UQO Working Papers · WP02 · v1.2
Decoding Real Estate Descriptions: Semantic Embeddings and Hedonic Pricing of Residential Properties in Quebec
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Listing text carries priceable information, but much of it is language
Twenty named semantic scores lift adjusted R² from 0.452 to 0.511; 40% of listings are in English and score lower against French references, and net of language and 415 FSA fixed effects the block adds 2.0 points (3.9 with a multilingual encoder).
Abstract
Broker-written listing descriptions contain information about housing quality that structured hedonic variables omit. This paper proposes an interpretable way to use that information: sentence-transformer embeddings of 17,087 single-family listings from Quebec are projected, by cosine similarity, onto 20 researcher-defined reference descriptions (luxury, needs renovation, waterfront, motivated seller, and so on), and the 20 similarity scores enter a semi-logarithmic hedonic model as named regressors. Adding the semantic block to a structural model raises the adjusted R² from 0.452 to 0.511, a gain confirmed out of sample in random and spatially grouped cross-validation.
Three qualifications follow. First, 40% of the descriptions are written in English and score systematically lower against the French references; the language of the listing alone explains 86% of the variance of the mean similarity and accounts for more than half of the apparent semantic gain. Net of language and of 415 forward-sortation-area fixed effects, the semantic block adds 2.0 percentage points of R² with the baseline encoder and 3.9 points with a multilingual encoder, and remains jointly significant with standard errors clustered by area. Second, the individual implicit-price gradients are sensitive to the phrasing of the references and to the encoder: only a core of dimensions (Luxury, Land & Nature, Family-Friendly, Quiet & Peaceful, Needs Renovation) keeps its sign and significance across specifications, while others, including Motivated Seller, New Construction and Entry-Level, do not.
Third, the coefficients are conditional associations in listing prices, not causal effects of wording. The paper documents what reference-based projection can and cannot deliver for hedonic inference and provides a replication pipeline that reproduces every reported number from the raw database.
Key findings
17,087
Quebec single-family listings with price and description
20
named reference descriptions used as semantic regressors
0.452 → 0.511
adjusted R² gain from the semantic block
39.7%
listings written in English, scoring lower on French references
86%
variance of mean similarity explained by language alone
2.0 pp
R² added net of language and 415 FSA fixed effects
Scientific contributions
- Named, auditable text features for hedonic models
- Cosine projection of listing embeddings onto 20 reference descriptions turns an opaque embedding into named regressors; the block is jointly significant in every specification (F(21, 17,059) = 99.5 in the baseline).
- Language, not content, drives half the gain
- 57.6% of listings are French, 39.7% English and 2.7% mixed; mean similarity is 0.434 for French versus 0.204 for English texts, and language indicators cut the semantic ΔR² from 0.047 to 0.017.
- The gain survives location controls and spatial holdouts
- With 415 FSA fixed effects and area-clustered errors the block still adds 2.0 points of R² (3.9 with a multilingual encoder); five-fold FSA-grouped cross-validation gives R²_oos of 0.522 versus 0.484 for the structural model.
- Only five dimensions are robust
- Luxury, Land & Nature, Family-Friendly, Quiet & Peaceful and Needs Renovation keep sign and significance across all specifications and both encoders; the previously emphasised Motivated Seller discount (−8.5%) becomes insignificant once language is controlled.
- Individual gradients are fragile to wording and encoder
- Perturbing the reference texts leaves the similarities almost unchanged (correlation 0.93–0.99) and the block ΔR² at 0.042–0.054, yet the vector of 20 coefficients correlates only 0.02–0.81 with the baseline.
- Inference needs spatial clustering
- FSA-clustered standard errors are on average 50% larger (22–107%) than HC3 errors, so dimension-level conclusions change while block-level conclusions do not.
Methodology
Methodology (5)
Reference-based cosine projection
Listing embeddings are projected onto 20 researcher-defined reference descriptions; each similarity score enters the log-price equation as a named regressor whose coefficient is read per standard deviation of similarity.
Hedonic models A–E and extended specifications
Structural, length, semantic and full OLS models with HC3 errors, then 14 extended specifications with language indicators, 415 FSA or 1,061 municipality fixed effects, FSA-clustered errors, unit-corrected lot area and language subsamples.
Out-of-sample validation
Five-fold random and FSA-grouped cross-validation reporting R²_oos, RMSE and MdAPE for structural, semantic, PCA-embedding and multilingual models.
Sensitivity to references and encoder
Three perturbations of the reference phrasing and a full re-encoding with a multilingual model, with block tests by domain and dimension-level stability checks.
Robustness battery
Quantile regressions, 1,000-replication bootstrap, 1%/99% trimming, VIF diagnostics, Lasso and elastic-net selection, and PCA of the raw embeddings as a comparison.
Research data
- Quebec single-family listings
- 17,087 listings with a valid price and free-text description, extracted from a 46,479-listing SQLite database (snapshot of 7 January 2026); structural fields, lot size, forward sortation area and language detected by stop-word rule.
- Reference descriptions and embeddings
- 20 French reference texts embedded with all-MiniLM-L6-v2 and, in the revision, paraphrase-multilingual-MiniLM-L12-v2; cosine similarities between listing and reference embeddings form the 17,087 × 20 semantic matrix.
Reproducibility
- Eight numbered Python scripts from the raw SQLite database to models, robustness, figures and LaTeX tables; the 2026-10-17 revision adds scripts 07–08 (language, fixed effects, clustering, out-of-sample, reference sensitivity, multilingual encoder)
- Baseline v1.1 results reproduced exactly and kept as benchmark; all revision tables and figure 12 are generated, never typed
- 12 figures and 15 tables regenerated from results/*.csv; embedding caches for both encoders
- Full LaTeX source of the 61-page paper with pre-revision PDF and review report archived in docs/
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.2current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 61 p. |