Sharon Levy

Sharon Levy, Emily Allaway, Melanie Subbiah et al.

Understanding what constitutes safe text is an important issue in natural language processing and can often prevent the deployment of models deemed harmful and unsafe. One such type of safety that has been scarcely studied is commonsense physical safety, i.e. text that is not explicitly violent and requires additional commonsense knowledge to comprehend that it leads to physical harm. We create the first benchmark dataset, SafeText, comprising real-life scenarios with paired safe and physically unsafe pieces of advice. We utilize SafeText to empirically study commonsense physical safety across various models designed for text generation and commonsense reasoning tasks. We find that state-of-the-art large language models are susceptible to the generation of unsafe text and have difficulty rejecting unsafe advice. As a result, we argue for further studies of safety and the assessment of commonsense physical safety in models before release.

Matthew Ho, Aditya Sharma, Justin Chang et al.

As large language models (LLMs) grow larger and more sophisticated, assessing their "reasoning" capabilities in natural language grows more challenging. Recent question answering (QA) benchmarks that attempt to assess reasoning are often limited by a narrow scope of covered situations and subject matters. We introduce WikiWhy, a QA dataset built around a novel auxiliary task: explaining why an answer is true in natural language. WikiWhy contains over 9,000 "why" question-answer-rationale triples, grounded on Wikipedia facts across a diverse set of topics. Each rationale is a set of supporting statements connecting the question to the answer. WikiWhy serves as a benchmark for the reasoning capabilities of LLMs because it demands rigorous explicit rationales for each answer to demonstrate the acquisition of implicit commonsense knowledge, which is unlikely to be easily memorized. GPT-3 baselines achieve only 38.7% human-evaluated correctness in the end-to-end answer & explain condition, leaving significant room for future improvements.

Alex Mei, Anisha Kabir, Sharon Levy et al.

An increasingly prevalent problem for intelligent technologies is text safety, as uncontrolled systems may generate recommendations to their users that lead to injury or life-threatening consequences. However, the degree of explicitness of a generated statement that can cause physical harm varies. In this paper, we distinguish types of text that can lead to physical harm and establish one particularly underexplored category: covertly unsafe text. Then, we further break down this category with respect to the system's information and discuss solutions to mitigate the generation of text in each of these subcategories. Ultimately, our work defines the problem of covertly unsafe language that causes physical harm and argues that this subtle yet dangerous issue needs to be prioritized by stakeholders and regulators. We highlight mitigation strategies to inspire future researchers to tackle this challenging problem and help improve safety within smart systems.

Samhita Honnavalli, Aesha Parekh, Lily Ou et al.

Women are often perceived as junior to their male counterparts, even within the same job titles. While there has been significant progress in the evaluation of gender bias in natural language processing (NLP), existing studies seldom investigate how biases toward gender groups change when compounded with other societal biases. In this work, we investigate how seniority impacts the degree of gender bias exhibited in pretrained neural generation models by introducing a novel framework for probing compound bias. We contribute a benchmark robustness-testing dataset spanning two domains, U.S. senatorship and professorship, created using a distant-supervision method. Our dataset includes human-written text with underlying ground truth and paired counterfactuals. We then examine GPT-2 perplexity and the frequency of gendered language in generated text. Our results show that GPT-2 amplifies bias by considering women as junior and men as senior more often than the ground truth in both domains. These results suggest that NLP applications built using GPT-2 may harm women in professional capacities.

Kai Nakamura, Sharon Levy, Yi-Lin Tuan et al.

A pressing challenge in current dialogue systems is to successfully converse with users on topics with information distributed across different modalities. Previous work in multiturn dialogue systems has primarily focused on either text or table information. In more realistic scenarios, having a joint understanding of both is critical as knowledge is typically distributed over both unstructured and structured forms. We present a new dialogue dataset, HybriDialogue, which consists of crowdsourced natural conversations grounded on both Wikipedia text and tables. The conversations are created through the decomposition of complex multihop questions into simple, realistic multiturn dialogue interactions. We propose retrieval, system state tracking, and dialogue response generation tasks for our dataset and conduct baseline experiments for each. Our results show that there is still ample opportunity for improvement, demonstrating the importance of building stronger dialogue systems that can reason over the complex setting of information-seeking dialogue grounded on tables and text.

Sharon Levy, Neha Anna John, Ling Liu et al.

Studies in bias and fairness in natural language processing have primarily examined social biases within a single language and/or across few attributes (e.g. gender, race). However, biases can manifest differently across various languages for individual attributes. As a result, it is critical to examine biases within each language and attribute. Of equal importance is to study how these biases compare across languages and how the biases are affected when training a model on multilingual data versus monolingual data. We present a bias analysis across Italian, Chinese, English, Hebrew, and Spanish on the downstream sentiment analysis task to observe whether specific demographics are viewed more positively. We study bias similarities and differences across these languages and investigate the impact of multilingual vs. monolingual training data. We adapt existing sentiment bias templates in English to Italian, Chinese, Hebrew, and Spanish for four attributes: race, religion, nationality, and gender. Our results reveal similarities in bias expression such as favoritism of groups that are dominant in each language's culture (e.g. majority religions and nationalities). Additionally, we find an increased variation in predictions across protected groups, indicating bias amplification, after multilingual finetuning in comparison to multilingual pretraining.

Alon Albalak, Sharon Levy, William Yang Wang

Open-retrieval question answering systems are generally trained and tested on large datasets in well-established domains. However, low-resource settings such as new and emerging domains would especially benefit from reliable question answering systems. Furthermore, multilingual and cross-lingual resources in emergent domains are scarce, leading to few or no such systems. In this paper, we demonstrate a cross-lingual open-retrieval question answering system for the emergent domain of COVID-19. Our system adopts a corpus of scientific articles to ensure that retrieved documents are reliable. To address the scarcity of cross-lingual training data in emergent domains, we present a method utilizing automatic translation, alignment, and filtering to produce English-to-all datasets. We show that a deep semantic retriever greatly benefits from training on our English-to-all data and significantly outperforms a BM25 baseline in the cross-lingual setting. We illustrate the capabilities of our system with examples and release all code necessary to train and deploy such a system.

Sophie Groenwold, Lily Ou, Aesha Parekh et al.

The growth of social media has encouraged the written use of African American Vernacular English (AAVE), which has traditionally been used only in oral contexts. However, NLP models have historically been developed using dominant English varieties, such as Standard American English (SAE), due to text corpora availability. We investigate the performance of GPT-2 on AAVE text by creating a dataset of intent-equivalent parallel AAVE/SAE tweet pairs, thereby isolating syntactic structure and AAVE- or SAE-specific language for each pair. We evaluate each sample and its GPT-2 generated text with pretrained sentiment classifiers and find that while AAVE text results in more classifications of negative sentiment than SAE, the use of GPT-2 generally increases occurrences of positive sentiment for both. Additionally, we conduct human evaluation of AAVE and SAE text generated with GPT-2 to compare contextual rigor and overall quality.