An infographic visualizing the key findings from the WMT25 Legal Domain Test Suite, highlighting the performance of Large Language Models (LLMs) and the future of automated legal text translation.
Sentences Analyzed
From the WMT25 English-Hindi legal dataset, ranging from 5 to 55 words in length.
Top BLEU Score
Achieved by Gemini-2.5-Pro, demonstrating superior lexical accuracy.
Top CHRF++ Score
Also by Gemini-2.5-Pro, indicating high character-level fidelity, crucial for legal texts.
This chart compares the top 10 Machine Translation systems based on the COMET score, a metric highly correlated with human judgment. LLM-based systems clearly dominate the top ranks.
The chart visualizes the COMET scores, which predict human judgments of translation quality. Higher scores indicate better performance. Gemini-2.5-Pro leads, followed closely by a mix of specialized NMT and other large language models, showcasing the competitive landscape.
This radar chart provides a multi-faceted view of the top 5 systems, comparing their performance across three critical and complementary evaluation metrics: BLEU, METEOR, and CHRF++.
Each axis represents a different quality metric. A larger area indicates a more balanced and robust performance. Gemini-2.5-Pro shows strong, well-rounded capabilities, while other systems exhibit varying strengths. For instance, some systems excel in lexical overlap (BLEU) but are weaker in character-level accuracy (CHRF++).
Input a legal phrase or sentence in English and let the Gemini API provide a translation and a brief quality analysis.
The methodology for assessing translation quality is shifting from simple word-matching to more nuanced, human-aligned frameworks. This is crucial in the legal field where meaning and precision are paramount.
Traditional Metrics
(e.g., BLEU)
Advanced Metrics
(e.g., METEOR, CHRF++)
Neural & Human-Aligned
(e.g., COMET, MQM)
Early metrics focused on matching words and phrases, which is useful for terminology but misses overall fluency and meaning.
Metrics like METEOR and CHRF++ improved evaluation by considering synonyms, word stems, and character sequences, offering a more accurate quality signal.
Modern frameworks like COMET and MQM aim to replicate human judgment, identifying specific error types and providing actionable feedback for improvement.
While LLMs show immense promise, their deployment in high-stakes legal environments requires addressing key challenges related to reliability and trustworthiness.
LLMs can generate plausible but factually incorrect or legally unsound content. In a legal context, this poses a significant risk, potentially leading to contractual disputes or miscarriages of justice.
Legal reasoning must be transparent and auditable. "Black-box" AI is insufficient. The future lies in Neuro-Symbolic AI, which combines LLM fluency with logic-based systems to provide clear, traceable, and justifiable outputs that align with legal standards.