The Importance of Modeling Social Factors of Language: Theory and Practice

By Dirk Hovy, Diyi Yang. 2021

218 citations recorded by the source corpus.

Natural language processing (NLP) applications are now more powerful and ubiquitous than ever before. With rapidly developing (neural) models and ever-more available data, current NLP models have access to more information than any human speaker during their life. Still, it would be hard to argue that NLP models have reached human-level capacity. In this position paper, we argue that the reason for the current limitations is a focus on information content while ignoring language’s social factors. We show that current NLP systems systematically break down when faced with interpreting the social factors of language. This limits applications to a subset of information-related tasks and prevents NLP from reaching human-level performance. At the same time, systems that incorporate even a minimum of social factors already show remarkable improvements. We formalize a taxonomy of seven social factors based on linguistic theory and exemplify current failures and emerging successes for each of them. We suggest that the NLP community address social factors to get closer to the goal of human-like language understanding.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling (2021, background): “Some of the issues detailed in this paper may be attributed to a lack of language understanding, especially the social meaning of language (Hovy & Spruit, 2016; Flek, 2020; Hovy & Yang, 2021; Nguyen et al., 2021).”
Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LMs (2022, background): “learning of Theory of Mind abilities (Bender and Koller, 2020; Bisk et al., 2020; Hovy and Yang, 2021), similar to how visual groundings can help with learning physical knowledge (Zhang et al.”
HONEST: Measuring Hurtful Sentence Completion in Language Models (2021, background): “More generally, language models generating taboo words and insults is the result of NLP systems not incorporating social norms (Hovy and Yang, 2021).”
Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection (2021, background): “However, given this subjectivity, the open question remains: whose perspective should be considered when using toxicity detection models? To try answering this question, we urge researchers and practitioners to consider all stakeholders and end users on which toxicity detection systems might be depl”
Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models (2022, background): “Finally, every application is shaped by a social context [60, 77, 92, 53], which includes a range of factors such as language and cultural norms in the community using a system.”
What Does it Mean for a Language Model to Preserve Privacy? (2022, background): “Although this task has advanced in the past few years, the proposed models are plagued by the same challenges as the chatbots and the problem is far from solved [41].”
Taxonomy of Risks posed by Language Models (2022, background): “ther complicated by the context dependency of what constitutes profoundly offensive speech [80, 92, 105].”
SafetyKit: First Aid for Measuring Safety in Open-domain Conversational Systems (2022, background): “As the very notion of offensiveness is highly dependent on social context (Hovy and Yang, 2021), this will be insufficient for measuring the appropriateness of a model’s responses in other dialects, cultures, and languages (Schmidt and Wiegand, 2017).”
Challenges and Strategies in Cross-Cultural NLP (2022, background): “However, the impact of social and cultural factors on linguistic form and stylistic variations is rarely discussed (Hovy and Yang, 2021).”
Five sources of bias in natural language processing (2021, background): “And language carries a lot of secondary information about the speaker, their self-identification, and membership in socio-demographic groups (Flek, 2020; Hovy & Spruit, 2016; Hovy & Yang, 2021).”
When Combating Hype, Proceed with Caution (2021, result): “Similar claims can be found in many other citing works (Utama et al., 2020; van Noord et al., 2020; Hovy and Yang, 2021; Sayers et al., 2021; PetiStantić et al., 2021; Jang and Lukasiewicz, 2021).”
Towards Measuring the Representation of Subjective Global Opinions in Language Models (2023, unknown): “While we believe that it is important to consider social contexts when developing AI systems (Hovy & Yang, 2021; Tamkin et al., 2021), we do not make definitive claims about ideal levels of cultural representation.”
Theory-Grounded Measurement of U.S. Social Stereotypes in English Language Models (2022, background): “relations between social groups and traits, like associating the group “Muslim” with the trait threatening, or “man” with confident (e.g., Bender et al., 2021; Nozza et al., 2021; Hovy and Yang, 2021).”
Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection (2022, background): “Because natural language varies with social and demographic variables (Rickford, 1985; Eckert, 1989; Labov, 2006; Blodgett et al., 2016; Hovy and Yang, 2021; Lucy and Bamman, 2021, inter alia ), we can also ask whose language will be excluded.”
Cultural Incongruencies in Artificial Intelligence (2022, background): “…tasks are inherently shaped by the culture within which they are embedded in, including not only the cultural-linguistic dependencies [16, 17] of tasks such as inferring emotion, sentiment, offensiveness, but also image and symbol interpretation—including gestures, facial expressions,…”