# RECORD 4c31e259cf5625db8bd0cbbcae78c68fed2c8164 # TITLE: Text-augmented house valuation: using embedding vectors and SHAP values # YEAR: 2026.0 VENUE: Aestimum DOI: 10.36253/aestim-19824 # AUTHORS: Lee Changro # PAGES: 14 CANDIDATE TABLE PAGES (1-based): [6, 8, 14] # ABSTRACT: With the recent advent of large language models, text data can now be readily processed and analyzed across various sectors; traditionally, this was not the case, and text data had proven challenging to analyze. The housing market uses and features a vast amount of textual data, such as property descriptions and neighborhood trend reports, but has yet to fully exploit this potential. This study has two objectives: first, to enhance house valuation accuracy by integrating text information into valuation models; and second, to interpret how text information contributes to improved performance. We convert the environmental descriptions of each house into embedding vectors, which are then incorporated into regression and extreme gradient boosting (XGB) models to estimate house prices. Additionally, we interpret the XGB model’s results using Shapley Additive explanations (SHAP) values. The embedding vectors significantly improved the performances of both the regression and XGB models, and the embedding vector values aligned well with the semantic meaning of the environmental descriptions of each house. This study contributes to the literature by deepening our understanding of embedding vectors in the context of house valuations and housing markets. ===== PAGE 1 (score 4, numeric-paren 0, table-words 0, var-words 4) ===== 1 AESTIMUM JUST ACCEPTED MANUSCRIPT Text-augmented house valuation: using embedding vectors and SHAP values Lee Changro Department of Real Estate, Kangwon National University, South Korea Email: spatialstat@naver.com This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences between this version and the Version of Record. Please cite this article as: Changro, L. (2026). Text-augmented house valuation: using embedding vectors and SHAP values. Aestimum, Just Accepted. DOI: 10.36253/aestim-19824 JUST ACCEPTED ===== PAGE 6 (score 31, numeric-paren 15, table-words 2, var-words 11) ===== 6 AESTIMUM JUST ACCEPTED MANUSCRIPT Figure 2. Flow of research. To convert the environmental descriptions into embedding vectors, three embedding models are employed: a word- embedding model (glove-wiki-gigaword-50) and two sentence-embedding models (paraphrase-MiniLM-L12-v2 and paraphrase-mpnet-base-v2). The glove-wiki-gigaword-50 is pre-trained and represents words as 50-dimensional vectors using data from Wikipedia and the Gigaword corpus. Paraphrase-MiniLM-L12-v2 and paraphrase-mpnet-base-v2, are referred to as sentence embeddings A and B, respectively; both are pretrained models from the Sentence Transformer library. Sentence embedding A is a lightweight transformer model that produces 384-dimensional embedding vectors, whereas sentence embedding B, which is an enhanced version of BERT and RoBERTa, generates 768-dimensional embedding vectors4. Evaluating and selecting the relevant embedding models is challenging (Torregrossa et al., 2021). In this study, model selection was guided by two complementary criteria: semantic similarity and downstream valuation performance. To illustrate the relative representational quality of the three embedding models, the most common environmental description in the dataset, “Remote and rural area” (455 houses, 27% of the total), was used as a reference, and the five most similar descriptions were identified using cosine distance and presented in Table 3. While relying on a single reference sentence may not fully capture the diversity of environmental descriptions across the dataset, this comparison was intended as an interpretable demonstration rather than the sole basis for selection. Descriptions in word embedding showed less similarity to the reference, whereas those in sentence embeddings A and B were more similar. Ultimately, sentence embedding B was chosen for use in the subsequent valuation models as it outperformed sentence embedding A in a downstream model performance (i.e., XGB). Table 3. Top five environmental descriptions similar to “Remote and rural area”. No. Word embedding Sentence embedding A Sentence embedding B 1 Pure farming area (0.321) Rural area near a village (0.074) Rural area (0.062) 2 Existing industrial area (0.362) Rural area (0.083) Rural area on a mountainside (0.149) 3 High-density commercial district (0.415) Rural area on a mountainside (0.167) Pure farmland (0.212) 4 Existing residential area (0.428) Rural area nearby the coast (0.186) Rural area near the suburbs (0.268) 5 Emerging residential area (0.429) Rural area near the suburbs (0.213) Rural area nearby the coast (0.281) Note: Numerical values indicate the cosine distance (= 1 – cosine similarity), where a value closer to zero indicates greater similarity. 4 Details on these models are available at https://www.sbert.net/docs/sentence_transformer/pretrained_models.html. Prior to embedding, the environmental descriptions underwent text preprocessing steps. These included lowercasing all text, removing punctuation and special characters, and stripping leading and trailing whitespace. Tokenization was handled internally by each embedding model: the word embedding model tokenizes text at the word level, splitting descriptions into individual word tokens, while the two sentence embedding models employ sub-word tokenization via their respective built-in tokenizers from the Sentence Transformers library, which handle out-of-vocabulary terms more robustly. ===== PAGE 7 (score 12, numeric-paren 0, table-words 0, var-words 12) ===== 7 AESTIMUM JUST ACCEPTED MANUSCRIPT Sentence embedding B consists of 768-dimensional vectors which can be inefficient if used in their entirety in subsequent valuation models. To enhance efficiency, sentence embedding B is reduced to two dimensions using principal component analysis (PCA). The explained variance ratio for these two dimensions (two components) is 37%, with the first and second components accounting for 22% and 15%, respectively: hereafter referred to as embedding components 1 and 25. 3.3 Valuation models Three valuation models are used to estimate house prices: OLS without text information, OLS with text information, and XGB with text information. The first model acts as the baseline. XGB is a widely used machine learning model that has demonstrated excellent performance in various studies (Alshboul et al., 2022; Chen et al., 2015; Osman et al., 2021). An additional benefit of using XGB is the ease of computing Shapley Additive explanations (SHAP) values, which can be efficiently calculated for tree-based models, including gradient boosting. The SHAP value begins by examining all possible subsets of explanatory variables, commonly known as coalitions. For each explanatory variable, its marginal contribution to the prediction is calculated when added to each possible coalition of the other explanatory variables. The SHAP value for an explanatory variable is the average of its marginal contributions across all possible coalitions (Lundberg and Lee, 2017). A positive SHAP value indicates that the explanatory variable positively contributes to the prediction or pushes it toward a higher value, whereas a negative SHAP value indicates the opposite effect. The absolute value of SHAP value indicates the strength of the contribution of the explanatory variable. A SHAP value of zero indicates that the explanatory variable has no impact on the prediction6. To estimate house prices, Equation (1) was used: 𝑃𝑟𝑖𝑐𝑒 (𝑙𝑜𝑔−𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑) = ℎ𝑜𝑢𝑠𝑒 𝑡𝑦𝑝𝑒+ 𝑠𝑖𝑡𝑒 𝑎𝑟𝑒𝑎+ 𝑓𝑙𝑜𝑜𝑟 𝑎𝑟𝑒𝑎+ 𝑠𝑡𝑜𝑟𝑦+ 𝑟𝑜𝑎𝑑+ 𝑠𝑖𝑡𝑒 𝑠ℎ𝑎𝑝𝑒+ + 𝑏𝑢𝑖𝑙𝑑𝑖𝑛𝑔 𝑓𝑟𝑎𝑚𝑒+ 𝑝𝑟𝑜𝑝𝑒𝑟𝑡𝑦 𝑎𝑔𝑒+ ℎ𝑜𝑢𝑠𝑒 𝑏𝑒𝑎𝑟𝑖𝑛𝑔+ 𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑜𝑛𝑒𝑛𝑡 1 + + 𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑜𝑛𝑒𝑛𝑡 2 (1) From the explanatory variables available in the benchmark house data, 11 were selected after reviewing their coefficients and significance. House type indicates the subcategory of a house, such as detached, attached, rowhouse, mixed type, and other variations. Site area refers to the area of the house site, whereas floor area is the total floor space within the house. Story denotes the number of floors; road categorizes the size of the road, including main, middle, narrow roads, and other variations; and site shape describes whether the house site is regular or irregular. Building frame includes types like reinforced concrete (RC), wooden frames, steel frames, brick masonry, and others. Property age is the number of years since a house was constructed and house bearing indicates the direction that the house faces (south or non-south). The final two variables were the principal components derived from the embedding vectors. The XGB implementation details are as follows. The maximum depth of the decision tree was set to 3. Increasing this value can result in a more complex model and potentially cause overfitting. A learning rate of 1.0 was used. The number of boosting rounds was set to 20. In addition, the squared error was selected as the loss function7. 4. Results and discussion 5 The number of PCA components was selected based on an evaluation of a downstream model performance using XGB. Models incorporating 3, 5, and 7 components were compared against a two-component baseline using mean squared error (MSE). No meaningful improvement was observed beyond two components. Additionally, beyond the first two components, each additional component individually explained less than 10% of the variance, with diminishing marginal returns. Incorporating more components also increase the risk of overfitting, particularly given the relatively small sample size (763 in Seongdong-gu and 907 in Gangjin-gun). Thus, a two-component representation was retained as the reasonable specification. 6 Details on the SHAP value implementation are provided in Lundberg and Lee (2017). 7 These hyperparameters were determined through evaluation on a validation dataset, where key parameters (tree depth, learning rate, and number of boosting rounds) were varied and the resulting goodness of fit was assessed. The reported values reflect the configuration that yielded the best performance. We acknowledge that a more systematic tuning approach, such as grid search, was not employed, which remains a limitation of the current study. ===== PAGE 8 (score 34, numeric-paren 17, table-words 5, var-words 4) ===== 8 AESTIMUM JUST ACCEPTED MANUSCRIPT 4.1 Text-driven model improvement Table 4 presents the results of the OLS analysis. In both Seongdong-gu and Gangjin-gun, embedding components 1 and 2 significantly enhanced the model performance. The adjusted R² values increased from 0.508 to 0.558 and from 0.633 to 0.751 in both regions, respectively. This indicates that text information, specifically environmental descriptions, contributes to improving valuation accuracy. Table 4. Results of the OLS analysis. Variables Seongdong-gu Gangjin-gun Descriptions not included Descriptions included Descriptions not included Descriptions included Intercept 4.791* 4.581* 2.394* 2.224* House type: rowhouse - - 0.381+ 0.455+ House type: attached -0.082+ -0.010+ - - House type: detached 0.010+ 0.018+ 0.306+ 0.264+ House type: mixed type -0.468* -0.553* -0.414+ -0.623* House type: other -0.706* -0.611* 0.060+ 0.208+ Site area 0.009* 0.009* 0.000* 0.000* (Skipped for brevity) … … … … Embedding component 1 - -0.265* - -0.393* Embedding component 2 - 0.592* - -0.090* Adjusted R2 0.508+ 0.558+ 0.633+ 0.751+ * Significant at the 5.0% level. The complete table is provided in the Appendix (Table A1). Figure 3 shows the goodness-of-fit of the test dataset for the three models: OLS without text information, OLS with text information, and XGB with text information. In both regions, the OLS model incorporating environmental descriptions demonstrated a better fit than the OLS model without such information. Furthermore, the XGB model exhibited the best fit among the three models in both regions, making it a reliable choice for more detailed interpretations8. 8 In Seongdong-gu, the difference between OLS without text and with text may not be obvious, but the improvement from OLS with text to XGB with text is discernible: the latter model’s predictions align more closely with observed values. In Gangjin-gun, the increasing alignment between predictions and observed values (from OLS without text, to OLS with text, and finally to XGB with text) is noticeable. ===== PAGE 9 (score 4, numeric-paren 0, table-words 0, var-words 4) ===== 9 AESTIMUM JUST ACCEPTED MANUSCRIPT Figure 3. Goodness-of-fit. 4.2 SHAP value interpretation The XGB model demonstrated the best fit, as shown in Figure 3. Figure 4 presents the importance of the explanatory variables, as determined by the SHAP values of the XGB model. Seongdong-gu, a highly urbanized area with scarce land, identifies site area as the most critical factor in estimating house prices. In contrast, Gangjin-gun, a rural area with abundant land, finds building-related factors such as property age to be more significant in house price estimations. In both regions, embedding components, especially the first one (embedding component 1), rank third in Seongdong-gu and second in Gangjin-gun9. 9 As shown in Figure 4, embedding component 2 ranked sixth in both Seongdong-gu and Gangjin-gun, indicating a non-negligible role in house valuation. As noted earlier, the two embedding components together capture 37% of the variance in the sentence embeddings, with components 1 and 2 accounting for 22% and 15%, respectively. By construction, PCA produces orthogonal components, meaning embedding component 2 captures a dimension of linguistic variation in environmental descriptions that is not represented by component 1. The semantic interpretation of component 2 is less straightforward than that of component 1, and therefore we focus on the interpretation of component 1 in this study. ===== PAGE 14 (score 74, numeric-paren 56, table-words 2, var-words 6) ===== 14 AESTIMUM JUST ACCEPTED MANUSCRIPT Appendix Table A1. Results of the OLS analysis. Variables Seongdong-gu Gangjin-gun Descriptions not included Descriptions included Descriptions not included Descriptions included Intercept 4.791* 4.581* 2.394* 2.224* House type: rowhouse - - 0.381+ 0.455+ House type: attached -0.082+ -0.010+ - - House type: detached 0.010+ 0.018+ 0.306+ 0.264+ House type: mixed type -0.468* -0.553* -0.414+ -0.623* House type: other -0.706* -0.611* 0.060+ 0.208+ Site area 0.009* 0.009* 0.000* 0.000* Floor area -0.001* -0.002* 0.001* 0.001* Story 0.084* 0.087* 0.319* 0.118+ Road: middle 0.204* 0.204* 0.334+ 0.440+ Road: narrow 0.184* 0.226* -0.295+ -0.022+ Road: narrower 0.288* 0.314* -0.696+ -0.206+ Road: narrowest 0.120+ 0.177* -0.654+ -0.195+ Site shape: regular 0.144* 0.136* 0.204* 0.108* Building frame: brick 0.246* 0.284* 0.792* 0.734* Building frame: RC 0.375* 0.438* 1.475* 1.385* Building frame: low quality wood 0.197+ 0.239+ 0.153* 0.237* Building frame: high quality wood - - 1.730* 1.786* Building frame: panel - - 0.740* 0.862* Building frame: steel - - 1.142* 1.304* Property age -0.005* -0.005* -0.010* -0.011* House bearing: south -0.063* -0.038+ 0.021+ 0.051+ Embedding component 1 - -0.265* - -0.393* Embedding component 2 - 0.592* - -0.090* Adjusted R2 0.508+ 0.558+ 0.633+ 0.751+ *: Significant at the 5.0% level.