Emily Sheng

Ahmed Magooda, Alec Helyar, Kyle Jackson et al. · microsoft-research

We present a framework for the automated measurement of responsible AI (RAI) metrics for large language models (LLMs) and associated products and services. Our framework for automatically measuring harms from LLMs builds on existing technical and sociotechnical expertise and leverages the capabilities of state-of-the-art LLMs, such as GPT-4. We use this framework to run through several case studies investigating how different LLMs may violate a range of RAI-related principles. The framework may be employed alongside domain-specific sociotechnical expertise to create measurements for new harm areas in the future. By implementing this framework, we aim to enable more advanced harm measurement efforts and further the responsible use of LLMs.

Dhruv Agarwal, Emily Sheng, Chad Atalla et al.

Evaluating generative AI (GenAI) systems is challenging because many targets of evaluation are broad, contested concepts, such as "reasoning," "fairness," or "creativity." When these concepts are left underspecified, it becomes unclear what should be measured or how evaluation results should be interpreted. This problem reflects a missing step: systematization, that is, moving from a broad background concept to an explicit, structured account of the concept in measurable terms. To help address the fact that systematization is cognitively demanding and resource-intensive, we investigate whether AI assistance can support this process. To enable AI-assisted systematization and assess its quality, we introduce a structured representation of a systematized concept, a concept spec, and a validation worksheet. We then develop two AI-assisted systematizers: a direct, zero-shot approach and a multi-agent approach that more closely mirrors manual systematization approaches from existing literature. We use these systematizers to produce concept specs for two concepts -- hate-based rhetoric and digital empathy -- and evaluate resulting concept specs on content validity and information recoverability.

Kyra Yee, Alice Schoenauer Sebag, Olivia Redfield et al.

Harmful content detection models tend to have higher false positive rates for content from marginalized groups. In the context of marginal abuse modeling on Twitter, such disproportionate penalization poses the risk of reduced visibility, where marginalized communities lose the opportunity to voice their opinion on the platform. Current approaches to algorithmic harm mitigation, and bias detection for NLP models are often very ad hoc and subject to human bias. We make two main contributions in this paper. First, we design a novel methodology, which provides a principled approach to detecting and measuring the severity of potential harms associated with a text-based model. Second, we apply our methodology to audit Twitter's English marginal abuse model, which is used for removing amplification eligibility of marginally abusive content. Without utilizing demographic labels or dialect classifiers, we are still able to detect and measure the severity of issues related to the over-penalization of the speech of marginalized communities, such as the use of reclaimed speech, counterspeech, and identity related terms. In order to mitigate the associated harms, we experiment with adding additional true negative examples and find that doing so provides improvements to our fairness metrics without large degradations in model performance.

Emily Sheng, Josh Arnold, Zhou Yu et al.

Dialogue systems in the form of chatbots and personal assistants are being increasingly integrated into people's lives. Modern dialogue systems may consider adopting anthropomorphic personas, mimicking societal demographic groups to appear more approachable and trustworthy to users. However, the adoption of a persona can result in the adoption of biases. In this paper, we present the first large-scale study on persona biases in dialogue systems and conduct analyses on personas of different social classes, sexual orientations, races, and genders. We define persona biases as harmful differences in responses (e.g., varying levels of offensiveness, agreement with harmful statements) generated from adopting different demographic personas. Furthermore, we introduce an open-source framework, UnitPersonaBias, to explore and aggregate persona biases in dialogue systems. By analyzing the Blender and DialoGPT dialogue systems, we observe that adopting personas can actually decrease harmful responses, compared to not using any personas. Additionally, we find that persona choices can affect the degree of harms in generated responses and thus should be systematically evaluated before deployment. We also analyze how personas can result in different amounts of harm towards specific demographics.

Emma Harvey, Emily Sheng, Su Lin Blodgett et al. · microsoft-research

The NLP research community has made publicly available numerous instruments for measuring representational harms caused by large language model (LLM)-based systems. These instruments have taken the form of datasets, metrics, tools, and more. In this paper, we examine the extent to which such instruments meet the needs of practitioners tasked with evaluating LLM-based systems. Via semi-structured interviews with 12 such practitioners, we find that practitioners are often unable to use publicly available instruments for measuring representational harms. We identify two types of challenges. In some cases, instruments are not useful because they do not meaningfully measure what practitioners seek to measure or are otherwise misaligned with practitioner needs. In other cases, instruments - even useful instruments - are not used by practitioners due to practical and institutional barriers impeding their uptake. Drawing on measurement theory and pragmatic measurement, we provide recommendations for addressing these challenges to better meet practitioner needs.

Emily Corvi, Hannah Washington, Stefanie Reed et al.

Representational harms are widely recognized among fairness-related harms caused by generative language systems. However, their definitions are commonly under-specified. We make a theoretical contribution to the specification of representational harms by introducing a framework, grounded in speech act theory (Austin, 1962), that conceptualizes representational harms caused by generative language systems as the perlocutionary effects (i.e., real-world impacts) of particular types of illocutionary acts (i.e., system behaviors). Building on this argument and drawing on relevant literature from linguistic anthropology and sociolinguistics, we provide new definitions of stereotyping, demeaning, and erasure. We then use our framework to develop a granular taxonomy of illocutionary acts that cause representational harms, going beyond the high-level taxonomies presented in previous work. We also discuss the ways that our framework and taxonomy can support the development of valid measurement instruments. Finally, we demonstrate the utility of our framework and taxonomy via a case study that engages with recent conceptual debates about what constitutes a representational harm and how such harms should be measured.

Sunipa Dev, Emily Sheng, Jieyu Zhao et al.

Recent studies show that Natural Language Processing (NLP) technologies propagate societal biases about demographic groups associated with attributes such as gender, race, and nationality. To create interventions and mitigate these biases and associated harms, it is vital to be able to detect and measure such biases. While existing works propose bias evaluation and mitigation methods for various tasks, there remains a need to cohesively understand the biases and the specific harms they measure, and how different measures compare with each other. To address this gap, this work presents a practical framework of harms and a series of questions that practitioners can answer to guide the development of bias measures. As a validation of our framework and documentation questions, we also present several case studies of how existing bias measures in NLP -- both intrinsic measures of bias in representations and extrinsic measures of bias of downstream applications -- can be aligned with different harms and how our proposed documentation questions facilitates more holistic understanding of what bias measures are measuring.

Emily Sheng, Kai-Wei Chang, Premkumar Natarajan et al.

Technology for language generation has advanced rapidly, spurred by advancements in pre-training large models on massive amounts of data and the need for intelligent agents to communicate in a natural manner. While techniques can effectively generate fluent text, they can also produce undesirable societal biases that can have a disproportionately negative impact on marginalized populations. Language generation presents unique challenges for biases in terms of direct user interaction and the structure of decoding techniques. To better understand these challenges, we present a survey on societal biases in language generation, focusing on how data and techniques contribute to biases and progress towards reducing biases. Motivated by a lack of studies on biases from decoding techniques, we also conduct experiments to quantify the effects of these techniques. By further discussing general trends and open challenges, we call to attention promising directions for research and the importance of fairness and inclusivity considerations for language generation applications.

Emily Sheng, David Uthus

There is a growing collection of work analyzing and mitigating societal biases in language understanding, generation, and retrieval tasks, though examining biases in creative tasks remains underexplored. Creative language applications are meant for direct interaction with users, so it is important to quantify and mitigate societal biases in these applications. We introduce a novel study on a pipeline to mitigate societal biases when retrieving next verse suggestions in a poetry composition system. Our results suggest that data augmentation through sentiment style transfer has potential for mitigating societal biases.

Emily Sheng, Kai-Wei Chang, Premkumar Natarajan et al.

Ad hominem attacks are those that target some feature of a person's character instead of the position the person is maintaining. These attacks are harmful because they propagate implicit biases and diminish a person's credibility. Since dialogue systems respond directly to user input, it is important to study ad hominems in dialogue responses. To this end, we propose categories of ad hominems, compose an annotated dataset, and build a classifier to analyze human and dialogue system responses to English Twitter posts. We specifically compare responses to Twitter topics about marginalized communities (#BlackLivesMatter, #MeToo) versus other topics (#Vegan, #WFH), because the abusive language of ad hominems could further amplify the skew of power away from marginalized populations. Furthermore, we propose a constrained decoding technique that uses salient $n$-gram similarity as a soft constraint for top-$k$ sampling to reduce the amount of ad hominems generated. Our results indicate that 1) responses from both humans and DialoGPT contain more ad hominems for discussions around marginalized communities, 2) different quantities of ad hominems in the training data can influence the likelihood of generating ad hominems, and 3) we can use constrained decoding techniques to reduce ad hominems in generated dialogue responses.

Emily Sheng, Kai-Wei Chang, Premkumar Natarajan et al.

We present a general approach towards controllable societal biases in natural language generation (NLG). Building upon the idea of adversarial triggers, we develop a method to induce societal biases in generated text when input prompts contain mentions of specific demographic groups. We then analyze two scenarios: 1) inducing negative biases for one demographic and positive biases for another demographic, and 2) equalizing biases between demographics. The former scenario enables us to detect the types of biases present in the model. Specifically, we show the effectiveness of our approach at facilitating bias analysis by finding topics that correspond to demographic inequalities in generated text and comparing the relative effectiveness of inducing biases for different demographics. The second scenario is useful for mitigating biases in downstream applications such as dialogue generation. In our experiments, the mitigation technique proves to be effective at equalizing the amount of biases across demographics while simultaneously generating less negatively biased text overall.

Emily Sheng, Kai-Wei Chang, Premkumar Natarajan et al.

We present a systematic study of biases in natural language generation (NLG) by analyzing text generated from prompts that contain mentions of different demographic groups. In this work, we introduce the notion of the regard towards a demographic, use the varying levels of regard towards different demographics as a defining metric for bias in NLG, and analyze the extent to which sentiment scores are a relevant proxy metric for regard. To this end, we collect strategically-generated text from language models and manually annotate the text with both sentiment and regard scores. Additionally, we build an automatic regard classifier through transfer learning, so that we can analyze biases in unseen text. Together, these methods reveal the extent of the biased nature of language model generations. Our analysis provides a study of biases in NLG, bias metrics and correlated human judgments, and empirical evidence on the usefulness of our annotated dataset.

Emily Sheng, Prem Natarajan

In biomedical literature, it is common for entity boundaries to not align with word boundaries. Therefore, effective identification of entity spans requires approaches capable of considering tokens that are smaller than words. We introduce a novel, subword approach for named entity recognition (NER) that uses byte-pair encodings (BPE) in combination with convolutional and recurrent neural networks to produce byte-level tags of entities. We present experimental results on several standard biomedical datasets, namely the BioCreative VI Bio-ID, JNLPBA, and GENETAG datasets. We demonstrate competitive performance while bypassing the specialized domain expertise needed to create biomedical text tokenization rules.

Emily Sheng, Prem Natarajan, Jonathan Gordon et al.

Characterizing the content of a technical document in terms of its learning utility can be useful for applications related to education, such as generating reading lists from large collections of documents. We refer to this learning utility as the "pedagogical value" of the document to the learner. While pedagogical value is an important concept that has been studied extensively within the education domain, there has been little work exploring it from a computational, i.e., natural language processing (NLP), perspective. To allow a computational exploration of this concept, we introduce the notion of "pedagogical roles" of documents (e.g., Tutorial and Survey) as an intermediary component for the study of pedagogical value. Given the lack of available corpora for our exploration, we create the first annotated corpus of pedagogical roles and use it to test baseline techniques for automatic prediction of such roles.