datasets
NormBank
Why NormBank?
NormBank is built to ground flexible normative reasoning for interactive, assistive, and collaborative AI systems. Unlike prior commonsense resources, NormBank grounds each inference within a multivalent sociocultural frame, which includes the setting (e.g., restaurant), the agents’ contingent roles (waiter, customer), their attributes (age, gender), and other physical, social, and cultural constraints (e.g., the temperature or the country of operation).
What’s ‘in the box?’
Constraints apply in different combinations to frame social norms. Under these manipulations, norms are non-monotonic — one can cancel an inference by updating its frame even slightly.
In total, NormBank contains 63k unique constraints from a taxonomy that includes the Setting (i.e., the loci of scripted social interactions, such as banks, classrooms, homes, hospitals) as well as…
-
Roles: setting-specific clusters of identity and responsibility (e.g., customer, server)
-
Attributes: properties of individual agents that determine their social norms (e.g., age, gender)
-
Environmental Factors: physical and situational signals that can trigger associative priming of social norms (e.g., time, country of origin, noise level, safety, privacy, cleanliness)
-
What can I do with this data?
NormBank is not designed for any particular narrow task; it is designed as a general-purpose knowledge resource that can ground social reasoning through downstream tasks. First, social scientists and engineers can NormBank as a static reference for discrete knowledge access about social norms that can inform downstream studies and design choices. Second, NLP researchers can use it to ground challenging NLP benchmark tasks. Third and most importantly, machine learning teams can incorporate NormBank into pre-training data for better transfer learning and downstream performance on social reasoning tasks.
🤗 Implicit Hate
Why Implicit Hate?
It is important to consider the subtle tricks that many extremists use to mask their threats and abuse. These more implicit forms of hate speech may easily go undetected by keyword detection systems, and even the most advanced architectures can fail if they have not been trained on implicit hate speech (Caselli et al. 2020).
What’s ‘in the box?’
This dataset contains 22,056 tweets from the most prominent extremist groups in the United States; 6,346 of these tweets contain implicit hate speech. We decompose the implicit hate class using the following taxonomy (distribution shown on the left).
-
Grievance: frustration over a minority group’s perceived privilege.
-
Incitement: implicitly promoting known hate groups and ideologies (e.g. by flaunting in-group power).
-
Inferiority: implying some group or person is of lesser value than another.
-
Irony: using sarcasm, humor, and satire to demean someone.
-
Stereotypes: associating a group with negative attribute using euphemisms, circumlocution, or metaphorical language.
-
Threats: making an indirect commitment to attack someone’s body, well-being, reputation, liberty, etc.
-
Other
Each of the 6,346 implicit hate tweets also has free-text annotations for target demographic group and an implied statement to describe the underlying message (see banner image above).
What can I do with this data?
State-of-the-art neural models may be able to learn from our data how to (1) classify this more difficult class of hate speech and (3) explain implicit hate by generating descriptions of both the target and the implied message. As our paper baselines show, neural models still have a ways to go, especially with classifying implicit hate categories, but overall, the results are promising, especially with implied statement generation, an admittedly challenging task.
We hope you can extend our baselines and further our efforts to understand and address some of these most pernicious forms of language that plague the web, especially among extremist groups.
🤗 MIC
Why MIC?
Open-domain or “chit-chat” conversational agents often reflect insensitive, hurtful, or contradictory viewpoints that erode a user’s trust in the integrity of the system. Moral integrity is one important pillar for building trust.
MIC is a dataset that can help us understand chatbot behaviors through their latent values and moral statements.
What’s ‘in the box?’
MIC contains 114k annotations, with 99k distinct “Rules of Thumb” (RoTs) that capture the moral assumptions of 38k chatbot replies to open-ended prompts. These RoTs represent diverse moral viewpoints, with the following distribution of underlying moral foundations:
-
Care: wanting someone or something to be safe, healthy, and happy. (58k chatbot replies)
-
Fairness: wanting to see individuals or groups treated equally or equitably. (24k)
-
Liberty: wanting people to be free to make their own decisions. (22k)
-
-
-
What can I do with this data?
We can train encoder-decoder models to automatically generate RoT explanations for chatbot behaviors. This could facilitate explainable downstream applications. For example, we could train RL systems that demote chatbot replies which fall into certain moral classes or train safety classifiers that guide systems towards the desired behaviors, with sensitivity towards ideological and political difference.
In the RoT generation task, we find our best models match the quality, fluency, and relevance of human annotations, but they still generate irrelevant RoTs nearly 28% of the time. This suggests that the proposed generation task is not yet solved and that MIC can continue to serve as resource for ongoing work in developing morally-consistent conversational agents.
🤗 Positive Reframing
Why Positive Frames?
This work was inspired by the need to escape the negative patterns of thinking that began to overwhelm the authors during the COVID-19 pandemic. We realized that what we needed was not some naive belief that everything would be okay if we ignored our problems. Instead, we needed reframing, or a shift in focus, with less weight on the negative things we can’t control, and more weight on the positive things about ourselves and our situation which we can control.
Positive reframing induces a complementary positive viewpoint (e.g. glass-half-full), which nevertheless supports the underlying content of the original sentence (see diagram above). The reframe implicates rather than contradicts the source, and the transformation is motivated by theoretically justified strategies from positive psychology (see What’s ‘in the box?’).
Our work shows how NLP can help lead the way by automatically reframing overly negative text using strategies from positive psychology.
What’s ‘in the box?’
The Positive Psychology Frames dataset contains 8,349 reframed sentence pairs, where the original sentence is drawn from a negative tweet (#stressed), and a reframed copy is provided by a crowdworker who was trained in the methods of positive psychology. Our positive psychology frames taxonomy is defined below (with the distribution of labels shown on the left).
-
Growth Mindset: Viewing a challenging event as an opportunity for the author specifically to grow or improve themselves.
-
-
-
Optimism: Focusing on things about the situation itself, in that moment, that are good (not just forecasting a better future).
-
Self-Affirmation: Talking about what strengths the author already has, or the values they admire, like love, courage, perseverance, etc.
-
Thankfulness: Expressing thankfulness or gratitude with key words like appreciate, glad that, thankful for, good thing, etc.
What can I do with this data?
State-of-the-art neural models can learn from our data how to (1) shift a negatively distorted text into a more positive perspective using a combination of strategies from positive psychology; and (2) recognize or classify the psychological strategies that are used to reframe a given source.
As our paper baselines show, neural models still have a long ways to go before they can reliably generate positive perspectives. We see particular errors from insubstantial changes, contradictions to the premise, self-contradictions, and hallucinations. Overall, our suggests that our dataset can serve as a useful benchmark for building natural language generation systems with positive perspectives.