E-MIMIC promotes the values of respect, openness, and equity
fostering a strong collaboration between linguists and data scientists to develop AI tools that bridge the gap between language and technology
Why
→ Diversity shapes our world — and inclusive communication gives space for every story to be shared.
→ By leveraging generative AI, we can support the widespread adoption of inclusive communication practices across different platforms, formats, and audiences.
Research objective 1
Detect and rewrite non-inclusive language in formal Italian communications using AI-powered models.
Research objective 2
Build expert-annotated datasets and inclusive language guidelines to train and evaluate NLP tools.
Research objective 3
Develop an accessible writing assistant tool to support inclusive communication in public and administrative contexts.
Inclusively: Communicating Without Excluding
Formal communications such as public calls, announcements, or regulations are supposed to exhibit respect for diversity in terms of gender, race, age, and disability. However, human writers often lack adequate inclusive writing skills. For instance, they tend to overuse the masculine as a neutral form, mainly because they are self-trained on biased text examples.
To overcome this issue, we propose to leverage Generative Artificial Intelligence to support inclusive language writing. Focusing on formal Italian communications, we have designed and developed an AI-assisted tool for non-inclusive text detection and reformulation.
Natural Language Processing algorithms designed with continuous human oversight and feedback
NLP you can trust – driven by human insight, refined by feedback
Our Work
From Guidelines to Models: Advancing Inclusive NLP
A human-in-the-loop approach combining linguistic expertise and NLP to detect and reformulate non-inclusive language.
Guidelines Definition & Corpus Creation
In close collaboration with linguistic experts, we established a set of linguistic criteria for Italian inclusive language. These guidelines served as the foundation for building a real-world corpus of Italian administrative texts, annotated specifically for two key tasks: (1) Detection of non-inclusive language and (2) Reformulation into inclusive alternatives
Human-Centered Validation Design
We implemented two human evaluation workflows to assess model quality:
Detection Evaluation: Linguists analyze whether the models correctly identify the words influencing inclusiveness using explainability methods.
Reformulation Evaluation: Experts verify if the proposed rewordings address inclusivity issues, retain grammatical accuracy, and preserve the meaning of the original sentence.
Model Development & Performance Evaluation
We fine-tuned: (1) Two transformer-based classifiers to detect non-inclusive sentences; (2) Two sequence-to-sequence models to generate inclusive reformulations
We then conducted a comprehensive evaluation combining standard metrics and expert-driven assessments