This paper presents the results of RiMOM in the Ontology Alignment Evaluation Initiative (OAEI) 2016. RiMOM participated in all three tracks of Instance Matching this year. In this paper, we first describe the overall framework of our system (RiMOM). Then we detail the techniques used in the framework for instance matching. Last, we give a thorough analysis on our results and discuss some future work on RiMOM.
of Wikipedia. To make full use of these data, we formalize the instance matching as a binary classification problem, and use the reference mappings to train a classifier, which will determine whether an instance pair is equivalent or not.
Faced with challenges in large-scale instance matching, we propose an novel data
integration framework RiMOM-2016 (the latest version of RiMOM), which is based
on our former ontology and instance matching system RiMOM [
This section describes the overall framework of RiMOM2016. The overview of the
instance matching system is shown in Fig. 1. The system includes seven modules,
i.e., Preprocess, Predicate Alignment, Mathcher Choosing, Candidate Pair Generation,
Matching Score Calculation, Instance Alignment and Validation. The sequences of the
process are shown in the Fig. 1. We illustrate the process as follows.
1. Preprocess: The system begins with Preprocess, which loads the ontologies and
parameters into system. In the meantime, preprocessor can get some meta data
about the two ontologies, which will be used in the later processes, Predicate
alignment and Matcher choosing
2. Predicate Alignment: In this process, we will get the alignments of the predicates
between the two ontologies.
3. Matcher choosing: The system will choose the most suitable one or more
matchers according to the meta data of the ontologies.
4. Candidate Pairs Generation: In this step, we get candidate pairs when the
instances have the same literal objects on some discriminatory predicates.
5. Matching Score Calculation & Instance Alignment: This procedure is the most
striking difference with the last version of RiMOM. In RiMOM-2016, we get
alignments in a supervised or an unsupervised way which depends on whether there
exist reference alignments or not. In case of unsupervised method, we calculate
similarities between two instances on each property, and then we aggregate these
similarities according to the degree of identifying obtained in step 1. On the
contrary, we conduct a supervised method when there exist reference alignments. For
each instance pairs, we also calculate the similarities as unsupervised way. Then
we construct a similarity vector for each pairs and train a logistic regression model
[
Measure if there is validation data set. 1.2
This year we participate in all of three subtasks in the Instance Matching track. We will describe specific techniques in this section.
Data Preprocessing: First, we remove some stop words like ”a, of, the”, etc. Afterwards, we calculate the TF-IDF values of words in each knowledge base. We also calculate some information of each predicate, in order to obtain the degree of identifying of predicates which will be used in similarity aggregation.
Predicate Alignment: The predicates can express rich semantics, and there exist
one-to-one, one-to-many, or many-to-many relationships among these predicates. It is
apparent that we should get the alignments of the predicates before we calculate the
similarity of instances. In RiMOM-2016, we use an object-based method to align
predicates, which is similar with RiMOM-2015 [
Blocking: This step aims to pick a relatively small set of candidate pairs from all pairs. Due to the large scale of knowledge bases, it is impossible to calculate matching scores of all instance pairs. In our method, we firstly generate the inverted index on the objects. instance pairs are selected into the candidate set when they have common objects. This method may reduce the recall slightly, but it also reduce the scale of computation significantly.
Multi-Strategy: We implement several matchers in our system, e.g. label-based approach and structure-based approach. In the preprocess step, we will compare the schema of the two ontologies. If the range of predicates is similar, the label-based approach will play a key role in the matching process. Otherwise, the literal properties are not similar (e.g. the two ontologies are defined in different languages or the intersection of values is really small), label-based approach will not be effective. In this case, we will get some supplementary information (e.g. machine translation, WordNet), or use structure-based appraoch (or use the structure similarity as a feature). In addition, we will use a learning-based method if we have data for training.
track, some of subtasks are defined in the same language, while others use multilingual data sets (e.g. SABINE Task).
Unsupervised method: we use a object-based method to get alignments, it is defined as follows: 0 fpn (i1; i2) = Sim(Oip1n ; Oip2n ) where i1 and i2 are instances from two data sets respectively. Oip1n represent the object value of instance i1 on property pn. Sim(Oip1n ; Op0n ) represent the similarity of i2 object values between these two instances on property pn and its corresponding property p0n. The computing method of this similarity depends on the data type. For example, we use Levenshtein distance for type:text and indicator function for type:int.
Sim(i1; i2) = !1 fp1 (i1; i2) + !2 fp2 (i1; i2) + ::: + !n fpn (i1; i2) (2)
For each property pj , we calculate the similarity according to equation 1 and aggregate them by weights !j which indicate the importance of properties.
Supervised method: In equation 2, the weight wi is determined by meta-data of ontology or manual. Intuitively, it could be improved by a learning-based method if we have some existing alignments. So, basically, we formulate this instance matching problem as a binary classification problem. For a pair of instance i1 and i2, the feature n vector f = ffpi gi=1. Thus, we can use a sigmoid function to compute the probability that instances i1 is equivalent with i2.
P (i1 i2) =
1 1 + ew f(i1;i2)
If i1 i2, P (i1 i2) > 0:5; otherwise P (i1 i2) < 0:5 In this case, the weights w can be determined by the maximum likelihood estimation technique for logistic regression. The assumption in this model is that we can use the machine learning method to determine which property is more important for instance matching problem. 1.3
The RiMOM system and configuration files (2016 version) can be found at https:// drive.google.com/file/d/0BzqVVt4Q8YUuaHpseWJOZkI4MnM/view? usp=sharing. (1) (3)
The Instance Matching track contains three tracks and seven subtasks. RiMOM-2016 participate in all of these tracks, and we will present the results and related analysis in this section. 2.1
There are two subtasks in this track: Inter-linguistic mapping and Data linking. Table 1 is the result for Inter-linguistic mapping task and Table 2 is for Data linking task. Inter-linguistic mapping is a cross-lingual task between English and Italian. As shown in the result, RiMOM preform well in this task. Data linking task requires participants to link the entity to DBpedia, and RiMOM get high Recall but low Precision in this task. There are two subtasks in this track: UOBM and SPIMBENCH. Each subtask contains two data set in different size: sandbox is small data set while mainbox is a large one. Table 3 4 5 6 show the final results in this track. We think RiMOM produce satisfactory results in all of the subtasks.
Tool Precision Recall F-measure LogMapIm 0.701 0.207 0.320
AML 0.785 0.577 0.665 RiMOM 0.771 0.877 0.821 Table 3. The result for UOBM sandbox
Tool Precision Recall F-measure LogMapIm 0.625 0.023 0.044
AML 0.509 0.515 0.512
RiMOM 0.443 0.516 0.477 Table 4. The result for UOBM mainbox 2.3
This track contains three subtasks: 9-heterogeneities, 4-heterogeneities, and False Positive Trap. Table. 7 shows the final result in this track.
Sub-task Precision Recall F-measure 9-heterogeneities 0.813 0.813 0.813 4-heterogeneities 0.746 0.746 0.746 False Positive Trap 0.707 0.707 0.707
Table 7. The result for DOREMUS Track 2.4
Our system can only align two ontologies at a time, and we think it will be a significant improvement if we can develop a system which is able to align several ontologies simultaneously. In addition, in cross-lingual environment, our system still rely on the machine translation. In this case, we hope to develop a method which is language-independent. 3
In this paper, we present the system of RiMOM in OAEI 2016 Campaign. We participate all of the three tracks in instance matching track this year. We described specific techniques we used in the task. In our project, we design a new framework to align instances in different languages. The results turn out that our method is effective.
In the future, we will make great efforts to improve our system continuously. 4
The work is supported by 973 Program (No.2014CB340504), NSFC-ANR (No.61261130588),and NSFC key project(No.61533018), Tsinghua University Initiative Scientific Research Program (No.20131089256) and THU-NUS NExT Co-Lab.