This work asks whether an unsupervised representation learned directly from patent text can overcome the cross‑code discovery limitation of traditional classification‑code‑driven retrieval systems, and whether a hyperbolic embedding space offers an advantage. We trained a hyperbolic (Poincaré) VAE and a matched Euclidean VAE on 45,000 US patents, running 96 experiments (6 random seeds). On 3,000 query sets, the hyperbolic model retrieved genuinely cross‑category prior art in the top‑10 results 18.76% of the time, a 52% improvement over the Euclidean baseline’s 12.34%, a practically meaningful gain. The hyperbolic model also diverged more from official CPC classifications (NMI = 0.2392 vs. 0.3206); we explain this geometrically: the two objectives capture different goals, and the hyperbolic one aligns better with prior‑art search needs. In a follow‑up study across 10 technology domains (an additional 50,000 patents), we found that a domain’s classification breadth (number of distinct CPC sub‑groups covered) negatively correlates with performance (Pearson r = ‑0.642, p = 0.045), while patents per subgroup positively correlate (r = 0.715, p = 0.020). These signals let practitioners gauge where the method is most valuable. All findings are reported with full statistical transparency, and we release a complete, 91‑test automated implementation.
Review