Fusing Semantic, Lexical, and Domain Perspectives for Recipe Similarity Estimation

Denica Kjorvezir; Danilo Najkov; Eva Valencič; Erika Jesenko; Barbara Koroišić Seljak; Tome Eftimov; Riste Stojanov

cs.CLMar 10, 2026

Fusing Semantic, Lexical, and Domain Perspectives for Recipe Similarity Estimation

Authors: Denica Kjorvezir, Danilo Najkov, Eva Valencič, Erika Jesenko, Barbara Koroišić Seljak, Tome Eftimov, Riste Stojanov

Organizations: Jožef Stefan Institute, Ljubljana, Slovenia · S.Cyril and Methodius University, Skopje, N. Macedonia · Biotechnical Faculty, University of Ljubljana, Slovenia

Abstract

This research focuses on developing advanced methods for assessing similarity between recipes by combining different sources of information and analytical approaches. We explore the semantic, lexical, and domain similarity of food recipes, evaluated through the analysis of ingredients, preparation methods, and nutritional attributes. A web-based interface was developed to allow domain experts to validate the combined similarity results. After evaluating 318 recipe pairs, experts agreed on 255 (80%). The evaluation of expert assessments enables the estimation of which similarity aspects--lexical, semantic, or nutritional--are most influential in expert decision-making. The application of these methods has broad implications in the food industry and supports the development of personalized diets, nutrition recommendations, and automated recipe generation systems.

Figures & tables

Figure 1: Graphical representation of the semantic similarity algorithm

Figure 2: Example of a recipe from the Recipe1M dataset

Figure 3: Visual representation of the algorithm used for calculating lexical similarity

Figure 4: Graphical representation of the nutritional similarity between two recipes

Figure 5: Visual representation of the algorithm used for nutritional similarity between recipes based on ingredients

Metric	Score
Semantic (DistilRoBERTa)	0.9015
Semantic (MiniLM-L6)	0.8720
Lexical	1.0000
Nutritional (per recipe)	1.0000
Nutritional (per ingredients)	1.0000

Table 1: Similarity Scores for Case 1

Metric	Score
Semantic (DistilRoBERTa)	0.8689
Semantic (MiniLM-L6)	0.7803
Lexical	0.0091
Nutritional (per recipe)	0.0407
Nutritional (per ingredients)	0.9271

Table 2: Similarity Scores for Case 2

Metric	Score
Semantic (DistilRoBERTa)	0.5086
Semantic (MiniLM-L6)	0.6114
Lexical	0.0000
Nutritional (per recipe)	0.9997
Nutritional (per ingredient)	0.5639

Table 3: Similarity Scores for Case 4

Metric	Score
Semantic (DistilRoBERTa)	0.4875
Semantic (MiniLM-L6)	0.4761
Lexical	0.0000
Nutritional (per recipe)	0.0442
Nutritional (per ingredient)	0.2740

Table 4: Similarity Scores for Case 5

Metric	Score
Semantic (DistilRoBERTa)	0.5200
Semantic (MiniLM-L6)	0.7127
Lexical	0.1125
Nutritional (per recipe)	0.3892
Nutritional (per ingredient)	0.6725

Table 5: Similarity Scores for Case 6

Metric	Mean	Median	Std. Dev.	Skew	Min	Max
DistilRoBERTa (Semantic)	0.6789	0.6739	0.0734	0.7216	0.4327	1.0000
MiniLM-L6 (Semantic)	0.6971	0.6930	0.0683	0.6854	0.4727	1.0000
Nutritional (Recipe)	0.6339	0.6893	0.2929	-0.4831	0.0000	1.0000
Ingredients Jaccard (Lexical)	0.1072	0.0769	0.1391	2.4875	0.0000	1.0000
Ingredients Nutritional	0.7733	0.8000	0.1661	-0.8777	0.0000	1.0000

Table 6: Descriptive Statistics of Core Similarity Metrics

Figure 6: Distribution of Similarity Scores (KDE Plot)

	DistilRoBERTa	MiniLM-L6	Nutritional (Recipe)	Jaccard	Nutritional (Ingredient)
DistilRoBERTa	1.00	0.85	0.25	0.54	0.21
MiniLM-L6	0.85	1.00	0.26	0.52	0.22
Nutritional	0.25	0.26	1.00	0.23	0.26
Jaccard	0.54	0.52	0.23	1.00	0.39
Ingredients Nutritional	0.21	0.22	0.26	0.39	1.00

Table 7: Correlation Heatmap of Similarity Metrics

Figure 7: Correlation Heatmap of Similarity Metrics

Metric	Criteria	Count (N)	Percentage
Nutritional	$\text{nutr}>0.95$ , $\text{roberta}<0.6$ , $\text{jaccard}<0.1$	1,738	1.03%
Semantic	$\text{roberta}>0.85$ , $\text{jaccard}<0.1$ , $\text{nutr}<0.2$	12	0.01%
Lexical	$\text{jaccard}>0.3$ , $\text{roberta}<0.6$	89	0.05%

Table 8: Failure Cases Across Metrics

Lexical Bin	Avg. Semantic	Avg. Nutritional	Final Ensemble Score
0.0–0.1	0.66	0.60	0.45
0.1–0.2	0.68	0.64	0.52
0.2–0.3	0.71	0.68	0.57
0.3–0.4	0.74	0.75	0.64
0.4–0.5	0.78	0.86	0.70
0.5–0.6	0.81	0.91	0.75

Table 9: Average Metric Scores by Lexical Jaccard Bins

Statistic	Value
Mean Absolute Difference	0.0344
Max Absolute Difference	0.1927
Correlation	0.85

Table 10: DistilRoBERTa vs. MiniLM-L6 Semantic Scores

Figure 8: Semantic Model Comparison: DistilRoBERTa vs. MiniLM-L6 (Scatter Plot)

Figure 9: Example web-based tool for evaluating the results

Class	Precision	Recall	F1-score
0 (Not Similar)	0.90	0.95	0.92
1 (Similar)	0.86	0.75	0.80

Table 11: Logistic Regression Model Classification Report

Feature	Coefficient	Normalized Importance
Semantic (Preparation steps)	0.087156	4.1%
Domain (Nutritional values)	0.063678	3.0%
Lexical (Ingredients / Jaccard)	1.986991	92.9%

Table 12: Logistic Regression Coefficients and Normalized Importance

Class	Precision	Recall	F1-score
0 (Not Similar)	0.89	0.91	0.90
1 (Similar)	0.76	0.72	0.74

Table 13: Random Forest Classification Report

Feature	Normalized Importance
Lexical (Ingredients / Jaccard)	42.6%
Domain (Nutritional Values)	37.3%
Semantic (Preparation Steps)	20.0%

Table 14: Feature Importance in Random Forest Model (Normalized Embeddings)

Figure 10: The constructed Random Forest model

Semantic Textual Similarity

Semantic Textual Similarity measures how closely two texts match in meaning, supporting comparison, retrieval, clustering, and evaluation.

See all 44 papers

A framework for recipe data structure with applications for culinary and nutritional insights
Aug 12, 2026Mansi Goel, Sumit Bhagat, Saloni Srivastava +3Knowledge Representation Document Information Extraction
Department of Computational Biology, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Infosys Centre for Artificial Intelligence, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Center of Excellence in Healthcare, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India +2
CGU-ILALab at FoodBench-QA 2026: Comparing Traditional and LLM-based Approaches for Recipe Nutrient Estimation
Apr 28, 2026Wei-Chun Chen, Yu-Xuan Chen, I-Fang Chung +1LLM Evaluation Cost-Aware Inference
Institute of Biomedical Informatics, National Yang Ming Chiao Tung University, Taipei, Taiwan · Institute of Health Data Science, Chang Gung University, Taoyuan, Taiwan · Department of Artificial Intelligence, Chang Gung University, Taoyuan, Taiwan +1
Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings
May 21, 2026Jakub Radzikowski, Josef ChenGraph Embedding

Aug 12, 2026J/K move · Enter open · S save

A framework for recipe data structure with applications for culinary and nutritional insights

Mansi Goel, Sumit Bhagat, Saloni Srivastava, Malav Patel+2

Department of Computational Biology, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Infosys Centre for Artificial Intelligence, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Center of Excellence in Healthcare, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India+2

Knowledge Representation Document Information Extraction

Open paper AskPDF arXiv

Cooking is a complex process that transforms raw ingredients into delicious and nutritious dishes, yet the recipes that encode this process remain largely free text; readable by people but not directly computable. Existing recipe collections capture fragments of this information, but no shared representation links a recipe's structured ingredient composition, its geo-cultural provenance, and its nutritional profile within a single queryable schema. We address this representation gap by formalizing a framework for recipe data structure that decomposes each recipe into typed ingredient entities, grounds those entities in a reference nutritional database, and annotates them with geo-cultural and dietary context. We present RecipeDB2, a structured compilation of 128,942 recipes with 35,474 ingredients from 32 regions and 99 countries. Ingredient phrases are parsed into seven culinary attributes using a transformer-based named-entity model; ingredients are linked to the USDA reference tables through a BERT embedding strategy (F1 = 87.90 on a manually adjudicated set of the 200 most frequent ingredients), yielding 148 nutritional parameters per mapped ingredient; a Random Forest classifier propagates 34 ingredient categories across the full vocabulary; and a deterministic, conservative rule set assigns each recipe a dietary style. Through RecipeDB2 (https://cosylab.iiitd.edu.in/recipedb2/), we demonstrate a scalable framework for making recipes computable, turning culinary heritage (long treated as an artistic rather than a quantitative object) into a data-driven analysis.

Abstract

Semantic Textual Similarity

Explore similar work

A framework for recipe data structure with applications for culinary and nutritional insights

CGU-ILALab at FoodBench-QA 2026: Comparing Traditional and LLM-based Approaches for Recipe Nutrient Estimation

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings