LEAF: Predicting the Environmental Impact of Food Products based on their Name

Bas KrahmerACL 2024

TL;DR

This paper introduces LEAF, a novel NLP model that predicts the environmental impact of food products using only their names. LEAF significantly outperforms zero-shot classification methods, offering a scalable solution to a critical information gap for consumers and policymakers.

What Problem Does This Paper Address?

Globally, food consumption is a major contributor to greenhouse gas emissions, accounting for up to 30% of the total. While Life Cycle Assessments (LCAs) can quantify this impact, this information is rarely available to consumers at the point of purchase, hindering informed decision-making. Existing methods for estimating environmental impact are often complex or unavailable for everyday products. This paper tackles the challenge of making environmental impact data accessible by leveraging natural language processing to predict it from product names alone.

Methods

The core of the LEAF system is a set of models that use pretrained sentence embedding models (specifically, distiluse-multilingual-base-v2 and bge-m3) to generate embeddings for product names. These embeddings are then fed into task-specific "readout heads." The authors explore three configurations:

  • LEAFC: A classification head that predicts one of 2518 Agribalyse classes, using cross-entropy loss.
  • LEAFr: A regression head that predicts a continuous environmental footprint (EF) score, using mean-squared error loss.
  • LEAFh: A hybrid head that combines both classification and regression losses, allowing for simultaneous optimization.

Crucially, the base embedding models are frozen, and only the readout heads are trained. This approach leverages the powerful multilingual representations of the base models without requiring extensive fine-tuning.

Datasets

The primary dataset used is Open Food Facts (OFF). This is a large, crowd-sourced, open-access database containing information on food products. The authors filtered this dataset to include only products with an associated Agribalyse class, resulting in 2518 classes and 2518 Agribalyse classes with associated LCA data. The EF score, a key metric, is derived from this LCA data.

Key Insights and Ideas

  • Product names are surprisingly informative: The study demonstrates that a product's name alone contains enough signal to predict its environmental impact with high accuracy, even across multiple languages.
  • Classification outperforms regression for this task: While regression models predict a continuous EF score, classification models that map to discrete environmental impact categories (like Eco-Score) achieve higher accuracy and lower error.
  • Multilingual capabilities are essential: Given the diverse origins of food products and the global nature of consumption, using multilingual embedding models is key to LEAF's success.
  • Large language models (LLMs) can be powerful baselines, but are not always superior: While GPT-3.5 showed promise, LEAF models, despite having significantly fewer parameters, achieved better performance in terms of accuracy and error, especially in a zero-shot setting.

Key Results

The LEAF models demonstrated strong performance:

  • The LEAFC model (classification head) achieved an accuracy of 0.731 and a Mean Absolute Error (MAE) of 0.071 on the test set.
  • The LEAFC, M3 configuration (using the larger bge-m3 embedding model) further improved accuracy to 0.772 with an MAE of 0.057.
  • LEAF models significantly outperformed the OpenAI GPT-3.5 baseline, which achieved an accuracy of 0.374 (and a corrected accuracy of 0.302 when accounting for hallucinations).
  • Multilingual analysis showed LEAF performing best across the top 5 most frequent languages in the dataset.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Class Specificity: The current Agribalyse classes abstract away fine-grained differences (e.g., local vs. imported apples), which can significantly impact environmental footprint.
  • Fixed Consumption Location: Models assume consumption in France, which may not reflect global supply chains and transportation impacts.
  • Data Granularity: Redundancy in LCA data for similar product variations (e.g., different types of almonds) could be addressed by creating more specific class distributions.

Future work could involve:

  • Incorporating more fine-grained LCA data or exploring EF score explainability.
  • Accounting for varying consumption locations.
  • Integrating additional data sources like ingredient lists, country of origin, and transportation methods.
  • Developing models that handle uncertainty and combine diverse data sources.

Summary updated

BibTeX
@inproceedings{krahmer2024leaf,
  author = {Bas Krahmer},
  title = {{LEAF: Predicting the Environmental Impact of Food Products based on their Name}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure