel L. Oberski, Dong Nguyen

Many existing benchmarks of large (multimodal) language models (LLMs) focus on measuring LLMs' academic proficiency, often with also an interest in comparing model performance with human test takers'. While such benchmarks have proven key to the development of LLMs, they suffer from several limitations, including questionable measurement quality (e.g., Do they measure what they are supposed to in a reliable way?), lack of quality assessment on the item level (e.g., Are some items more important or difficult than others?) and unclear human population reference (e.g., To whom can the model be compared?). In response to these challenges, we propose leveraging knowledge from psychometrics -- a field dedicated to the measurement of latent variables like academic proficiency -- into LLM benchmarking. We make four primary contributions. First, we reflect on current LLM benchmark developments and contrast them with psychometrics-based test development. Second, we introduce PATCH: a novel framework for {P}sychometrics-{A}ssis{T}ed ben{CH}marking of LLMs. PATCH addresses the aforementioned limitations. In particular, PATCH enables valid comparison between LLMs and human populations. Third, we demonstrate PATCH by measuring several LLMs' proficiency in 8th grade mathematics against 56 human populations. We show that adopting a psychometrics-based approach yields evaluation outcomes that diverge from those based on current benchmarking practices. Fourth, we release 4 high-quality datasets to support measuring and comparing LLM proficiency in grade school mathematics and science with human populations.

Dong Nguyen, Laura Rosseel

Spelling variation (e.g. funnnn vs. fun) can influence the social perception of texts and their writers: we often have various associations with different forms of writing (is the text informal? does the writer seem young?). In this study, we focus on the social perception of spelling variation in online writing in English and study to what extent this perception is aligned between humans and large language models (LLMs). Building on sociolinguistic methodology, we compare LLM and human ratings on three key social attributes of spelling variation (formality, carefulness, age). We find generally strong correlations in the ratings between humans and LLMs. However, notable differences emerge when we analyze the distribution of ratings and when comparing between different types of spelling variation.

Yupei Du, Yingjin Song, Hugh Mee Wong et al.

Data pruning, selecting small but impactful subsets, offers a promising way to efficiently scale NLP model training. However, existing methods often involve many different design choices, which have not been systematically studied. This limits future developments. In this work, we decompose data pruning into two key components: the data representation and the selection algorithm, and we systematically analyze their influence on the selection of instances. Our theoretical and empirical results highlight the crucial role of representations: better representations, e.g., training gradients, generally lead to a better selection of instances, regardless of the chosen selection algorithm. Furthermore, different selection algorithms excel in different settings, and none consistently outperforms the others. Moreover, the selection algorithms do not always align with their intended objectives: for example, algorithms designed for the same objective can select drastically different instances, highlighting the need for careful evaluation.

Qixiang Fang, Dong Nguyen, Daniel L Oberski

Text embedding models from Natural Language Processing can map text data (e.g. words, sentences, documents) to supposedly meaningful numerical representations (a.k.a. text embeddings). While such models are increasingly applied in social science research, one important issue is often not addressed: the extent to which these embeddings are valid representations of constructs relevant for social science research. We therefore propose the use of the classic construct validity framework to evaluate the validity of text embeddings. We show how this framework can be adapted to the opaque and high-dimensional nature of text embeddings, with application to survey questions. We include several popular text embedding methods (e.g. fastText, GloVe, BERT, Sentence-BERT, Universal Sentence Encoder) in our construct validity analyses. We find evidence of convergent and discriminant validity in some cases. We also show that embeddings can be used to predict respondent's answers to completely new survey questions. Furthermore, BERT-based embedding techniques and the Universal Sentence Encoder provide more valid representations of survey questions than do others. Our results thus highlight the necessity to examine the construct validity of text embeddings before deploying them in social science research.

Thuy T. Do, Du Nguyen, Anh Le et al.

Hydroxychloroquine (HCQ) is used to prevent or treat malaria caused by mosquito bites. Recently, the drug has been suggested to treat COVID-19, but that has not been supported by scientific evidence. The information regarding the drug efficacy has flooded social networks, posting potential threats to the community by perverting their perceptions of the drug efficacy. This paper studies the reactions of social network users on the recommendation of using HCQ for COVID-19 treatment by analyzing the reaction patterns and sentiment of the tweets. We collected 164,016 tweets from February to December 2020 and used a text mining approach to identify social reaction patterns and opinion change over time. Our descriptive analysis identified an irregularity of the users' reaction patterns associated tightly with the social and news feeds on the development of HCQ and COVID-19 treatment. The study linked the tweets and Google search frequencies to reveal the viewpoints of local communities on the use of HCQ for COVID-19 treatment across different states. Further, our tweet sentiment analysis reveals that public opinion changed significantly over time regarding the recommendation of using HCQ for COVID-19 treatment. The data showed that high support in the early dates but it significantly declined in October. Finally, using the manual classification of 4,850 tweets by humans as our benchmark, our sentiment analysis showed that the Google Cloud Natural Language algorithm outperformed the Valence Aware Dictionary and sEntiment Reasoner in classifying tweets, especially in the sarcastic tweet group.

Anna Wegmann, Dong Nguyen

Style is an integral part of natural language. However, evaluation methods for style measures are rare, often task-specific and usually do not control for content. We propose the modular, fine-grained and content-controlled similarity-based STyle EvaLuation framework (STEL) to test the performance of any model that can compare two sentences on style. We illustrate STEL with two general dimensions of style (formal/informal and simple/complex) as well as two specific characteristics of style (contrac'tion and numb3r substitution). We find that BERT-based methods outperform simple versions of commonly used style measures like 3-grams, punctuation frequency and LIWC-based approaches. We invite the addition of further tasks and task instances to STEL and hope to facilitate the improvement of style-sensitive measures.

Yupei Du, Qixiang Fang, Dong Nguyen

Various measures have been proposed to quantify human-like social biases in word embeddings. However, bias scores based on these measures can suffer from measurement error. One indication of measurement quality is reliability, concerning the extent to which a measure produces consistent results. In this paper, we assess three types of reliability of word embedding gender bias measures, namely test-retest reliability, inter-rater consistency and internal consistency. Specifically, we investigate the consistency of bias scores across different choices of random seeds, scoring rules and words. Furthermore, we analyse the effects of various factors on these measures' reliability scores. Our findings inform better design of word embedding gender bias measures. Moreover, we urge researchers to be more critical about the application of such measures.

Alexander Robertson, Farhana Ferdousi Liza, Dong Nguyen et al.

The semantics of emoji has, to date, been considered from a static perspective. We offer the first longitudinal study of how emoji semantics changes over time, applying techniques from computational linguistics to six years of Twitter data. We identify five patterns in emoji semantic development and find evidence that the less abstract an emoji is, the more likely it is to undergo semantic change. In addition, we analyse select emoji in more detail, examining the effect of seasonality and world events on emoji semantics. To aid future work on emoji and semantics, we make our data publicly available along with a web-based interface that anyone can use to explore semantic change in emoji.

Paul Röttger, Bertram Vidgen, Dong Nguyen et al.

Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck's utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.

Nicole Peinelt, Dong Nguyen, Maria Liakata

Question paraphrase identification is a key task in Community Question Answering (CQA) to determine if an incoming question has been previously asked. Many current models use word embeddings to identify duplicate questions, but the use of topic models in feature-engineered systems suggests that they can be helpful for this task, too. We therefore propose two ways of merging topics with word embeddings (early vs. late fusion) in a new neural architecture for question paraphrase identification. Our results show that our system outperforms neural baselines on multiple CQA datasets, while an ablation study highlights the importance of topics and especially early topic-embedding fusion in our architecture.

Dong Nguyen, Maria Liakata, Simon DeDeo et al.

In this article we describe our experiences with computational text analysis. We hope to achieve three primary goals. First, we aim to shed light on thorny issues not always at the forefront of discussions about computational text analysis methods. Second, we hope to provide a set of best practices for working with thick social and cultural concepts. Our guidance is based on our own experiences and is therefore inherently imperfect. Still, given our diversity of disciplinary backgrounds and research practices, we hope to capture a range of ideas and identify commonalities that will resonate for many. And this leads to our final goal: to help promote interdisciplinary collaborations. Interdisciplinary insights and partnerships are essential for realizing the full potential of any computational text analysis that involves social and cultural concepts, and the more we are able to bridge these divides, the more fruitful we believe our work will be.

Dong Nguyen, Barbara McGillivray, Taha Yasseri

The Internet facilitates large-scale collaborative projects and the emergence of Web 2.0 platforms, where producers and consumers of content unify, has drastically changed the information market. On the one hand, the promise of the "wisdom of the crowd" has inspired successful projects such as Wikipedia, which has become the primary source of crowd-based information in many languages. On the other hand, the decentralized and often un-monitored environment of such projects may make them susceptible to low quality content. In this work, we focus on Urban Dictionary, a crowd-sourced online dictionary. We combine computational methods with qualitative annotation and shed light on the overall features of Urban Dictionary in terms of growth, coverage and types of content. We measure a high presence of opinion-focused entries, as opposed to the meaning-focused entries that we expect from traditional dictionaries. Furthermore, Urban Dictionary covers many informal, unfamiliar words as well as proper nouns. Urban Dictionary also contains offensive content, but highly offensive content tends to receive lower scores through the dictionary's voting system. The low threshold to include new material in Urban Dictionary enables quick recording of new words and new meanings, but the resulting heterogeneous content can pose challenges in using Urban Dictionary as a source to study language innovation.

Dong Nguyen, Thomas Demeester, Dolf Trieschnigg et al.

A publicly available dataset for federated search reflecting a real web environment has long been absent, making it difficult for researchers to test the validity of their federated search algorithms for the web setting. We present several experiments and analyses on resource selection on the web using a recently released test collection containing the results from more than a hundred real search engines, ranging from large general web search engines such as Google, Bing and Yahoo to small domain-specific engines. First, we experiment with estimating the size of uncooperative search engines on the web using query based sampling and propose a new method using the ClueWeb09 dataset. We find the size estimates to be highly effective in resource selection. Second, we show that an optimized federated search system based on smaller web search engines can be an alternative to a system using large web search engines. Third, we provide an empirical comparison of several popular resource selection methods and find that these methods are not readily suitable for resource selection on the web. Challenges include the sparse resource descriptions and extremely skewed sizes of collections.

Dong Nguyen, Jacob Eisenstein

Quantifying the degree of spatial dependence for linguistic variables is a key task for analyzing dialectal variation. However, existing approaches have important drawbacks. First, they are based on parametric models of dependence, which limits their power in cases where the underlying parametric assumptions are violated. Second, they are not applicable to all types of linguistic data: some approaches apply only to frequencies, others to boolean indicators of whether a linguistic variable is present. We present a new method for measuring geographical language variation, which solves both of these problems. Our approach builds on Reproducing Kernel Hilbert space (RKHS) representations for nonparametric statistics, and takes the form of a test statistic that is computed from pairs of individual geotagged observations without aggregation into predefined geographical bins. We compare this test with prior work using synthetic data as well as a diverse set of real datasets: a corpus of Dutch tweets, a Dutch syntactic atlas, and a dataset of letters to the editor in North American newspapers. Our proposed test is shown to support robust inferences across a broad range of scenarios and types of data.

Thomas Demeester, Robin Aly, Djoerd Hiemstra et al.

Evaluation of search engines relies on assessments of search results for selected test queries, from which we would ideally like to draw conclusions in terms of relevance of the results for general (e.g., future, unknown) users. In practice however, most evaluation scenarios only allow us to conclusively determine the relevance towards the particular assessor that provided the judgments. A factor that cannot be ignored when extending conclusions made from assessors towards users, is the possible disagreement on relevance, assuming that a single gold truth label does not exist. This paper presents and analyzes the Predicted Relevance Model (PRM), which allows predicting a particular result's relevance for a random user, based on an observed assessment and knowledge on the average disagreement between assessors. With the PRM, existing evaluation metrics designed to measure binary assessor relevance, can be transformed into more robust and effectively graded measures that evaluate relevance towards a random user. It also leads to a principled way of quantifying multiple graded or categorical relevance levels for use as gains in established graded relevance measures, such as normalized discounted cumulative gain (nDCG), which nowadays often use heuristic and data-independent gain values. Given a set of test topics with graded relevance judgments, the PRM allows evaluating systems on different scenarios, such as their capability of retrieving top results, or how well they are able to filter out non-relevant ones. Its use in actual evaluation scenarios is illustrated on several information retrieval test collections.

Dong Nguyen, A. Seza Doğruöz, Carolyn P. Rosé et al.

Language is a social phenomenon and variation is inherent to its social nature. Recently, there has been a surge of interest within the computational linguistics (CL) community in the social dimension of language. In this article we present a survey of the emerging field of "Computational Sociolinguistics" that reflects this increased interest. We aim to provide a comprehensive overview of CL research on sociolinguistic themes, featuring topics such as the relation between language and social identity, language use in social interaction and multilingual communication. Moreover, we demonstrate the potential for synergy between the research communities involved, by showing how the large-scale data-driven methods that are widely used in CL can complement existing sociolinguistic studies, and how sociolinguistics can inform and challenge the methods and assumptions employed in CL studies. We hope to convey the possible benefits of a closer collaboration between the two communities and conclude with a discussion of open challenges.