This position paper outlines our proposed approach to automate the process of creating mappings between di erent data formats in the transportation domain. The approach exploits the word2vec model, in combination with graphs for nding meaningful equivalent relationships between concepts in di erent data formats. The modern vision of transportation is that of "mobility as a service", in which users can seamlessly build door-to-door trips including several travel modes through a single entry point, with a uni ed interface and payment methods. To realize this vision, a wide range of diverse actors of the transportation ecosystem must communicate, interact, and cooperate with one another. Divergence of transportation standards and heterogeneity of data representations, formats and models are the main obstacles towards making such an interoperable system a reality. Hence, solutions are needed that bridge this fragmentation, hide the peculiarities of di erent standards and allow for the communication and exchange of data among heterogeneous, non-integrated systems. In line with this objective, The SPRINT (Semantics for PerfoRmant and scalable INteroperability of multimodal Transport) project aims at developing tools and technologies that facilitate interoperability in the transport domain. The core idea underlying the project is to go beyond pure \syntactic" interoperability|where interested parties are forced to adopt a uni ed set of formats for data exchange|and instead leverage \semantic" interoperability, which enables di erent systems to communicate with each other through their native standards, by mapping their concepts to a common ontology, which provides an unambiguous and homogeneous view of data. One of the speci c goals of the SPRINT project is to enhance and automate the conversion process realized ? This work was supported by Shift2Rail and the EU Horizon 2020 research and innovation programme under grant agreement No: 826172 (SPRINT). y Copyright 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Semantic Model
(Reference Ontology)
t
f o
gin ssec
iL rP
by the ST4RT Converter [
A ST4RT converter (whose main principles and processes are depicted in
Fig. 1) is a software artefact that acts as an adapter between two distinct formats.
Given a suitable mapping between the source/target data and a reference formal
ontology, a ST4RT converter rst transforms data expressed in the source
format into an intermediate representation based on the reference ontology. Then,
following a similar procedure in the reverse direction, the converter translates
the intermediate representation into the target data model. This approach has
the notable advantage of exempting participating parties from harmonizing the
syntax and structure of their data; a meaningful communication is achieved only
if they agree on the concepts and semantics behind their terminology and syntax.
As shown in Fig. 1(a), the ST4RT Converter relies on dual lifting and lowering
processes mapping given standards to/from the reference ontology. The rst step
in whole approach is the annotation process, through which the source and
target standards are semantically annotated to state the mappings between their
data models and the reference ontology. Figure 1(b) shows the whole
conversion work ows at design-time (left) and run-time (right). At design-time, given
the structured data model of a standard, corresponding java classes are
automatically generated, which are the basis for annotation process. At this stage,
java classes, attributes and methods must be annotated to map each term to its
equivalent concept in the reference ontology [
Lifting & Lowering Specification
(Standard A)
XML (Source)
Run-time
Java Object (Source) XML (Target)
Java Object (Target) gni ss re ec woL orP
Design Time XSDs Annotations
Java Classes (a) (b) Fig. 1. a) ST4RT approach to semantic interoperability, b) Converter Work ow at design- and run-time, composed of annotation, lifting and lowering processes. an instance of the target data model. Except for the mapping process in the annotation phase, all steps are accomplished in an automated manner.
This position paper presents our proposed approach to make the annotation phase of the conversion process more e cient, automated, and user-friendly. As mentioned above, so far the annotation step is carried out manually, which hampers the e ciency and overall performance of the process. Human users who are expert in both the reference ontology and the desired standards are required to establish the mappings with the concepts appearing in the source and target speci cations, which is a time- and e ort-consuming procedure. In the proposed approach, we aim at making the annotation-creation process more automated by taking advantage of machine-learning algorithms and methodologies.
The rest of this paper brie y describes related works (Sect. 2), then outlines the proposed method (Sect. 3), and concludes with a brief discussion (Sect. 4). 2
The word2vec model [
As explained in Section 1, in order to lift /lower a given standard to/from the reference semantic model we need to state the mapping between their concepts and structures in the annotation phase. This section describes the proposed method for the automatic generation of such mappings.
De nitions Let S and R be, respectively, the source standard and the reference semantic model. We indicate by XS (resp., XR) the structure used by S (resp., R), and by OS (resp., OR) the set of the vocabularies in XS (resp., XR). For example, XS could be de ned through XSD or OWL. If MS is an instance of XS and MR is an instance of XR, we say that MS and MR are equivalent if they are used for the same purpose in S and R|i.e., they are semantically equal.
We consider instances Mi (where i 2 fS; Rg) in which we can identify a root concept. Then, Mi is de ned using a sub-tree WXi of Xi, which is based on a vocabulary VOi that is a subset of Oi (i.e., VOi Oi). We write
Mi 2 T (VOi ; WXi ) to indicate that Mi is built on sub-tree WXi using the terms of VOi .
Given Ms, VOS , WXS , OR and XR, we aim to de ne a Map method that maps the concepts appearing in MS to concepts of XR, thus building an instance MR of R that is equivalent to MS:
MR = Map(MS; VOS ; WXS ; OR; XR) Assumptions For applying the proposed method, we assume that the following four premises are true. Although some of them might not be true in general, our aim is to automate the mapping process in most cases, sacri cing completeness to obtain e ciency, thus we deem these simpli cations to be acceptable and general enough. We brie y discuss in Sect. 4 how the last one can be relaxed. Assumption 1 The language in both sides of the mapping is English. Assumption 2 Given that our method targets mappings between standards in the same domain and both involved systems cover the concepts of the domain, we assume that for each concept in the source system, we have at least one corresponding concept in the reference system.
Assumption 3 The corresponding instances in the source (i.e., the given standard) and target (i.e., the reference semantic model) formats include the same equivalent concepts; that is, for each concept in the source instance, there exists exactly one concept in target format (one to one relationship between concepts). Assumption 4 All concepts exist in the word2vec model.
Procedure Figure 2 depicts an overall work ow of the proposed procedure. In order to map the source data to the reference data format, the rst step is to decompose the source data to its components: VOS , which is the set of terms that exist in MS, and the tree representation WXS of the given structured data. Then, the semantic equivalent of the main concept in the source data should be determined, where the main concept is the root of the tree. According to Assumption 2, there should be at least one concept in the reference data model corresponding to the source main concept. To detect it, the extracted root of the tree structure in the source data (WXS ), would be embedded to its 300dimensional vector space employing the word2vec model. If the main concept is a phrase, its atomic parts should be embedded individually and then averaged. Since the word2vec model identi es semantically close concepts based on their relative distances in the 300-dimensional vector space, we search the space for the vector that is nearest to the one of the source main concept and tag it as the equivalent concept in the reference system.
After determining the equivalent main concepts, the structures corresponding to that particular main concept in the source and reference systems (WXS and WXR ) are retrieved. Then, inside the tree structures of each data format, all the Decomposition Main Concept Identification (Word2vec) Retrieve the structure with equivalent
root Translation (Word2vec) PreBooking
Source Standard
Passenger
Name Path
Date Registering
Pre Pre Booking Registering
Source Destination
MR PreRegister
Reference Semantic
RouteSource RouteDestination RouteDestination TravelerName MS PreBook PassengerName PreBook Path Source PreBook Path Desti.
PreBook Date
MS Structure
PreRegister
RouteSource Passenger Traveler
Name Name
RouteSource ? RouteDestination ? RouteDestination ? TravelerName ?
MR PreRegister PreRegister PreRegister PreRegister
RouteSource RouteDest.
Time TravelerName possible routes from the root to the leaf nodes are extracted. To reach a single leaf node, there is only one path from the root. Each extracted route consists of the set of all the terms from the root to that particular leaf node. According to Assumption 3,1 the structures of the equivalent main concepts in the source and reference data contain the same number of attributes, hence the number of leaf nodes in both data formats are equal, and so is the number of routes.
For each route in the tree structure of the source data, its corresponding route in the reference data structure should be identi ed. To do so, a vector for each of the words in the current route is obtained using the word2vec model, and then the average of such vectors must be calculated. We expect that the vectors that are the result of the averaging operation in the corresponding routes in the source and reference data fall closer to each other in the vector space. This is possible due to the properties of the word2vec model, in which combinations of words can be meaningfully represented through vector addition. Similarly, the average operation of the vectors preserves their semantics, since it just scales the vector magnitude by a positive number, leaving the direction unchanged. We assign the route from the source data structure to the nearest average of the existing routes in the reference data structure. Then, the attribute name of the 1 As mentioned in Sect 3, we trade-o generality for e ciency. In future works we will look to relax the assumptions underlying the approach. leaf node from the source data is mapped to the corresponding leaf node in the reference data format and their attribute values are transferred accordingly. 4
This section outlines some of the possible challenges that might arise while applying the proposed method. The rst one is related to the extracted concepts and occurs because the individual concepts in the structure are typically a combination of two or more words, for example pre booking or PreBooking. These kinds of compound words usually do not exist in the word2vec model as a single word. To address this issue, it might be necessary to perform a pre-processing step, which could possibly be splitting the compound words to their atomic stems and then computing their average.
Another challenge could be due to the absence of some terms of the source or reference ontologies, which prevents the averaging in the route matching step. To tackle this problem, one possible approach is further training the existing word2vec model, using the transfer learning technique. To this end, a su cient number of unstructured texts containing the missing words are necessary. Although, instead of unstructured text, it might be possible to perform transfer learning using instances of structured data formats containing the missing words, either by attening the structured text to make it unstructured, or by adding extra layers to the word2vec model.
Finally, to validate the method, a possible approach could consist in preparing a dataset containing a set of pairs (M1; M2) of equivalent instances in di erent data formats (hence, with the same main concepts). Then, the Map method should be applied to one element from each pair (say, M1) and the result should be compared to the true data structure and the terms of the other element of the pair (i.e., M2). Subsequently, the direction of the Map method should be reversed (i.e., it should be applied to M2) and the same process should be repeated. The method is validated if both mapped data are equivalent to the corresponding true data formats.