Interdisciplinary research and development projects in medical engineering bene t from well selected collaboration partners. The process of nding such partners from often unfamiliar elds is di cult, but can be supported by an expert pro le that is based on patent analysis and classifying the patents to competence elds in medical engineering. Patent analysis and categorization are di cult and require the analysis of the semantic content. Hence, we propose a twofold approach using a large controlled vocabulary, a smaller competence eld ontology, and an alignment between them to assign patents to a certain competence eld. The approach has two parts: a Topic Map approach and a Publication approach. We evaluate these approaches and its components in several ways. Furthermore, we compare four di erent ways to assign a patent to a competence eld and show that the semantic wealth of a large biomedical ontology is bene cial to the classi cation task.
Ontology matching has been an active research area for more than 10 years
[
While a classication scheme or taxonomy can be easily represented as an
ontology, representing the content of a patent as an ontology or describing the
patent with elements of an ontology is more challenging. Patents have their own
specic language and use a terminology that is dierent from a typical research
publication. Patents are classied using the International Patent Classication
(IPC) system; however, this is too general for a detailed patent analysis [
We are aiming at building a recommender system for research projects in
medical engineering (ME) [
To address the problem of patent terminology, we exploit explicit references to scientic publications and their semantic annotations. In ME, most of the publications appear in journals or conferences that are indexed by PubMed 4. PubMed uses MeSH5, a rich controlled vocabulary with a hierarchical structure, to annotate the publications. Thus, to retrieve a MeSH annotation for a patent, we lookup the references to research articles in PubMed and retrieve the corresponding MeSH terms.
Using references to scientic publications is only one aspect in our approach for patent classication. The overall approach, depicted in Figure 1 consists of two complementary sub-approaches: the Topic Map Approach (TMA) and the Publications Approach (PBA). Both approaches utilize two ontologies - a competence eld (CF) ontology and an ontology with comprehensive medical knowledge (MeSH) - and an alignment between them.
For the Publication Approach, excerpts of publication databases, as well
as their associated MeSH terms are imported into our Data Lake (DL) system
Constance [
Thus, for both approaches, we have a list of related concepts from the MeSH ontology. To establish a link to the competence eld ontology, which we have created to describe the innovation areas in medical engineering (see section 2), we use ontology matching.
There are several questions arising when we analyze the presented approach. Creating an alignment between ontologies and the use of a huge medical ontology in this context require a high amount of resources in terms of memory and CPU power. Hence, we need to know if the eort using it is worth it. Furthermore,
Topic Map Patent
Data Lake System Patent Categorization
Labeled Patent
MeSH Ontology
Competence
Field Ontology
Alignment Ontology Matching Topic Map Approach Publication Approach it is of interest if the quality and size of the alignment between the ontologies have an impact on the results. A special problem is to rate the quality of the alignment without a reference alignment. To answer these questions we present the following contributions in this paper:
We analyze and select medical ontologies to use them as a basis for the creation of the CF ontology and as a single point of entry to identify the semantics of patents and publications.
We describe the process of designing the competence eld ontology and rate its quality based on approved methodologies.
We create dierent alignments between the CF ontology and the medical ontology with dierent matcher congurations and compare their quality. We compare the results of four dierent approaches to categorize a patent: (1) Topic Map Approach with direct comparison of terms with concepts of the CF ontology (i.e., using no ontology matching techniques), (2) Publication Approach, (3) Topic Map Approach, (4) combination of Topic Map Approach and Publication Approach. Approach (2) and (3) use the alignment computed by ontology matching.
The rest of this paper is structured as follows. In Section 2 we explain the
design of the CF ontology. Furthermore, the selection process of the utilized
medical ontologies is explained (rst results about these issues were reported in
[
Our assumption is that a huge medical ontology (or a set of them) and mappings to a smaller competence eld ontology (CFO) will help to more easily classify patents into competence elds. The idea is somehow similar to a smart multilevel lter. First we retrieve terms describing the content of a patent (either from the topic map or the cited publications). These terms are compared to concept names in a huge medical taxonomy using string similarity measures. The most similar ones are selected, which results in a potentially long list of concepts. Afterwards we lter further and search for mappings from the concepts and their predecessors to concepts of the smaller competence eld ontology using more intelligent matchers. This leads to scores which identify the membership condence to the competence elds.
To implement this approach two foremost things have to be done: (1) we
have to model the competence eld ontology and (2) we need to evaluate and
select comprehensive medical ontologies. For the design of ontologies there exist
several acknowledged methodologies, such as METHONTOLOGY [
To create the CFO, we started from the descriptions in [
In parallel we searched for one or multiple large biomedical taxonomies. We
need these taxonomies for two things. First, we want to extend the basic CFO
we created before with more terms to describe the competence elds in more
detail. Second, we need the large ontology as entry point to nd terms describing
the patents and with the alignment to the CFO we can determine the
corresponding competence elds. This corresponds to the sixth scenario of the NeOn
methodology, namely reusing, merging and reengineering ontological resources .
The rst step in this scenario is the ontological resource reuse process , starting
with the Ontology Search [
Imaging Techniques
NCIT
NCIT + MeSH Prostheses & Telemedicine Operative & In-Vitro Special Implants Interventional Diagnostics Therapies &
Dev. and Sys. Diagnosis Sys.
NCIT + MeSH + SNOMEDCT NCIT + MSH + SNOMEDCT + RHMeSH
Complete
the Ontology Assessment and Comparison steps [
Hence, we decided to analyze the coverage by adding one ontology after another to see the gain of adding further ontologies. We used the most promising ontologies identied before and started with the NCI Thesaurus. Figure 2 shows the results.
It can be noted, that we gain about 10% coverage using all ontologies. The biggest gain is achieved by adding the MeSH ontology. Thus, we decided to use the NCIT and the MeSH ontologies to extend the CFO, as this was a good compromise between coverage and complexity. For the matching of the biomedical ontology to the CFO we rst picked only one ontology to keep the computational overhead during runtime low. If it does not give us satisfying results, we will add more ontologies and also align them with the CFO. One possibility would be also to use the UMLS which is a superset of many medical ontologies, but it is really large, which could lead to performance problems. For now, we selected the Robert Hoehndorf MeSH 9 as it has a good coverage and is available in the OWL format.
The next steps to develop the CFO are the ontology aligning and ontology merging step and the ontological resource engineering process . We proceeded in these steps as follows. We took the extracted terms, the so far found concepts from the coverage analysis, and the detailed description of the innovation elds, and carried out an extended search in the MeSH Browser 10 and the NCIT Brow
ser11 for these and related concepts. We analyzed the hierarchical structure of
each of the found concepts and decided for each concept if it is adopted into
the CFO. Where applicable we also adopted the inheritance relationship of
concepts. We extended and restructured the CFO in cycles, i.e., according to [
The ontology has been implemented in OWL using the NeOn toolkit 12. We evaluated the CFO also in tests in the complete process of patent categorization. We noticed that the initial results were not satisfying because some competence elds were not represented well in the CFO. Hence, we did a frequency analysis of the MeSH terms from the Publications Approach. We made a ranked list of MeSH concepts based on how often they have been searched for, but did not lead to matches in the CFO. Based on this list we added more useful concepts to the CFO (no trivial, misleading terms, such as Human, but for example Gene Expression Regulation ). The current CFO consists of 529 concepts and can be downloaded at http://dbis.rwth-aachen.de/cms/projects/mi-mappa/CFO.owl . 3
As explained above, we are using three dierent basic approaches and one combined approach to classify patents. Figure 4 gives an overview of the dierent approaches. #1: TMD (Topic Map with Direct Mapping): In this approach, we match the terms extracted from the topic maps directly with the competence 11 https://ncit.nci.nih.gov/ncitbrowser/ 12 http://neon-toolkit.org eld ontology. This can be seen as a base line as it does not use a semantically rich ontology as intermediate component, but only uses string matching to match terms and ontology elements. #2: PBA (Publication-Based Approach): This approach uses the MeSH terms attached to publications which are referenced by a patent. Then, we use an alignment between the CFO and MeSH to compute a score for the relationship between a patent and a competence eld. #3: TMA (Topic Map Approach): Here, we also use topic mapping (as in approach #1) to create initial clusters of patents and extract terms occurring frequently in these clusters. These terms are then matched with the concepts of the MeSH ontology. Using the same alignment as in the second approach, a relationship to the CFO is established. #4: COM (Combined Approach of #2 & #3): This is a combination of PBA and TMA, with an emphasis on the results of PBA.
As the approaches TMD and TMA are based on topics, we rst briey explain this part, before we present how we did the alignment between of CFO and MeSH, and describe the publication-based approach. 3.1
Topic Mapping
A basic set of patents is used to build a topic map. Firstly, the corpus of
documents is preprocessed (stemming, removing stop words, etc.) and a
DocumentTerm-Matrix (DTM) is created. The matrix is input to a Latent Dirichlet
Allocation (LDA) algorithm with the Gibbs sampling algorithm for estimation and
variational expectation maximization [
We evaluated dierent numbers of topics and dierent numbers of terms
extracted for each topic (e.g., 10, 30, 50, etc.). As computation of the subsequent
steps increases with a higher number of topics and terms, we used 50 topics and
50 terms for our evaluation in Section 4. As the TMD approach matches the
terms directly with the CFO, no further processing on the extracted terms is
done in this case. We just do a similarity calculation using a normalized Longest
Common Subsequence [
For the categorization of the input patent with the TMA, the topic with the highest probability in the topic map (or multiple topics if they have the same probability) is retrieved. Each term characterizing the topic is compared with all concepts in the medical ontology resulting in a set of matching concepts. For each of the concepts in the set direct mappings and mappings of parent concepts are collected from the alignment and it is determined to which competence eld the matching concept in the CF ontology belongs. From the similarities average scores are calculated for each term and each competence eld. Based on this, an average score is calculated from all terms for the topic(s) of the patent. Hence, for each patent we have a score for each of the competence elds and normalize these, such that all scores add up to 1. 3.2
Ontology Matching
To rate how strong a patent or publication is related to a certain competence
eld, we need to match the describing terms either extracted from publications
or from the topic map to terms describing the competence elds. In preparation
to this step, we create an alignment between the selected MeSH ontology and the
CFO. The alignment constitutes of a set of mappings between the concepts of the
two ontologies. This means, for each mapping we have a pair of concepts and a
similarity value. As we do not try to re-invent the wheel, we used
AgreementMakerLight [
Currently, we are also testing other settings and their impact on the quality of patent classication results. First experiments show, that slightly relaxed lter settings (e.g., not using a cardinality lter) increases the number of mappings and therefore, also improves the classication result. 3.3
Publication-based Approach We queried the web service of EPMC 13 to retrieve the metadata of the papers referenced in our patent dataset. To extract the references from the patent 13 European PubMed Central, https://europepmc.org/ data, we use a pattern-based approach similar to the FreeCite citation parser 14. Luckily, the patent data is semi-structured such that the citations can be clearly identied. Nevertheless, for a large fraction of the patents, we are not able to retrieve MeSH terms from referenced publications (either because the referenced publication does not appear in PubMed or the citation is incorrect).
The retrieved metadata for each referenced publication is then stored in
our Data Lake system Constance [
Subsequently, we use a process which is similar to the TMA. In both cases, we have a list of MeSH terms as input. For each of the terms in the list, the mappings are determined as before and average scores per competence eld are calculated and normalized for each patent. 3.4
Combined Approach In the combined approach (COM), if both approaches TMA and PBA deliver results, the results are combined and overall scores for each competence eld are determined. In all cases, we assign at most three competence elds to a patent. In most cases, only one competence eld is assigned to a patent as the other competence elds do not exceed a certain threshold. Thus, we take the intersection of competence elds computed by TMA and PBA. If this is not empty, we take this result (because both approaches are sure about a result). If the intersection is empty, we take the competence elds with the highest scores from TMA and PBA. 4
In our experimental setup, we compare the aforementioned approaches. For the
analysis of patents, we need a comprehensive data basis with high data quality.
In the course of the mi-Mappa project, a subset of the PATSTAT database (2016
Spring edition, version 5.07) published by the European Patent Oce (EPO) is
used. For our purposes, we selected patents issued by a German (DE) or British
(UK) authority after 2004, which are from the medical domain (CPC class A61),
and which have an English abstract and title. This results in a set of 26,814
patents. For about 4,500 patents of this set, we are able to retrieve MeSH terms
for the referenced publications. From this set, we randomly selected 59 patents
to do a manual assignment to competence elds to evaluate our approaches. A
more extensive expert evaluation is currently being setup. In addition, we plan
also to evaluate our approach to the results of our project partners who apply a
supervised learning approach using Support Vector Machines [
For TMA, we experimented with various congurations for the number of topics and their associated terms. We observe that with a relatively small number of topics and terms, e.g. 10 or 20, the terms are extremely broad-based and do 14 http://freecite.library.brown.edu/ not provide meaningful matches with the MeSH ontology or the CFO. Therefore, based on the results, we chose the number of topics, as well as the number of terms to be 50 for our default test conguration.
Fig. 5 summarizes the ndings from our experiments for the aforementioned approaches. It is obvious that all three of our proposed approaches #2, #3, and #4 perform signicantly better than the baseline approach #1. All three evaluation parameters, i.e., precision, recall, and the f-measure are worse for the baseline approach. In contrast, when the MeSH ontology is used for matching the ontology terms (#3 ), the precision and f-score are 0:375 and 0:38, respectively, which are more than the doubled values of corresponding values produced by #1. The PBA performs even better, resulting in precision, recall, and f-score values of 0:46, 0:47 and 0:44, respectively. However, the combined approach #4 signicantly outperforms all the others, and results in precision, recall, and fmeasure values of 0:53, 0:55 and 0:53, respectively. Indeed, we found that in the case of the TMD-approach #1, there were a lot of erroneous matches, which led to non-distinctive results for the CFO assignment. These results arm the superiority of techniques which use a comprehensive biomedical ontology and ontology matching for patent classication tasks. 5
There are only few works that apply ontology matching in the context of
patent analysis. Semantic similarities (based on ontology matching) and case-based
reasoning have been applied in the design of invention processes which use
patent analysis to study related works. Patent analysis using ontologies has been
applied especially for patent search [
Patent analysis is a complex topic as patents use their own language and terminology. Even for humans used to research publications, patents are dicult to understand. Thus, typical approaches for classifying patents might fail.
In this paper, we investigated an ontology-based approach to assign patents to competence elds in medical engineering. We developed two dierent approaches and a combined approach that are based on a large biomedical ontology, its alignment to the competence eld ontology designed by us, and other ontology matching techniques. We have shown that these more elaborated approaches outperform an approach that directly matches terms of patents with the competence eld ontology.
However, the overall f-measure of about 55% for the combined approach is not yet satisfying. One problem is the small set of patents for which we have assigned competence elds that we can use as a ground truth. This will be extended with a larger expert evaluation in which patents will be classied by several experts. Even humans might disagree on the assignment of a patent to a competence eld; therefore, we will have multiple expert opinions for one patent. We will also work on ne tuning and optimizing our approach. So far, we focused on the quality of the result, and did not worry too much about the performance. Still, we think that the area of patent classication is an interesting eld which could benet more from the results in ontology matching.
This work has been supported by the Klaus Tschira Stiftung gGmbH in the context of the mi-Mappa project ( http://www.dbis.rwth-aachen.de/mi-Mappa/ , project no. 00.263.2015). We thank our project partners from the Institute of Applied Medical Engineering at the Helmholtz Institute of RWTH Aachen University & Hospital, especially Dr. Robert Farkas, for the fruitful discussions of the approach and for providing the patent data.