TL;DR
This paper maps the landscape of Natural Language Processing for Social Good (NLP4SG), defining the field, analyzing its current research trends, and identifying under-explored areas. It provides a crucial overview for researchers and practitioners aiming to leverage NLP for positive societal impact.
What Problem Does This Paper Address?
The field of Natural Language Processing for Social Good (NLP4SG) is rapidly growing, encompassing a wide range of applications from combating hate speech to aiding individuals with learning disabilities. However, there's a lack of clarity regarding what constitutes NLP4SG, which areas are being addressed, who the key players are, and importantly, what "white spots" or neglected areas exist. This paper aims to fill this gap by providing a structured definition, analyzing existing research, and highlighting opportunities for future work to ensure NLP development genuinely contributes to societal well-being.
Methods
The authors propose a working definition of NLP4SG by adapting existing AI for Social Good (AI4SG) definitions. They then analyze a large corpus of NLP research papers (over 50,000 from the ACL Anthology and Google Scholar) to map the field. Their methodology involves:
- Keyword Matching: Compiling a comprehensive list of keywords related to social good areas (e.g., agriculture, healthcare, education) and other dimensions (ethics, collaboration).
- Automatic Annotation: Using these keywords to automatically tag papers as "explicit NLP4SG" if they mention these terms.
- Manual Validation: A meta-annotator validated the automatic tagging, achieving 95% agreement.
- Analysis of Trends: Examining the temporal evolution of NLP4SG publications, identifying key research areas, and comparing "explicit NLP4SG" papers with "non-explicit NLP4SG" papers in terms of authorship, citations, and publication venues.
Datasets
The primary dataset used is the NLP Scholar Dataset, which contains over 50,000 instances from the ACL Anthology and Google Scholar. This dataset was enriched with paper abstracts. The authors also reference keywords from existing works like Shi et al. (2020) and the UN Sustainable Development Goals.
Key Insights and Ideas
- Defining NLP4SG is Crucial: A clear definition is essential to guide research and measure impact, moving beyond "social good" as an accidental byproduct.
- Healthcare Dominates: Healthcare is the most frequently addressed area within NLP4SG research.
- Ethics is Key, Collaboration is Lagging: While ethical considerations are increasingly present, research on "systems of change and collaboration" is significantly less prominent.
- Explicit vs. Implicit NLP4SG: The study distinguishes between papers that explicitly mention NLP4SG terms and those that contribute to social good more implicitly, suggesting the latter might be underestimated.
- "Explicit NLP4SG" Papers Have Different Characteristics: These papers tend to have more authors but fewer citations and are often published in less prestigious venues compared to general NLP papers.
- Under-Researched Areas Exist: Areas like language disorders, environmental sustainability, urban planning, transportation, and public safety are identified as having less NLP4SG research.
Key Results
- Growth of NLP4SG: The percentage of NLP publications related to social good has been steadily increasing, particularly in the last decade. By 2020, over 20% of publications mentioned social good-related terms.
- Area Distribution: Healthcare, social care, media corrupted communication, and education are the most popular areas. Public safety, transportation, urban planning, environmental sustainability, and language disorders are less explored.
- Other Dimensions: Ethics and general social good are frequently mentioned, while collaboration is less so.
- Explicit NLP4SG Papers: These constitute 9.63% of the total NLP papers analyzed. They tend to have more authors (3.65 vs. 2.97) but fewer citations (25.87 vs. 42.03) compared to non-explicit NLP4SG papers.
- Venue Distribution: Explicit NLP4SG papers are more prevalent in workshops, shared tasks, and student research venues, with a decreasing presence in top-tier conferences and journals.
Limitations and Future Work
- Keyword Matching Limitations: The "explicit NLP4SG" tagging relies on keyword matching, which might miss papers that contribute to social good in more subtle ways. This means the results represent a "lower bound baseline."
- Eurocentric Perspective: The analysis primarily focuses on ACL publications, potentially leading to a Eurocentric view and missing contributions from other regions or publication venues.
- Defining "Positive Impact": What constitutes a "positive impact" is context-dependent and can vary across cultures and values.
- Future Work Suggestions:
- Develop more systematic methods for identifying NLP4SG research.
- Encourage a broader discussion on social good dimensions like ethics and collaboration.
- Investigate how NLP applications involve end-users and incorporate contextual knowledge.
- Promote the formation of an interconnected social good community across disciplines.
- Provide data statements and terms of use for developed technologies to prevent misuse.
Summary updated