Abstract
As the Natural Language Processing (NLP) community grows, so does its environmental footprint. While much attention has been paid to the energy consumption of training large-scale models, the "human cost" of international travel remains under-explored. This study analyzes over 60,000 papers from the ACL Anthology (1965–2020) to quantify the trade-off between carbon emissions and geographic diversity. We find that while total emissions have reached 6 million kg of $CO_2$ annually, in-person conferences provide a unique "regional boost" to diversity that is not easily replicated by journals or virtual formats.
Introduction: The Travel-Diversity Tension
The NLP community faces a difficult trade-off: international conferences are criticized for high emissions, yet they serve as the primary engine for networking and visibility. Moving entirely to virtual formats risks harming the diversity of the field by silencing researchers from underrepresented regions. This research moves the debate from anecdotal evidence to quantitative data by mapping the carbon cost of conference travel against participation diversity over five decades.
Methodology
The authors developed a custom NLP pipeline to transform the ACL Anthology into a geospatial dataset:
- Text Mining & Information Extraction: XML data and PDFs for 60,572 publications were parsed. A RoBERTa-based Named Entity Recognition (NER) pipeline identified author affiliations and addresses.
- Geocoding & Carbon Modeling: Locations were converted to coordinates via the Google Maps API. Historical flight emissions were estimated using a linear regression model trained on UK government GHG data, accounting for "Radial Forcing" (non-$CO_2$ climate impacts at high altitudes).
- Diversity Metric: The Gini coefficient was applied to measure the concentration of research participation across different countries.
Key Insights: The Efficiency Paradox and Regional Boost
The Scale of Travel
Emissions have grown exponentially, from negligible amounts in the 1970s to over 6 million kg of $CO_2$ in 2018. In that year alone, NLP conference travel emitted roughly 6,000 tonnes of $CO_2e$—equivalent to training 71 instances of GPT-3 from scratch. Despite aircraft becoming more fuel-efficient, the average carbon cost per paper has remained stable (0.8–1.2 tonnes of $CO_2$). This "efficiency paradox" occurs because technological gains are offset by researchers traveling significantly longer distances.
The "Regional Boost"
In-person conferences act as a powerful magnet for local talent. When a conference is held in a specific region, participation from that region spikes. Interestingly, the study found that non-physical venues (journals) often have lower diversity scores than many in-person conferences, suggesting that the social "pull" of a physical event is a key driver for global inclusion.
Results and Diversity Trends
- Diversity Decline: Geographic diversity in NLP grew steadily from 1970 to 2014 but has trended downward since 2015. Research is becoming increasingly concentrated in a few dominant hubs, primarily the US and China.
- Strategic Location Savings: By choosing conference locations strategically, the community has saved approximately 27% of potential emissions compared to random placement.
- Local vs. Global: Localized conferences (e.g., TALN, NoDaLiDa) emit only about 25% of the carbon per paper compared to flagship international events while maintaining high regional relevance.
Limitations and Future Outlook
The study assumes only the first author travels and that they depart from their institutional address, likely underestimating total emissions. Furthermore, diversity is defined purely by geography, excluding critical factors like gender, race, or socioeconomic status. As the field moves toward hybrid models post-pandemic, future work must determine if "virtual diversity" can truly replace the unique benefits of in-person regional engagement.
Summary updated