skip to main content


Title: How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?
Text-to-image generative models have achieved unprecedented success in generating high-quality images based on natural language descriptions. However, it is shown that these models tend to favor specific social groups when prompted with neutral text descriptions (e.g., ‘a photo of a lawyer’). Following Zhao et al. (2021), we study the effect on the diversity of the generated images when adding ethical intervention that supports equitable judgment (e.g., ‘if all individuals can be a lawyer irrespective of their gender’) in the input prompts. To this end, we introduce an Ethical NaTural Language Interventions in Text-to-Image GENeration (ENTIGEN) benchmark dataset to evaluate the change in image generations conditional on ethical interventions across three social axes – gender, skin color, and culture. Through CLIP-based and human evaluation on minDALL.E, DALL.E-mini and Stable Diffusion, we find that the model generations cover diverse social groups while preserving the image quality. In some cases, the generations would be anti-stereotypical (e.g., models tend to create images with individuals that are perceived as man when fed with prompts about makeup) in the presence of ethical intervention. Preliminary studies indicate that a large change in the model predictions is triggered by certain phrases such as ‘irrespective of gender’ in the context of gender bias in the ethical interventions. We release code and annotated data at https://github.com/Hritikbansal/entigen_emnlp.  more » « less
Award ID(s):
1927554
NSF-PAR ID:
10391940
Author(s) / Creator(s):
; ; ;
Date Published:
Journal Name:
Empirical Methods in Natural Language Processing
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Introduction Social media has created opportunities for children to gather social support online (Blackwell et al., 2016; Gonzales, 2017; Jackson, Bailey, & Foucault Welles, 2018; Khasawneh, Rogers, Bertrand, Madathil, & Gramopadhye, 2019; Ponathil, Agnisarman, Khasawneh, Narasimha, & Madathil, 2017). However, social media also has the potential to expose children and adolescents to undesirable behaviors. Research showed that social media can be used to harass, discriminate (Fritz & Gonzales, 2018), dox (Wood, Rose, & Thompson, 2018), and socially disenfranchise children (Page, Wisniewski, Knijnenburg, & Namara, 2018). Other research proposes that social media use might be correlated to the significant increase in suicide rates and depressive symptoms among children and adolescents in the past ten years (Mitchell, Wells, Priebe, & Ybarra, 2014). Evidence based research suggests that suicidal and unwanted behaviors can be promulgated through social contagion effects, which model, normalize, and reinforce self-harming behavior (Hilton, 2017). These harmful behaviors and social contagion effects may occur more frequently through repetitive exposure and modelling via social media, especially when such content goes “viral” (Hilton, 2017). One example of viral self-harming behavior that has generated significant media attention is the Blue Whale Challenge (BWC). The hearsay about this challenge is that individuals at all ages are persuaded to participate in self-harm and eventually kill themselves (Mukhra, Baryah, Krishan, & Kanchan, 2017). Research is needed specifically concerning BWC ethical concerns, the effects the game may have on teenagers, and potential governmental interventions. To address this gap in the literature, the current study uses qualitative and content analysis research techniques to illustrate the risk of self-harm and suicide contagion through the portrayal of BWC on YouTube and Twitter Posts. The purpose of this study is to analyze the portrayal of BWC on YouTube and Twitter in order to identify the themes that are presented on YouTube and Twitter posts that share and discuss BWC. In addition, we want to explore to what extent are YouTube videos compliant with safe and effective suicide messaging guidelines proposed by the Suicide Prevention Resource Center (SPRC). Method Two social media websites were used to gather the data: 60 videos and 1,112 comments from YouTube and 150 posts from Twitter. The common themes of the YouTube videos, comments on those videos, and the Twitter posts were identified using grounded, thematic content analysis on the collected data (Padgett, 2001). Three codebooks were built, one for each type of data. The data for each site were analyzed, and the common themes were identified. A deductive coding analysis was conducted on the YouTube videos based on the nine SPRC safe and effective messaging guidelines (Suicide Prevention Resource Center, 2006). The analysis explored the number of videos that violated these guidelines and which guidelines were violated the most. The inter-rater reliabilities between the coders ranged from 0.61 – 0.81 based on Cohen’s kappa. Then the coders conducted consensus coding. Results & Findings Three common themes were identified among all the posts in the three social media platforms included in this study. The first theme included posts where social media users were trying to raise awareness and warning parents about this dangerous phenomenon in order to reduce the risk of any potential participation in BWC. This was the most common theme in the videos and posts. Additionally, the posts claimed that there are more than 100 people who have played BWC worldwide and provided detailed description of what each individual did while playing the game. These videos also described the tasks and different names of the game. Only few videos provided recommendations to teenagers who might be playing or thinking of playing the game and fewer videos mentioned that the provided statistics were not confirmed by reliable sources. The second theme included posts of people that either criticized the teenagers who participated in BWC or made fun of them for a couple of reasons: they agreed with the purpose of BWC of “cleaning the society of people with mental issues,” or they misunderstood why teenagers participate in these kind of challenges, such as thinking they mainly participate due to peer pressure or to “show off”. The last theme we identified was that most of these users tend to speak in detail about someone who already participated in BWC. These videos and posts provided information about their demographics and interviews with their parents or acquaintances, who also provide more details about the participant’s personal life. The evaluation of the videos based on the SPRC safe messaging guidelines showed that 37% of the YouTube videos met fewer than 3 of the 9 safe messaging guidelines. Around 50% of them met only 4 to 6 of the guidelines, while the remaining 13% met 7 or more of the guidelines. Discussion This study is the first to systematically investigate the quality, portrayal, and reach of BWC on social media. Based on our findings from the emerging themes and the evaluation of the SPRC safe messaging guidelines we suggest that these videos could contribute to the spread of these deadly challenges (or suicide in general since the game might be a hoax) instead of raising awareness. Our suggestion is parallel with similar studies conducted on the portrait of suicide in traditional media (Fekete & Macsai, 1990; Fekete & Schmidtke, 1995). Most posts on social media romanticized people who have died by following this challenge, and younger vulnerable teens may see the victims as role models, leading them to end their lives in the same way (Fekete & Schmidtke, 1995). The videos presented statistics about the number of suicides believed to be related to this challenge in a way that made suicide seem common (Cialdini, 2003). In addition, the videos presented extensive personal information about the people who have died by suicide while playing the BWC. These videos also provided detailed descriptions of the final task, including pictures of self-harm, material that may encourage vulnerable teens to consider ending their lives and provide them with methods on how to do so (Fekete & Macsai, 1990). On the other hand, these videos both failed to emphasize prevention by highlighting effective treatments for mental health problems and failed to encourage teenagers with mental health problems to seek help and providing information on where to find it. YouTube and Twitter are capable of influencing a large number of teenagers (Khasawneh, Ponathil, Firat Ozkan, & Chalil Madathil, 2018; Pater & Mynatt, 2017). We suggest that it is urgent to monitor social media posts related to BWC and similar self-harm challenges (e.g., the Momo Challenge). Additionally, the SPRC should properly educate social media users, particularly those with more influence (e.g., celebrities) on elements that boost negative contagion effects. While the veracity of these challenges is doubted by some, posting about the challenges in unsafe manners can contribute to contagion regardless of the challlenges’ true nature. 
    more » « less
  2. Artificial Intelligence (AI) systems for mental healthcare (MHCare) have been ever-growing after realizing the importance of early interventions for patients with chronic mental health (MH) conditions. Social media (SocMedia) emerged as the go-to platform for supporting patients seeking MHCare. The creation of peer-support groups without social stigma has resulted in patients transitioning from clinical settings to SocMedia supported interactions for quick help. Researchers started exploring SocMedia content in search of cues that showcase correlation or causation between different MH conditions to design better interventional strategies. User-level Classification-based AI systems were designed to leverage diverse SocMedia data from various MH conditions, to predict MH conditions. Subsequently, researchers created classification schemes to measure the severity of each MH condition. Such ad-hoc schemes, engineered features, and models not only require a large amount of data but fail to allow clinically acceptable and explainable reasoning over the outcomes. To improve Neural-AI for MHCare, infusion of clinical symbolic knowledge that clinicans use in decision making is required. An impactful use case of Neural-AI systems in MH is conversational systems. These systems require coordination between classification and generation to facilitate humanistic conversation in conversational agents (CA). Current CAs with deep language models lack factual correctness, medical relevance, and safety in their generations, which intertwine with unexplainable statistical classification techniques. This lecture-style tutorial will demonstrate our investigations into Neuro-symbolic methods of infusing clinical knowledge to improve the outcomes of Neural-AI systems to improve interventions for MHCare:(a) We will discuss the use of diverse clinical knowledge in creating specialized datasets to train Neural-AI systems effectively. (b) Patients with cardiovascular disease express MH symptoms differently based on gender differences. We will show that knowledge-infused Neural-AI systems can identify gender-specific MH symptoms in such patients. (c) We will describe strategies for infusing clinical process knowledge as heuristics and constraints to improve language models in generating relevant questions and responses. 
    more » « less
  3. Abstract

    There are currently >1.3 million human –omics samples that are publicly available. This valuable resource remains acutely underused because discovering particular samples from this ever-growing data collection remains a significant challenge. The major impediment is that sample attributes are routinely described using varied terminologies written in unstructured natural language. We propose a natural-language-processing-based machine learning approach (NLP-ML) to infer tissue and cell-type annotations for genomics samples based only on their free-text metadata. NLP-ML works by creating numerical representations of sample descriptions and using these representations as features in a supervised learning classifier that predicts tissue/cell-type terms. Our approach significantly outperforms an advanced graph-based reasoning annotation method (MetaSRA) and a baseline exact string matching method (TAGGER). Model similarities between related tissues demonstrate that NLP-ML models capture biologically-meaningful signals in text. Additionally, these models correctly classify tissue-associated biological processes and diseases based on their text descriptions alone. NLP-ML models are nearly as accurate as models based on gene-expression profiles in predicting sample tissue annotations but have the distinct capability to classify samples irrespective of the genomics experiment type based on their text metadata. Python NLP-ML prediction code and trained tissue models are available athttps://github.com/krishnanlab/txt2onto.

     
    more » « less
  4. null (Ed.)
    Interest in physical therapy and individual exercises such as yoga/dance has increased alongside the well-being trend, and people globally enjoy such exercises at home/office via video streaming platforms. However, such exercises are hard to follow without expert guidance. Even if experts can help, it is almost impossible to give personalized feedback to every trainee remotely. Thus, automated pose correction systems are required more than ever, and we introduce a new captioning dataset named FixMyPose to address this need. We collect natural language descriptions of correcting a “current” pose to look like a “target” pose. To support a multilingual setup, we collect descriptions in both English and Hindi. The collected descriptions have interesting linguistic properties such as egocentric relations to the environment objects, analogous references, etc., requiring an understanding of spatial relations and commonsense knowledge about postures. Further, to avoid ML biases, we maintain a balance across characters with diverse demographics, who perform a variety of movements in several interior environments (e.g., homes, offices). From our FixMyPose dataset, we introduce two tasks: the pose-correctional-captioning task and its reverse, the target-pose-retrieval task. During the correctional-captioning task, models must generate the descriptions of how to move from the current to the target pose image, whereas in the retrieval task, models should select the correct target pose given the initial pose and the correctional description. We present strong cross-attention baseline models (uni/multimodal, RL, multilingual) and also show that our baselines are competitive with other models when evaluated on other image-difference datasets. We also propose new task-specific metrics (object-match, body-part-match, direction-match) and conduct human evaluation for more reliable evaluation, and we demonstrate a large human-model performance gap suggesting room for promising future work. Finally, to verify the sim-to-real transfer of our FixMyPose dataset, we collect a set of real images and show promising performance on these images. Data and code are available: https://fixmypose-unc.github.io. 
    more » « less
  5. Abstract

    Classifying images using supervised machine learning (ML) relies on labeled training data—classes or text descriptions, for example, associated with each image. Data‐driven models are only as good as the data used for training, and this points to the importance of high‐quality labeled data for developing a ML model that has predictive skill. Labeling data is typically a time‐consuming, manual process. Here, we investigate the process of labeling data, with a specific focus on coastal aerial imagery captured in the wake of hurricanes that affected the Atlantic and Gulf Coasts of the United States. The imagery data set is a rich observational record of storm impacts and coastal change, but the imagery requires labeling to render that information accessible. We created an online interface that served labelers a stream of images and a fixed set of questions. A total of 1,600 images were labeled by at least two or as many as seven coastal scientists. We used the resulting data set to investigate interrater agreement: the extent to which labelers labeled each image similarly. Interrater agreement scores, assessed with percent agreement and Krippendorff's alpha, are higher when the questions posed to labelers are relatively simple, when the labelers are provided with a user manual, and when images are smaller. Experiments in interrater agreement point toward the benefit of multiple labelers for understanding the uncertainty in labeling data for machine learning research.

     
    more » « less