Entity matching is a critical task in integrating and linking entities across diferent Knowledge Graphs (KGs). Each entity matching task involves a pair of KGs, and the nature of these KGs, such as their size, schema, data quality, and domain, can categorize them into diferent Linking Problem Types (LPTs). Selecting the most appropriate matcher for diferent types of LPTs can substantially enhance the accuracy and efectiveness of entity matching. This research aims to empirically evaluate matchers for each LPT and develop a framework to systematically associate matchers with specific LPTs, enhancing both accuracy and eficiency in the entity matching process.
To empirically evaluate the matchers, we take the pairs of KGs associated with each LPT and assess the performance of each matcher on these pairs, calculating the average precision (prec.), recall (rec.), and F-measure (fm.) across the pairs. This helps identifying the most efective matcher for a given LPT. Table 2 shows the performance of each matcher on each KG pair of Table 1 and sums-up their average performance.
We intentionally removed the KG pair “Starwars-swtor" KG pair from the evaluation process to use it as a test case. Note that the best matching results for the pair “Starwars-swtor" was achieved using BaselineAltLabel (fm. of 0.91).
Matcher
Pair BaseLineLabel1 BaseLineAltLabel1
The framework will utilize Algorithm 1 to systematically select the optimal matcher for each pair of KGs associated with specific LPTs. This process involves comparing the average performance scores of various matchers across the LPTs linked to the input KG pair and selecting the matcher with the highest score.
Example Execution of Algorithm 2 Consider the KG pair “Starwars-swtor" with LPTs 1.1.1.2 and 5.7. For these LPTs, the overall performance of each matcher is computed as: 1. BaseLineLabel Average Precision = 0.935, Average F-measure = 0.73, Average Recall = 0.606. 2. BaseLineAltLabel Average Precision = 0.8875, Average F-measure = 0.7865, Average Recall = 0.715.
3. LogMap Average Precision = 0.8885, Average F-measure = 0.73, Average Recall = 0.6225. 1These matchers utilizes respectively rdfs:label and skos:altLabel for matching entities.
To determine the best matcher, we compare the average F-measure scores. In this case, BaseLineAltLabel has the highest average performance. This outcome aligns with the initial performance results showing the algorithm’s utility in selecting the best matcher.
This research introduces a framework that systematically aligns specific LPTs with the most appropriate entity matching algorithms. Future work will include expanding the matcher evaluation phase to incorporate new pairs of KGs associated with additional LPTs and exploring the integration of other advanced matchers.
This work is partially supported by the French National Research Agency ANR DACE-DL project, grant number ANR-21-CE23-0019.