In the automotive industry, accurately estimating the cost of repairing car damages is crucial for both customers and service providers. The process of manually analyzing unstructured reports describing car damage and predicting repair prices is time-consuming and prone to errors. To address this challenge, prediction of repair prices from unstructured damage descriptions.
CEUR ceur-ws.org
The automotive industry faces numerous challenges [
In the following sections, we will delve deeper into the functionality and capabilities of . We will explore how the tool leverages machine learning and natural language processing techniques to analyze and extract crucial information from the unstructured insurance reports. Furthermore, we will present the results of estimation repair price.
In this section, we will provide a detailed overview of the two main phases of our tool.
The first phase of involves the crucial task of structuring the information extracted from the unstructured reports. In this phase, the tool utilizes natural language processing techniques to analyze and understand the textual descriptions of car damages. It extracts entities such as the type of damage (e.g., dented bumper, cracked windshield), its severity, and relevant attributes (e.g., location, size). Furthermore, the tool establishes relationships between these entities to provide a comprehensive understanding of the reported damages.
Named Entity Recognition (NER) is a task in natural language processing that involves
identifying and recognize named entities in text. There are several approaches to NER, including:
rule-based approach, dictionaries approach, machine learning and, deep learning [
To evaluate the performance of the models, we measure various metrics such as F1 score, precision, and recall. Table 1 shows how diferent models performed when they were tested on various entity types such as , , , , , and .
Relation Extraction (RE) is a task in natural language processing that focuses on identifying and extracting relationships between entities mentioned in text. In the context of the information structuring phase of , once named entities related to car damages have been recognized,
Entities the tool aims to establish relationships between these entities to gain a comprehensive understanding of the reported damages. For instance, it aims to extract the relationship ℎ between the car part and the damage. The RE task can be viewed as a classification task. The goal is to classify and categorize the relationships between entities into diferent predefined classes or categories.
In this task, we evaluate the performance of multiple machine learning models for recognizing and classifying relationships between entities. The models under consideration include Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Decision Trees (DT), and Random Forests (RF). The objective is to determine the most efective model for relation extraction. Our evaluation reveals that the RF model outperforms the others, providing better results in this context. This achievement was realized through the optimization of particular hyperparameters, specifically by setting the _ to 8 and the _ℎ to 10. Table 2 displays a comparative analysis of various models for relation extraction covering relationships such as ℎ , ℎ , , and ℎ .
In structuring the information phase, transforms the unstructured textual data into a structured format (Figure 1) that can be readily utilized for further analysis. This phase enables the accurate predictions and ensuring the reliability of the subsequent price estimation.
In the price prediction phase, machine learning algorithms and statistical techniques are employed to estimate the repair cost. These models utilize the structured data obtained from the
Information structuring to predict the price of repair cost. The price prediction models are trained using historical data, which encompasses structured data with actual price values. By analyzing the relationships between the input features and the prices in the training data, the models learn to make predictions on new, unseen data. In our study, we conducted a comparison of several regression models to predict repair costs. The performance of these models was evaluated using diferent evaluation metrics, including mean absolute error, mean squared error, root mean squared error, and R-squared score. Among the evaluated models, model demonstrated the best performance, achieving an mean absolute error of 0.92 with the following hyperparameters: n_estimators=100, learning_rate=0.1, max_depth=5. This indicates that, on average, the predicted repair costs deviate from the actual costs by only 0.92 euros, Figure 2 present an example of estimating repair costs with model.
In Figure 1 and Figure 2, we present the interface of our system, illustrating the lfow from structured information extraction to price prediction. These figures provide a visual representation of the diferent stages involved in our approach. We conducted all our experiments using a real dataset of damage reports, and we ran them on a machine equipped with 16 GB RAM and an Intel Core i7 − 12700 processor, ensuring ample computational resources for both training and testing our models. The tool was developed using Python programming language and the 2 framework. The demo of CarDRP is available online3.
In this study, we propose a tool for predicting repair costs using unstructured data, which operates in two phases. The first phase, information structuring transform unstructured textual data into a structured format to facilitates accurate predictions and enhances the reliability of subsequent price estimations. Furthermore, the price prediction phase employs machine learning specially regression models to forecast the repair costs. Looking forward, there are several perspectives to consider for future research. Firstly, it would be beneficial to explore the incorporation of additional data to train deep learning models and further improve prediction accuracy. Additionally, investigating the integration of semantic ontologies can provide a deeper understanding of the underlying concepts and relationships within the data. This integration can enhance the overall predictive capabilities of the system and enable more accurate estimations. 2https://streamlit.io/ 3https://drive.google.com/drive/folders/1o0NOBqxj3rxFURFBmuH-C6kjj_d85TOU