https://nlp.unibo.it/tag/legal-explanations/Legal ExplanationsHugo Blox Builder (https://hugoblox.com)en-usTue, 13 Oct 2026 00:00:00 +0000https://nlp.unibo.it/media/logo.svghttps://nlp.unibo.it/tag/legal-explanations/https://nlp.unibo.it/proposals/legal/tos-explanations/Tue, 13 Oct 2026 00:00:00 +0000https://nlp.unibo.it/proposals/legal/tos-explanations/<h2 id="context">Context</h2> <p>Privacy policies and terms of service often contain clauses that are unfair to customers or do not fully meet the General Data Protection Regulation (GDPR). These documents are long and technical, so their analysis is complex and slow. Artificial intelligence can help users understand them and make informed decisions.</p> <h2 id="objective">Objective</h2> <p>This proposal uses large language models to generate legal explanations of whether a clause in a digital contract can be considered fair. The model may classify the clause itself or motivate a label assigned by a person or another system. The thesis evaluates the impact of the available information, such as guidelines or other contracts, and of the technique, such as retrieval-augmented generation or few-shot learning.</p> <figure class="lt-viz lt-svg" style="--lt-svg-min: 640px"> <svg viewBox="0 0 800 270" role="img" aria-labelledby="tosexp-title" xmlns="http://www.w3.org/2000/svg"> <defs><marker id="tosexp-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse"><path class="lt-s-head" d="M0 0 L10 5 L0 10 z"/></marker></defs> <title id="tosexp-title">An illustrative clause of a Terms of Service document, We may change these terms at any time, goes to a large language model. The model receives information such as guidelines or other contracts, through retrieval-augmented generation or few-shot learning. It returns a label, fair or unfair, which is either its own or given by a person or another system, and a legal explanation of why the label holds.</title> <text class="lt-s-sub" x="20" y="22">A CLAUSE OF A TERMS OF SERVICE</text> <rect class="lt-s-tok" x="20" y="34" width="330" height="56" rx="10"/> <text class="lt-s-txt" x="185" y="67">We may change these terms at any time.</text> <text class="lt-s-sub" x="20" y="140">INFORMATION</text> <rect class="lt-s-box lt-s-pred" x="20" y="152" width="330" height="64" rx="12"/> <text class="lt-s-txt lt-s-strong" x="185" y="178">Guidelines or other contracts</text> <text class="lt-s-txt lt-s-small" x="185" y="200">through RAG or few-shot learning</text> <path class="lt-s-arrow" marker-end="url(#tosexp-arrow)" d="M354 62 H380 V112 H406"/> <path class="lt-s-arrow" marker-end="url(#tosexp-arrow)" d="M354 184 H380 V148 H406"/> <rect class="lt-s-box lt-s-sel" x="410" y="96" width="160" height="68" rx="12"/> <text class="lt-s-txt lt-s-strong" x="490" y="125">Large language</text> <text class="lt-s-txt lt-s-strong" x="490" y="145">model</text> <text class="lt-s-sub" x="620" y="22">OUTPUT</text> <path class="lt-s-arrow" d="M574 130 H596 V62 V184"/> <path class="lt-s-arrow" marker-end="url(#tosexp-arrow)" d="M596 62 H616"/> <path class="lt-s-arrow" marker-end="url(#tosexp-arrow)" d="M596 184 H616"/> <rect class="lt-s-label" x="620" y="40" width="160" height="44" rx="22"/> <text class="lt-s-txt lt-s-strong lt-s-labeltxt" x="700" y="67">fair or unfair</text> <text class="lt-s-txt lt-s-small" x="700" y="110">its own label, or one given</text> <text class="lt-s-txt lt-s-small" x="700" y="126">by a person or a system</text> <rect class="lt-s-tok" x="620" y="156" width="160" height="56" rx="10"/> <text class="lt-s-txt lt-s-strong" x="700" y="180">Explanation</text> <text class="lt-s-txt lt-s-small" x="700" y="200">why the label holds</text> <text class="lt-s-txt lt-s-small" x="400" y="254">The thesis measures how the type of information and the technique change the explanations.</text> </svg> <figcaption>The explanation task on the illustrative clause of the unfair clause detection proposal, built from the description of the proposal.</figcaption> </figure>