Mark Dredze

Shijie Wu, Ozan Irsoy, Steven Lu et al. · deepmind

The use of NLP in the realm of financial technology is broad and complex, with applications ranging from sentiment analysis and named entity recognition to question answering. Large Language Models (LLMs) have been shown to be effective on a variety of tasks; however, no LLM specialized for the financial domain has been reported in literature. In this work, we present BloombergGPT, a 50 billion parameter language model that is trained on a wide range of financial data. We construct a 363 billion token dataset based on Bloomberg's extensive data sources, perhaps the largest domain-specific dataset yet, augmented with 345 billion tokens from general purpose datasets. We validate BloombergGPT on standard LLM benchmarks, open financial benchmarks, and a suite of internal benchmarks that most accurately reflect our intended usage. Our mixed dataset training leads to a model that outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks. Additionally, we explain our modeling choices, training process, and evaluation methodology. We release Training Chronicles (Appendix C) detailing our experience in training BloombergGPT.

Moniba Keymanesh, Adrian Benton, Mark Dredze

Expressing natural language descriptions of structured facts or relations -- data-to-text generation (D2T) -- increases the accessibility of structured knowledge repositories. Previous work shows that pre-trained language models(PLMs) perform remarkably well on this task after fine-tuning on a significant amount of task-specific training data. On the other hand, while auto-regressive PLMs can generalize from a few task examples, their efficacy at D2T is largely unexplored. Furthermore, we have an incomplete understanding of the limits of PLMs on D2T. In this work, we conduct an empirical study of both fine-tuned and auto-regressive PLMs on the DART multi-domain D2T dataset. We consider their performance as a function of the amount of task-specific data and how these data are incorporated into the models: zero and few-shot learning, and fine-tuning of model weights. In addition, we probe the limits of PLMs by measuring performance on subsets of the evaluation data: novel predicates and abstractive test examples. To improve the performance on these subsets, we investigate two techniques: providing predicate descriptions in the context and re-ranking generated candidates by information reflected in the source. Finally, we conduct a human evaluation of model errors and show that D2T generation tasks would benefit from datasets with more careful manual curation.

David Mueller, Nicholas Andrews, Mark Dredze

Traditional multi-task learning architectures train a single model across multiple tasks through a shared encoder followed by task-specific decoders. Learning these models often requires specialized training algorithms that address task-conflict in the shared parameter updates, which otherwise can lead to negative transfer. A new type of multi-task learning within NLP homogenizes multi-task architectures as a shared encoder and language model decoder, which does surprisingly well across a range of diverse tasks. Does this new architecture suffer from task-conflicts that require specialized training algorithms? We study how certain factors in the shift towards text-to-text models affects multi-task conflict and negative transfer, finding that both directional conflict and transfer are surprisingly constant across architectures.

Shijie Wu, Benjamin Van Durme, Mark Dredze

Pretrained multilingual encoders enable zero-shot cross-lingual transfer, but often produce unreliable models that exhibit high performance variance on the target language. We postulate that this high variance results from zero-shot cross-lingual transfer solving an under-specified optimization problem. We show that any linear-interpolated model between the source language monolingual model and source + target bilingual model has equally low source language generalization error, yet the target language generalization error reduces smoothly and linearly as we move from the monolingual to bilingual model, suggesting that the model struggles to identify good solutions for both source and target languages using the source language alone. Additionally, we show that zero-shot solution lies in non-flat region of target language error generalization surface, causing the high variance.

Jonathan Liu, Haoling Qiu, Jonathan Lasko et al.

Recent research has shown that hallucinations, omissions, and biases are prevalent in everyday use-cases of LLMs. However, chatbots used in medical contexts must provide consistent advice in situations where non-medical factors are involved, such as when demographic information is present. In order to understand the conditions under which medical chatbots fail to perform as expected, we develop an infrastructure that 1) automatically generates queries to probe LLMs and 2) evaluates answers to these queries using multiple LLM-as-a-judge setups and prompts. For 1), our prompt creation pipeline samples the space of patient demographics, histories, disorders, and writing styles to create realistic questions that we subsequently use to prompt LLMs. In 2), our evaluation pipeline provides hallucination and omission detection using LLM-as-a-judge as well as agentic workflows, in addition to LLM-as-a-judge treatment category detectors. As a baseline study, we perform two case studies on inter-LLM agreement and the impact of varying the answering and evaluation LLMs. We find that LLM annotators exhibit low agreement scores (average Cohen's Kappa $κ=0.118$), and only specific (answering, evaluation) LLM pairs yield statistically significant differences across writing styles, genders, and races. We recommend that studies using LLM evaluation use multiple LLMs as evaluators in order to avoid arriving at statistically significant but non-generalizable results, particularly in the absence of ground-truth data. We also suggest publishing inter-LLM agreement metrics for transparency. Our code and dataset are available here: https://github.com/BBN-E/medic-neurips-2025-demo.

Xiaolei Huang, Franck Dernoncourt, Mark Dredze

Clinical notes in Electronic Health Records (EHR) present rich documented information of patients to inference phenotype for disease diagnosis and study patient characteristics for cohort selection. Unsupervised user embedding aims to encode patients into fixed-length vectors without human supervisions. Medical concepts extracted from the clinical notes contain rich connections between patients and their clinical categories. However, existing unsupervised approaches of user embeddings from clinical notes do not explicitly incorporate medical concepts. In this study, we propose a concept-aware unsupervised user embedding that jointly leverages text documents and medical concepts from two clinical corpora, MIMIC-III and Diabetes. We evaluate user embeddings on both extrinsic and intrinsic tasks, including phenotype classification, in-hospital mortality prediction, patient retrieval, and patient relatedness. Experiments on the two clinical corpora show our approach exceeds unsupervised baselines, and incorporating medical concepts can significantly improve the baseline performance.

Carlos Aguirre, Mark Dredze, Philip Resnik

Stressors are related to depression, but this relationship is complex. We investigate the relationship between open-ended text responses about stressors and depressive symptoms across gender and racial/ethnic groups. First, we use topic models and other NLP tools to find thematic and vocabulary differences when reporting stressors across demographic groups. We train language models using self-reported stressors to predict depressive symptoms, finding a relationship between stressors and depression. Finally, we find that differences in stressors translate to downstream performance differences across demographic groups.

Keith Harrigian, Mark Dredze

Social media allows researchers to track societal and cultural changes over time based on language analysis tools. Many of these tools rely on statistical algorithms which need to be tuned to specific types of language. Recent studies have shown the absence of appropriate tuning, specifically in the presence of semantic shift, can hinder robustness of the underlying methods. However, little is known about the practical effect this sensitivity may have on downstream longitudinal analyses. We explore this gap in the literature through a timely case study: understanding shifts in depression during the course of the COVID-19 pandemic. We find that inclusion of only a small number of semantically-unstable features can promote significant changes in longitudinal estimates of our target outcome. At the same time, we demonstrate that a recently-introduced method for measuring semantic shift may be used to proactively identify failure points of language-based models and, in turn, improve predictive generalization.

Kaiser Sun, Mark Dredze

The development of large language models leads to the formation of a pre-train-then-align paradigm, in which the model is typically pre-trained on a large text corpus and undergoes a tuning stage to align the model with human preference or downstream tasks. In this work, we investigate the relationship between pre-training and fine-tuning by fine-tuning multiple intermediate pre-trained model checkpoints. Our results on 18 datasets suggest that i) continual pre-training improves the model in a latent way that unveils after fine-tuning; ii) with extra fine-tuning, the datasets that the model does not demonstrate capability gain much more than those that the model performs well during the pre-training stage; iii) although model benefits significantly through supervised fine-tuning, it may forget previously known domain knowledge and the tasks that are not seen during fine-tuning; iv) the model resembles high sensitivity to evaluation prompts after supervised fine-tuning, but this sensitivity can be alleviated by more pre-training.

Keith Harrigian, Mark Dredze

Self-disclosed mental health diagnoses, which serve as ground truth annotations of mental health status in the absence of clinical measures, underpin the conclusions behind most computational studies of mental health language from the last decade. However, psychiatric conditions are dynamic; a prior depression diagnosis may no longer be indicative of an individual's mental health, either due to treatment or other mitigating factors. We ask: to what extent are self-disclosures of mental health diagnoses actually relevant over time? We analyze recent activity from individuals who disclosed a depression diagnosis on social media over five years ago and, in turn, acquire a new understanding of how presentations of mental health status on social media manifest longitudinally. We also provide expanded evidence for the presence of personality-related biases in datasets curated using self-disclosed diagnoses. Our findings motivate three practical recommendations for improving mental health datasets curated using self-disclosed diagnoses: 1) Annotate diagnosis dates and psychiatric comorbidities; 2) Sample control groups using propensity score matching; 3) Identify and remove spurious correlations introduced by selection bias.

Keith Harrigian, Tina Tang, Anthony Gonzales et al.

Diabetic eye disease is a major cause of blindness worldwide. The ability to monitor relevant clinical trajectories and detect lapses in care is critical to managing the disease and preventing blindness. Alas, much of the information necessary to support these goals is found only in the free text of the electronic medical record. To fill this information gap, we introduce a system for extracting evidence from clinical text of 19 clinical concepts related to diabetic eye disease and inferring relevant attributes for each. In developing this ophthalmology phenotyping system, we are also afforded a unique opportunity to evaluate the effectiveness of clinical language models at adapting to new clinical domains. Across multiple training paradigms, we find that BERT language models pretrained on out-of-distribution clinical data offer no significant improvement over BERT language models pretrained on non-clinical data for our domain. Our study tempers recent claims that language models pretrained on clinical data are necessary for clinical NLP tasks and highlights the importance of not treating clinical language data as a single homogeneous domain.

Hanjie Chen, Zhouxiang Fang, Yash Singla et al.

LLMs have demonstrated impressive performance in answering medical questions, such as achieving passing scores on medical licensing examinations. However, medical board exams or general clinical questions do not capture the complexity of realistic clinical cases. Moreover, the lack of reference explanations means we cannot easily evaluate the reasoning of model decisions, a crucial component of supporting doctors in making complex medical decisions. To address these challenges, we construct two new datasets: JAMA Clinical Challenge and Medbullets.\footnote{Datasets and code are available at \url{https://github.com/HanjieChen/ChallengeClinicalQA}.} JAMA Clinical Challenge consists of questions based on challenging clinical cases, while Medbullets comprises simulated clinical questions. Both datasets are structured as multiple-choice question-answering tasks, accompanied by expert-written explanations. We evaluate seven LLMs on the two datasets using various prompts. Experiments demonstrate that our datasets are harder than previous benchmarks. In-depth automatic and human evaluations of model-generated explanations provide insights into the promise and deficiency of LLMs for explainable medical QA.

David Mueller, Mark Dredze, Nicholas Andrews

Despite the widespread adoption of multi-task training in deep learning, little is understood about how multi-task learning (MTL) affects generalization. Prior work has conjectured that the negative effects of MTL are due to optimization challenges that arise during training, and many optimization methods have been proposed to improve multi-task performance. However, recent work has shown that these methods fail to consistently improve multi-task generalization. In this work, we seek to improve our understanding of these failures by empirically studying how MTL impacts the optimization of tasks, and whether this impact can explain the effects of MTL on generalization. We show that MTL results in a generalization gap (a gap in generalization at comparable training loss) between single-task and multi-task trajectories early into training. However, we find that factors of the optimization trajectory previously proposed to explain generalization gaps in single-task settings cannot explain the generalization gaps between single-task and multi-task models. Moreover, we show that the amount of gradient conflict between tasks is correlated with negative effects to task optimization, but is not predictive of generalization. Our work sheds light on the underlying causes for failures in MTL and, importantly, raises questions about the role of general purpose multi-task optimization algorithms.

Carlos Aguirre, Kuleen Sasse, Isabel Cachola et al.

Recently, work in NLP has shifted to few-shot (in-context) learning, with large language models (LLMs) performing well across a range of tasks. However, while fairness evaluations have become a standard for supervised methods, little is known about the fairness of LLMs as prediction systems. Further, common standard methods for fairness involve access to models weights or are applied during finetuning, which are not applicable in few-shot learning. Do LLMs exhibit prediction biases when used for standard NLP tasks? In this work, we explore the effect of shots, which directly affect the performance of models, on the fairness of LLMs as NLP classification systems. We consider how different shot selection strategies, both existing and new demographically sensitive methods, affect model fairness across three standard fairness datasets. We discuss how future work can include LLM fairness evaluations.

Sebastian Gehrmann, Claire Huang, Xian Teng et al.

To responsibly develop Generative AI (GenAI) products, it is critical to define the scope of acceptable inputs and outputs. What constitutes a "safe" response is an actively debated question. Academic work puts an outsized focus on evaluating models by themselves for general purpose aspects such as toxicity, bias, and fairness, especially in conversational applications being used by a broad audience. In contrast, less focus is put on considering sociotechnical systems in specialized domains. Yet, those specialized systems can be subject to extensive and well-understood legal and regulatory scrutiny. These product-specific considerations need to be set in industry-specific laws, regulations, and corporate governance requirements. In this paper, we aim to highlight AI content safety considerations specific to the financial services domain and outline an associated AI content risk taxonomy. We compare this taxonomy to existing work in this space and discuss implications of risk category violations on various stakeholders. We evaluate how existing open-source technical guardrail solutions cover this taxonomy by assessing them on data collected via red-teaming activities. Our results demonstrate that these guardrails fail to detect most of the content risks we discuss.

Shiyue Zhang, Shijie Wu, Ozan Irsoy et al.

Autoregressive language models are trained by minimizing the cross-entropy of the model distribution Q relative to the data distribution P -- that is, minimizing the forward cross-entropy, which is equivalent to maximum likelihood estimation (MLE). We have observed that models trained in this way may "over-generalize", in the sense that they produce non-human-like text. Moreover, we believe that reverse cross-entropy, i.e., the cross-entropy of P relative to Q, is a better reflection of how a human would evaluate text generated by a model. Hence, we propose learning with MixCE, an objective that mixes the forward and reverse cross-entropies. We evaluate models trained with this objective on synthetic data settings (where P is known) and real data, and show that the resulting models yield better generated text without complex decoding strategies. Our code and models are publicly available at https://github.com/bloomberg/mixce-acl2023

Keith Harrigian, Carlos Aguirre, Mark Dredze

Data-driven methods for mental health treatment and surveillance have become a major focus in computational science research in the last decade. However, progress in the domain, in terms of both medical understanding and system performance, remains bounded by the availability of adequate data. Prior systematic reviews have not necessarily made it possible to measure the degree to which data-related challenges have affected research progress. In this paper, we offer an analysis specifically on the state of social media data that exists for conducting mental health research. We do so by introducing an open-source directory of mental health datasets, annotated using a standardized schema to facilitate meta-analysis.

Miriam Wanner, Seth Ebner, Zhengping Jiang et al.

As generated text becomes more commonplace, it is increasingly important to evaluate how well-supported such text is by external knowledge sources. Many approaches for evaluating textual support rely on some method for decomposing text into its individual subclaims which are scored against a trusted reference. We investigate how various methods of claim decomposition -- especially LLM-based methods -- affect the result of an evaluation approach such as the recently proposed FActScore, finding that it is sensitive to the decomposition method used. This sensitivity arises because such metrics attribute overall textual support to the model that generated the text even though error can also come from the metric's decomposition step. To measure decomposition quality, we introduce an adaptation of FActScore, which we call DecompScore. We then propose an LLM-based approach to generating decompositions inspired by Bertrand Russell's theory of logical atomism and neo-Davidsonian semantics and demonstrate its improved decomposition quality over previous methods.

Bang An, Shiyue Zhang, Mark Dredze

Efforts to ensure the safety of large language models (LLMs) include safety fine-tuning, evaluation, and red teaming. However, despite the widespread use of the Retrieval-Augmented Generation (RAG) framework, AI safety work focuses on standard LLMs, which means we know little about how RAG use cases change a model's safety profile. We conduct a detailed comparative analysis of RAG and non-RAG frameworks with eleven LLMs. We find that RAG can make models less safe and change their safety profile. We explore the causes of this change and find that even combinations of safe models with safe documents can cause unsafe generations. In addition, we evaluate some existing red teaming methods for RAG settings and show that they are less effective than when used for non-RAG settings. Our work highlights the need for safety research and red-teaming methods specifically tailored for RAG LLMs.

Miriam Wanner, Benjamin Van Durme, Mark Dredze

The decompose-then-verify strategy for verification of Large Language Model (LLM) generations decomposes claims that are then independently verified. Decontextualization augments text (claims) to ensure it can be verified outside of the original context, enabling reliable verification. While decomposition and decontextualization have been explored independently, their interactions in a complete system have not been investigated. Their conflicting purposes can create tensions: decomposition isolates atomic facts while decontextualization inserts relevant information. Furthermore, a decontextualized subclaim presents a challenge to the verification step: what part of the augmented text should be verified as it now contains multiple atomic facts? We conduct an evaluation of different decomposition, decontextualization, and verification strategies and find that the choice of strategy matters in the resulting factuality scores. Additionally, we introduce DnDScore, a decontextualization aware verification method which validates subclaims in the context of contextual information.

Fan Bai, Keith Harrigian, Joel Stremmel et al.

Clinical Question Answering (QA) systems enable doctors to quickly access patient information from electronic health records (EHRs). However, training these systems requires significant annotated data, which is limited due to the expertise needed and the privacy concerns associated with clinical data. This paper explores generating Clinical QA data using large language models (LLMs) in a zero-shot setting. We find that naive prompting often results in easy questions that do not reflect the complexity of clinical scenarios. To address this, we propose two prompting strategies: 1) instructing the model to generate questions that do not overlap with the input context, and 2) summarizing the input record using a predefined schema to scaffold question generation. Experiments on two Clinical QA datasets demonstrate that our method generates more challenging questions, significantly improving fine-tuning performance over baselines. We compare synthetic and gold data and find a gap between their training efficacy resulting from the quality of synthetically generated answers.

Sharon Levy, William D. Adler, Tahilin Sanchez Karver et al.

Large language models (LLMs) acquire beliefs about gender from training data and can therefore generate text with stereotypical gender attitudes. Prior studies have demonstrated model generations favor one gender or exhibit stereotypes about gender, but have not investigated the complex dynamics that can influence model reasoning and decision-making involving gender. We study gender equity within LLMs through a decision-making lens with a new dataset, DeMET Prompts, containing scenarios related to intimate, romantic relationships. We explore nine relationship configurations through name pairs across three name lists (men, women, neutral). We investigate equity in the context of gender roles through numerous lenses: typical and gender-neutral names, with and without model safety enhancements, same and mixed-gender relationships, and egalitarian versus traditional scenarios across various topics. While all models exhibit the same biases (women favored, then those with gender-neutral names, and lastly men), safety guardrails reduce bias. In addition, models tend to circumvent traditional male dominance stereotypes and side with 'traditionally female' individuals more often, suggesting relationships are viewed as a female domain by the models.

Sharon Levy, Tahilin Sanchez Karver, William D. Adler et al.

Chat-based large language models have the opportunity to empower individuals lacking high-quality healthcare access to receive personalized information across a variety of topics. However, users may ask underspecified questions that require additional context for a model to correctly answer. We study how large language model biases are exhibited through these contextual questions in the healthcare domain. To accomplish this, we curate a dataset of sexual and reproductive healthcare questions that are dependent on age, sex, and location attributes. We compare models' outputs with and without demographic context to determine group alignment among our contextual questions. Our experiments reveal biases in each of these attributes, where young adult female users are favored.

Yahan Li, Keith Harrigian, Ayah Zirikly et al.

Large language models with a transformer-based encoder/decoder architecture, such as T5, have become standard platforms for supervised tasks. To bring these technologies to the clinical domain, recent work has trained new or adapted existing models to clinical data. However, the evaluation of these clinical T5 models and comparison to other models has been limited. Are the clinical T5 models better choices than FLAN-tuned generic T5 models? Do they generalize better to new clinical domains that differ from the training sets? We comprehensively evaluate these models across several clinical tasks and domains. We find that clinical T5 models provide marginal improvements over existing models, and perform worse when evaluated on different domains. Our results inform future choices in developing clinical LLMs.

Heyuan Huang, Alexandra DeLucia, Vijay Murari Tiyyala et al.

While Large Language Models (LLMs) can generate fluent and convincing responses, they are not necessarily correct. This is especially apparent in the popular decompose-then-verify factuality evaluation pipeline, where LLMs evaluate generations by decomposing the generations into individual, valid claims. Factuality evaluation is especially important for medical answers, since incorrect medical information could seriously harm the patient. However, existing factuality systems are a poor match for the medical domain, as they are typically only evaluated on objective, entity-centric, formulaic texts such as biographies and historical topics. This differs from condition-dependent, conversational, hypothetical, sentence-structure diverse, and subjective medical answers, which makes decomposition into valid facts challenging. We propose MedScore, a new pipeline to decompose medical answers into condition-aware valid facts and verify against in-domain corpora. Our method extracts up to three times more valid facts than existing methods, reducing hallucination and vague references, and retaining condition-dependency in facts. The resulting factuality score substantially varies by decomposition method, verification corpus, and used backbone LLM, highlighting the importance of customizing each step for reliable factuality evaluation by using our generalizable and modularized pipeline for domain adaptation.

Alexandra DeLucia, Mark Dredze

Abstractive multi-document summarization (MDS) is the task of automatically summarizing information in multiple documents, from news articles to conversations with multiple speakers. The training approaches for current MDS models can be grouped into four approaches: end-to-end with special pre-training ("direct"), chunk-then-summarize, extract-then-summarize, and inference with GPT-style models. In this work, we evaluate MDS models across training approaches, domains, and dimensions (reference similarity, quality, and factuality), to analyze how and why models trained on one domain can fail to summarize documents from another (News, Science, and Conversation) in the zero-shot domain transfer setting. We define domain-transfer "failure" as a decrease in factuality, higher deviation from the target, and a general decrease in summary quality. In addition to exploring domain transfer for MDS models, we examine potential issues with applying popular summarization metrics out-of-the-box.

Kuleen Sasse, Carlos Aguirre, Isabel Cachola et al.

WARNING: This paper contains content that maybe upsetting or offensive to some readers. Dog whistles are coded expressions with dual meanings: one intended for the general public (outgroup) and another that conveys a specific message to an intended audience (ingroup). Often, these expressions are used to convey controversial political opinions while maintaining plausible deniability and slip by content moderation filters. Identification of dog whistles relies on curated lexicons, which have trouble keeping up to date. We introduce FETCH!, a task for finding novel dog whistles in massive social media corpora. We find that state-of-the-art systems fail to achieve meaningful results across three distinct social media case studies. We present EarShot, a strong baseline system that combines the strengths of vector databases and Large Language Models (LLMs) to efficiently and effectively identify new dog whistles.

Carlos Aguirre, Mark Dredze

Training supervised machine learning systems with a fairness loss can improve prediction fairness across different demographic groups. However, doing so requires demographic annotations for training data, without which we cannot produce debiased classifiers for most tasks. Drawing inspiration from transfer learning methods, we investigate whether we can utilize demographic data from a related task to improve the fairness of a target task. We adapt a single-task fairness loss to a multi-task setting to exploit demographic labels from a related task in debiasing a target task and demonstrate that demographic fairness objectives transfer fairness within a multi-task framework. Additionally, we show that this approach enables intersectional fairness by transferring between two datasets with different single-axis demographics. We explore different data domains to show how our loss can improve fairness domains and tasks.

Mahsa Yarmohammadi, Shijie Wu, Marc Marone et al.

Zero-shot cross-lingual information extraction (IE) describes the construction of an IE model for some target language, given existing annotations exclusively in some other language, typically English. While the advance of pretrained multilingual encoders suggests an easy optimism of "train on English, run on any language", we find through a thorough exploration and extension of techniques that a combination of approaches, both new and old, leads to better performance than any one cross-lingual strategy in particular. We explore techniques including data projection and self-training, and how different pretrained encoders impact them. We use English-to-Arabic IE as our initial example, demonstrating strong performance in this setting for event extraction, named entity recognition, part-of-speech tagging, and dependency parsing. We then apply data projection and self-training to three tasks across eight target languages. Because no single set of techniques performs the best across all tasks, we encourage practitioners to explore various configurations of the techniques described in this work when seeking to improve on zero-shot training.

Yuval Pinter, Amanda Stent, Mark Dredze et al.

Commonly-used transformer language models depend on a tokenization schema which sets an unchangeable subword vocabulary prior to pre-training, destined to be applied to all downstream tasks regardless of domain shift, novel word formations, or other sources of vocabulary mismatch. Recent work has shown that "token-free" models can be trained directly on characters or bytes, but training these models from scratch requires substantial computational resources, and this implies discarding the many domain-specific models that were trained on tokens. In this paper, we present XRayEmb, a method for retrofitting existing token-based models with character-level information. XRayEmb is composed of a character-level "encoder" that computes vector representations of character sequences, and a generative component that decodes from the internal representation to a character sequence. We show that incorporating XRayEmb's learned vectors into sequences of pre-trained token embeddings helps performance on both autoregressive and masked pre-trained transformer architectures and on both sequence-level and sequence tagging tasks, particularly on non-standard English text.

Elliot Schumacher, James Mayfield, Mark Dredze

Entity linking -- the task of identifying references in free text to relevant knowledge base representations -- often focuses on single languages. We consider multilingual entity linking, where a single model is trained to link references to same-language knowledge bases in several languages. We propose a neural ranker architecture, which leverages multilingual transformer representations of text to be easily applied to a multilingual setting. We then explore how a neural ranker trained in one language (e.g. English) transfers to an unseen language (e.g. Chinese), and find that while there is a consistent but not large drop in performance. How can this drop in performance be alleviated? We explore adding an adversarial objective to force our model to learn language-invariant representations. We find that using this approach improves recall in several datasets, often matching the in-language performance, thus alleviating some of the performance loss occurring from zero-shot transfer.

Zach Wood-Doughty, Isabel Cachola, Mark Dredze

Machine learning models that offer excellent predictive performance often lack the interpretability necessary to support integrated human machine decision-making. In clinical medicine and other high-risk settings, domain experts may be unwilling to trust model predictions without explanations. Work in explainable AI must balance competing objectives along two different axes: 1) Explanations must balance faithfulness to the model's decision-making with their plausibility to a domain expert. 2) Domain experts desire local explanations of individual predictions and global explanations of behavior in aggregate. We propose to train a proxy model that mimics the behavior of the trained model and provides fine-grained control over these trade-offs. We evaluate our approach on the task of assigning ICD codes to clinical notes to demonstrate that explanations from the proxy model are faithful and replicate the trained model behavior.

Aaron Mueller, Mark Dredze

Neural topic models can augment or replace bag-of-words inputs with the learned representations of deep pre-trained transformer-based word prediction models. One added benefit when using representations from multilingual models is that they facilitate zero-shot polylingual topic modeling. However, while it has been widely observed that pre-trained embeddings should be fine-tuned to a given task, it is not immediately clear what supervision should look like for an unsupervised task such as topic modeling. Thus, we propose several methods for fine-tuning encoders to improve both monolingual and zero-shot polylingual neural topic modeling. We consider fine-tuning on auxiliary tasks, constructing a new topic classification task, integrating the topic classification objective directly into topic model training, and continued pre-training. We find that fine-tuning encoder representations on topic classification and integrating the topic classification task directly into topic modeling improves topic quality, and that fine-tuning encoder representations on any task is the most important factor for facilitating cross-lingual transfer.

Carlos Aguirre, Keith Harrigian, Mark Dredze

Multiple studies have demonstrated that behavior on internet-based social media platforms can be indicative of an individual's mental health status. The widespread availability of such data has spurred interest in mental health research from a computational lens. While previous research has raised concerns about possible biases in models produced from this data, no study has quantified how these biases actually manifest themselves with respect to different demographic groups, such as gender and racial/ethnic groups. Here, we analyze the fairness of depression classifiers trained on Twitter data with respect to gender and racial demographic groups. We find that model performance systematically differs for underrepresented groups and that these discrepancies cannot be fully explained by trivial data representation issues. Our study concludes with recommendations on how to avoid these biases in future research.

Xiaolei Huang, Michael J. Paul, Robin Burke et al.

Language varies across users and their interested fields in social media data: words authored by a user across his/her interests may have different meanings (e.g., cool) or sentiments (e.g., fast). However, most of the existing methods to train user embeddings ignore the variations across user interests, such as product and movie categories (e.g., drama vs. action). In this study, we treat the user interest as domains and empirically examine how the user language can vary across the user factor in three English social media datasets. We then propose a user embedding model to account for the language variability of user interests via a multitask learning framework. The model learns user language and its variations without human supervision. While existing work mainly evaluated the user embedding by extrinsic tasks, we propose an intrinsic evaluation via clustering and evaluate user embeddings by an extrinsic task, text classification. The experiments on the three English-language social media datasets show that our proposed approach can generally outperform baselines via adapting the user factor.

Zach Wood-Doughty, Ilya Shpitser, Mark Dredze

Drawing causal conclusions from observational data requires making assumptions about the true data-generating process. Causal inference research typically considers low-dimensional data, such as categorical or numerical fields in structured medical records. High-dimensional and unstructured data such as natural language complicates the evaluation of causal inference methods; such evaluations rely on synthetic datasets with known causal effects. Models for natural language generation have been widely studied and perform well empirically. However, existing methods not immediately applicable to producing synthetic datasets for causal evaluations, as they do not allow for quantifying a causal effect on the text itself. In this work, we develop a framework for adapting existing generation models to produce synthetic text datasets with known causal effects. We use this framework to perform an empirical comparison of four recently-proposed methods for estimating causal effects from text data. We release our code and synthetic datasets.

Rachel Dorn, Alicia L. Nobles, Masoud Rouhizadeh et al.

The identification and removal/replacement of protected information from social media data is an understudied problem, despite being desirable from an ethical and legal perspective. This paper identifies types of potentially directly identifiable information (inspired by protected health information in clinical texts) contained in tweets that may be readily removed using off-the-shelf algorithms, introduces an English dataset of tweets annotated for identifiable information, and compiles these off-the-shelf algorithms into a tool (Nightjar) to evaluate the feasibility of using Nightjar to remove directly identifiable information from the tweets. Nightjar as well as the annotated data can be retrieved from https://bitbucket.org/mdredze/nightjar.

Elliot Schumacher, James Mayfield, Mark Dredze

Cross-language entity linking grounds mentions in multiple languages to a single-language knowledge base. We propose a neural ranking architecture for this task that uses multilingual BERT representations of the mention and the context in a neural network. We find that the multilingual ability of BERT leads to robust performance in monolingual and multilingual settings. Furthermore, we explore zero-shot language transfer and find surprisingly robust performance. We investigate the zero-shot degradation and find that it can be partially mitigated by a proposed auxiliary training objective, but that the remaining error can best be attributed to domain shift rather than language transfer.

Aaron Mueller, Zach Wood-Doughty, Silvio Amir et al.

The #MeToo movement on Twitter has drawn attention to the pervasive nature of sexual harassment and violence. While #MeToo has been praised for providing support for self-disclosures of harassment or violence and shifting societal response, it has also been criticized for exemplifying how women of color have been discounted for their historical contributions to and excluded from feminist movements. Through an analysis of over 600,000 tweets from over 256,000 unique users, we examine online #MeToo conversations across gender and racial/ethnic identities and the topics that each demographic emphasized. We found that tweets authored by white women were overrepresented in the movement compared to other demographics, aligning with criticism of unequal representation. We found that intersected identities contributed differing narratives to frame the movement, co-opted the movement to raise visibility in parallel ongoing movements, employed the same hashtags both critically and supportively, and revived and created new hashtags in response to pivotal moments. Notably, tweets authored by black women often expressed emotional support and were critical about differential treatment in the justice system and by police. In comparison, tweets authored by white women and men often highlighted sexual harassment and violence by public figures and weaved in more general political discussions. We discuss the implications of work for digital activism research and design including suggestions to raise visibility by those who were under-represented in this hashtag activism movement. Content warning: this article discusses issues of sexual harassment and violence.

Shijie Wu, Mark Dredze

Multilingual BERT (mBERT), XLM-RoBERTa (XLMR) and other unsupervised multilingual encoders can effectively learn cross-lingual representation. Explicit alignment objectives based on bitexts like Europarl or MultiUN have been shown to further improve these representations. However, word-level alignments are often suboptimal and such bitexts are unavailable for many languages. In this paper, we propose a new contrastive alignment objective that can better utilize such signal, and examine whether these previous alignment methods can be adapted to noisier sources of aligned data: a randomly sampled 1 million pair subset of the OPUS collection. Additionally, rather than report results on a single dataset with a single model run, we report the mean and standard derivation of multiple runs with different seeds, on four datasets and tasks. Our more extensive analysis finds that, while our new objective outperforms previous work, overall these methods do not improve performance with a more robust evaluation framework. Furthermore, the gains from using a better underlying model eclipse any benefits from alignment training. These negative results dictate more care in evaluating these methods and suggest limitations in applying explicit alignment objectives.

Shijie Wu, Mark Dredze

Multilingual BERT (mBERT) trained on 104 languages has shown surprisingly good cross-lingual performance on several NLP tasks, even without explicit cross-lingual signals. However, these evaluations have focused on cross-lingual transfer with high-resource languages, covering only a third of the languages covered by mBERT. We explore how mBERT performs on a much wider set of languages, focusing on the quality of representation for low-resource languages, measured by within-language performance. We consider three tasks: Named Entity Recognition (99 languages), Part-of-speech Tagging, and Dependency Parsing (54 languages each). mBERT does better than or comparable to baselines on high resource languages but does much worse for low resource languages. Furthermore, monolingual BERT models for these languages do even worse. Paired with similar languages, the performance gap between monolingual BERT and mBERT can be narrowed. We find that better models for low resource languages require more efficient pretraining techniques or more data.

David Mueller, Nicholas Andrews, Mark Dredze

Named-entities are inherently multilingual, and annotations in any given language may be limited. This motivates us to consider polyglot named-entity recognition (NER), where one model is trained using annotated data drawn from more than one language. However, a straightforward implementation of this simple idea does not always work in practice: naive training of NER models using annotated data drawn from multiple languages consistently underperforms models trained on monolingual data alone, despite having access to more training data. The starting point of this paper is a simple solution to this problem, in which polyglot models are fine-tuned on monolingual data to consistently and significantly outperform their monolingual counterparts. To explain this phenomena, we explore the sources of multilingual transfer in polyglot NER models and examine the weight structure of polyglot models compared to their monolingual counterparts. We find that polyglot models efficiently share many parameters across languages and that fine-tuning may utilize a large number of those parameters.

Andriy Mulyar, Elliot Schumacher, Masoud Rouhizadeh et al.

Clinical notes contain an extensive record of a patient's health status, such as smoking status or the presence of heart conditions. However, this detail is not replicated within the structured data of electronic health systems. Phenotyping, the extraction of patient conditions from free clinical text, is a critical task which supports avariety of downstream applications such as decision support and secondary use of medical records. Previous work has resulted in systems which are high performing but require hand engineering, often of rules. Recent work in pretrained contextualized language models have enabled advances in representing text for a variety of tasks. We therefore explore several architectures for modeling pheno-typing that rely solely on BERT representations of the clinical note, removing the need for manual engineering. We find these architectures are competitive with or outperform existing state of the art methods on two phenotyping tasks.

Shijie Wu, Mark Dredze

Pretrained contextual representation models (Peters et al., 2018; Devlin et al., 2018) have pushed forward the state-of-the-art on many NLP tasks. A new release of BERT (Devlin, 2018) includes a model simultaneously pretrained on 104 languages with impressive performance for zero-shot cross-lingual transfer on a natural language inference task. This paper explores the broader cross-lingual potential of mBERT (multilingual) as a zero shot language transfer model on 5 NLP tasks covering a total of 39 languages from various language families: NLI, document classification, NER, POS tagging, and dependency parsing. We compare mBERT with the best-published methods for zero-shot cross-lingual transfer and find mBERT competitive on each task. Additionally, we investigate the most effective strategy for utilizing mBERT in this manner, determine to what extent mBERT generalizes away from language specific features, and measure factors that influence cross-lingual transfer.

Zach Wood-Doughty, Ilya Shpitser, Mark Dredze

Causal understanding is essential for many kinds of decision-making, but causal inference from observational data has typically only been applied to structured, low-dimensional datasets. While text classifiers produce low-dimensional outputs, their use in causal inference has not previously been studied. To facilitate causal analyses based on language data, we consider the role that text classifiers can play in causal inference through established modeling mechanisms from the causality literature on missing data and measurement error. We demonstrate how to conduct causal analyses using text classifiers on simulated and Yelp data, and discuss the opportunities and challenges of future work that uses text data in causal inference.

Travis Wolfe, Mark Dredze, Benjamin Van Durme

Hand-engineered feature sets are a well understood method for creating robust NLP models, but they require a lot of expertise and effort to create. In this work we describe how to automatically generate rich feature sets from simple units called featlets, requiring less engineering. Using information gain to guide the generation process, we train models which rival the state of the art on two standard Semantic Role Labeling datasets with almost no task or linguistic insight.

Ann Irvine, Mark Dredze

This work presents a systematic theoretical and empirical comparison of the major algorithms that have been proposed for learning Harmonic and Optimality Theory grammars (HG and OT, respectively). By comparing learning algorithms, we are also able to compare the closely related OT and HG frameworks themselves. Experimental results show that the additional expressivity of the HG framework over OT affords performance gains in the task of predicting the surface word order of Czech sentences. We compare the perceptron with the classic Gradual Learning Algorithm (GLA), which learns OT grammars, as well as the popular Maximum Entropy model. In addition to showing that the perceptron is theoretically appealing, our work shows that the performance of the HG model it learns approaches that of the upper bound in prediction accuracy on a held out test set and that it is capable of accurately modeling observed variation.

Nanyun Peng, Mark Dredze

Many domain adaptation approaches rely on learning cross domain shared representations to transfer the knowledge learned in one domain to other domains. Traditional domain adaptation only considers adapting for one task. In this paper, we explore multi-task representation learning under the domain adaptation scenario. We propose a neural network framework that supports domain adaptation for multiple tasks simultaneously, and learns shared representations that better generalize for domain adaptation. We apply the proposed framework to domain adaptation for sequence tagging problems considering two tasks: Chinese word segmentation and named entity recognition. Experiments show that multi-task domain adaptation works better than disjoint domain adaptation for each task, and achieves the state-of-the-art results for both tasks in the social media domain.

Mark Dredze, Manuel García-Herranz, Alex Rutherford et al.

Data on human spatial distribution and movement is essential for understanding and analyzing social systems. However existing sources for this data are lacking in various ways; difficult to access, biased, have poor geographical or temporal resolution, or are significantly delayed. In this paper, we describe how geolocation data from Twitter can be used to estimate global mobility patterns and address these shortcomings. These findings will inform how this novel data source can be harnessed to address humanitarian and development efforts.

Mo Yu, Mark Dredze, Raman Arora et al.

Modern NLP models rely heavily on engineered features, which often combine word and contextual information into complex lexical features. Such combination results in large numbers of features, which can lead to over-fitting. We present a new model that represents complex lexical features---comprised of parts for words, contextual information and labels---in a tensor that captures conjunction information among these parts. We apply low-rank tensor approximations to the corresponding parameter tensors to reduce the parameter space and improve prediction speed. Furthermore, we investigate two methods for handling features that include $n$-grams of mixed lengths. Our model achieves state-of-the-art results on tasks in relation extraction, PP-attachment, and preposition disambiguation.