Online Datasets in Open Data Portals typically have minimal metadata and users wishing to consider their reuse in extended analyses are poorly served. One approach is to find and re-annotate the metadata according to subject-specific, community adopted vocabularies. In support of this we explore a multi-tiered framework combining the capabilities of a crawler, a tagger and a recommendation engine, as well as tools for the provisioning of data as discoverable services. We provide details of prototype scale implementations of these components and a cursory evaluation of the tagger for subject-specific metadata enrichment using the Global Agricultural Concept Scheme (GACS).
1.1
A rudimentary search for Open Data tagged with the term agriculture identified
datasets in a variety of country/geography specific portals [
Earlier approaches to publishing structured Open Data have leveraged
community adopted controlled vocabulary terms and dataset definitions expressed
in Resource Description Framework (RDF) serialization formats, known as as
Linked Open Data [
In recent years, the Global Open Data for Agriculture and Nutrition
(GODAN11) project has advocated for the publication of open data and the creation
of ecosystems where agricultural data is Findable, Accessible, Interoperable, and
Reusable (FAIR) [
The current state of the ODPs containing agricultural data provides a good motivation for the creation of a dedicated infrastructures that supports comprehensive Open Data exploration for potential reuse. Primarily, users want to i) search for and query across globally distributed agricultural datasets based on multiple keywords and defined relations, and ii) retrieve integrated data in 2 https://catalog.data.gov/dataset 3 https://data.gov.uk/ 4 https://www.data.gouv.fr/en/datasets/ 5 https://data.gov.au/dataset 6 https://open.canada.ca/data/en/dataset?portal_type=dataset 7 https://data.overheid.nl/data/dataset 8 https://africaopendata.org/dataset 9 https://opendatacharter.net/agriculture-open-data-package/ 10 https://www.w3.org/TR/sparql11-overview/ 11 https://www.godan.info/ a unified standard format so that they are compatible and readily usable with third party tools.
In order for an infrastructure to support these capabilities it needs to address the following tasks: (i) regular crawling of the Web for sites related to agriculture, (ii) screening of Open Data files and indexing them, (iii) downloading and scanning the files for key agriculture vocabulary terms, (iv) generating subject specific metadata for the data files, (v) recommending relevant datasets based on curated metadata, (vi) change management and revision of metadata, (vii) provision of data resources as discoverable Web services, and (viii) publishing data according to interoperability standards.
In this paper we propose a multi-tier framework, Section 2, for the harvesting of Open Data files, subject specific enrichment of metadata, and the provisioning of Open Data as services. Using the target use case of Open Data about agriculture and leverage of the Global Agricultural Concept Scheme (GACS) we provide details of prototype scale implementations, Section 3, and a cursory evaluation of the tagger in, Section 4. In Section 5 we briefly discuss the framework in the context of the target functionality and list future work. Section 6 contains concluding remarks. 2
The multi-tier framework presented in in Figure 1 provides a solution to support better discovery and reuse of Open Agricultural Data.
As shown in Figure 1, the Data Sources column displays two sources of data:
i) Open Agricultural Datasets which are generated and collected based on
typical agricultural activities, and ii) the Controlled Vocabulary of Agriculture and
Nutrition such as the Global Agricultural Concept Scheme (GACS) consisting
of standard vocabularies which are agricultural concepts mapped from three
well known sources: the AGROVOC multilingual agricultural thesaurus by the
Food and Agricultural Organization (FAO) of the United Nations, the CAB
Thesaurus by the Centre for Agriculture and Biosciences International (CABI),
and the NAL Thesaurus by the US National Agricultural Library [
In Phase 1, the country-specific ODPs hosting the Open Agricultural Datasets are crawled and indexed for further processing. The crawler uses seed URLs of the ODPs as inputs, fetches contents such as text, data, and hyperlinks from recursively-linked pages, parses and stores them as segments, from which an index is then created. Off-the-sheft crawlers12 and indexers13 can also be used for this purpose.
In Phase 2, the index is enriched and updated using a tagger Individual data files are downloaded and parsed, and relevant tags are added based on a custom scoring algorithm that ranks words matching to the controlled vocabulary.
In Phase 3, a Semantic Recommendation System is used to suggest relevant datasets to end users, which can then be further curated in preparation for 12 http://nutch.apache.org/ 13 http://lucene.apache.org/solr/ integration with other datasets. Further enrichment of metadata using mapping to external ontologies can be incorporated also.
In Phase 4, access to data as services is provided using SADI Semantic Web
services [
The crawler in Phase 1 recursively scans through the ODP pages and sub-pages describing each dataset and their URLs. The crawler saves this information locally in segments which are parsed and structured into fields by an indexer. An index of the datasets containing descriptions and metadata is created. The file formats currently supported by the crawler are Zip (.zip), Microsoft Excel (.xls, .xlsx), Portable Document Format (.pdf), Comma-separated values (.csv) and Text (.txt). Similar functionalities are provided by the recently introduced Dataset Search14 by GoogleTM. 3.2
The tagger in Phase 2 is used to enrich the descriptions of the datasets by adding metadata from expert-authored controlled vocabularies. The core features of the tagger are the use of (i) an in memory vocabulary graph generated from a controlled vocabulary file and (ii) a custom scoring algorithm based on lexical matching of terms in data files to the terms in the vocabularies.
The current implementation of the tagger uses the vocabularies from GACS to create a graph where the nodes in the graph are terms or concepts. Before a node is created in the vocabulary graph, stemming is applied so that each term is reduced to its root form. The concept hierarchies of the vocabulary contain both broader concepts (as in superclass in ontologies) which identify parent nodes and narrower concepts (as in subclass in ontologies) which identify child nodes. Scoring of Annotations The tagger reads each word from the input data file and applies stemming. It then searches for both an exact match and a stem match in the vocabulary graph. If a lexical match, with or without stemming, to a concept is detected, a score is added to the term and to each of its broader concept terms in the graph based on their depth in the hierarchy. The narrower concepts (more specific and lower down the hierarchy) are assigned lower scores to avoid the selection of concepts that don’t provide significant information. Once scoring is complete, the upper 3rd percentile of concepts are selected as annotations for the document. This provides a barrier excluding tags that are unrelated to the content of a document but are still contained in it, such as terms from sources and references. The current deployment of the tagger excludes matches to geographical locations because of their widespread use and marginal relevance in the current study.
ing provides additional tagging to the datasets a simple example is shown for illustration and intelligibility purposes. The tagger was run on the dataset titled 14 https://toolbox.google.com/datasetsearch Wheat/Barley and their Products15 hosted at the Open Government Portal16 maintained by the Government of Canada. This file contains mentions of Wheat and Barley but not Cereals.
Table 1 shows tags annotated to the Open Data file with and without the introduction of the scoring technique. Without the implemented scoring technique (tagging of lexical and stem-based matches to GACS) the tagger can identify only terms directly mentioned in the files. Using the adopted scoring technique the term Cereals, the parent term for Wheat and Barley in GACS is retrieved.
Tags with lexical/stem matching import, export, wheat, barley, permits
Tags with scoring cereals The GACS hierarchy17, shown below, for the preferred term wheat illustrates how the broader concept cereals is related to the narrower. Moreover, the scoring can be extended to retrieve multiple parent terms in the hierarchy including cases where multiple inheritance may occur.
... > crops > f ieldcrops > graincrops > cereals > wheat
Metadata and tags provided when the file was submitted to an ODP can be enriched in a systematic way by using the tagger, namely with lexically matched terms found in GACS. The scoring algroithm additionally provides subject specific tags that are broader in scope. In the subsequent phase of the framework only the enriched datasets, including the lexically matched tags and the broader augmented tags, are used by the recommendation engine to filter and categorize data according to users’ interests, Phase 3. 3.3
The recommendation engine in Phase 3 currently uses a content-based filtering method, where extensive tagging of data files and custom scoring of matched tags is employed to determine the level of similarity between files. The engine uses the initial preferences of a user, which can be obtained from tagging an online publication specified by the user.
Upon request for recommendation, all datasets are scored according to their relevance to the tags within the user’s profile. Scoring is done by multiplying the normalized weight of each tag by the normalized weight of a matching tag within the user’s profile. The cumulative score for each document is then compared pairwise and the highest scoring documents are returned to the user as a 15 https://open.canada.ca/data/en/dataset/3a4e7f9b-64d2-432f-8394-15f6814aad62 16 https://open.canada.ca/en 17 http://browser.agrisemantics.org/gacs/en/page/C212 recommendation. Additionally, a history of the suggested files is stored within the user’s profile to avoid repeat runs offering the same recommendations. The engine was tested for both programmatic functionality and the quality of the recommended datasets. Preliminary test results show the greater the numbers of annotations, the better the relevance of the recommended datasets. Extension of the recommendation engine will include the use of additional community developed ontologies and inferencing based on subsumption, transitivity. 4
The tagger was run on a machine running Ubuntu 17.10 server with a 4-core 3 GHz processor and 8 GB memory. During the experiment, the tagger tried to match data from 212 CSV datasets hosted on FAOStat18 and Data.gov19 to the beta version of GACS controlled vocabulary. The outcome of the initial experiments showed that the scoring worked surprisingly well for most of the datasets. As is to be expected, the tagger worked best when data files contained meaningful agriculture related terms and performed worst when data files contained terms mostly as names, identifiers and numeric values.
Table 2 shows an analysis of results derived after running the tagger on 5 random datasets. The Topics column indicates what type of information the data files contain, the Tags column indicates if the value of a score crossed the threshold to select any tags or not, and the Outcome indicates whether the matching performance of the tagger is best case, moderate case, worst case or resulted in a false positive. For some data files the selected tags were found to be false-positive as well as false-negative. Due to space constraints a rigorous analysis of the tagger is beyond the scope of this paper. However, in testing it was found that Open Datasets are very broad in scope and their composition is complex as they often are published as spreadsheets, invoices, and statistical reports. Often, the rows and columns can only be explained by an expert in the subject area or by the data provider. It is also difficult to interpret when numerical values with units are present.
Thus, although automatic tagging may work for some datasets, for many other datasets it is prone to errors. Therefore, it is recommended that the tags added automatically be verified manually by experts before approving the files for use in the recommender system in the subsequent phase. 5
We have outlined a framework designed to address the challenges described in Section 1.3. In addition, we have been able to corroborate the general feasibility of our approach in so far as harvesting, tagging and recommending files to users. At the current time the tools implemented in this framework 18 http://www.fao.org/faostat/en/#data 19 https://www.data.gov/ Title of the dataset Incidental catch at BC marine finfish aquaculture sites Adult Salmon Health (Snorkel Surveys) Cape Breton Highlands USDA FSA Farm Payment Name/ Address File for 2008 USDA FSA Farm Payment File for 2010 Pineapple - Average retail price per pound and per cup equivalent, 2013
Topics Tags Time, location, 321 Words were facility, common tried, 266 matches, Best and scientific and 9 top scoring case name of the fish tags
58 Words were waterbody, species, tried, 51 matches, Moderate age and quality and 2 top scoring case tags
Outcome Names and addresses Identifiers and numerical values Packages and market price
None matched None matched fertilizers
Worst
case
Worst
case
False
positive
are yet to mature and more experiments are required to assess and improve
their performance. The idea of harvesting files in ODPs likely motivated the
development of Dataset Search by GoogleTM where users are provided with an
overview of the metadata assigned by the original publisher of the datasets. In
our pilot studies, we were able to further enrich the metadata for individual files
providing agriculture specific tags from GACS that extend beyond the metadata
provided by the dataset publisher. Compared to the techniques described in
related work [
With end users in mind, the recommendation system we implemented was
designed to support users who are looking for recently published candidate data
files and consider them for reuse. In addition, it can support users wishing to
participate in crowdsourcing and provisioning of data as services. Indeed, the
greater goal for the framework includes the provision of Open Data as services
over which ad hoc queries can be run. This is possible if the data can be
sufficiently well structured, annotated with metadata and could support meaningful
queries across data sets. Given that our system is still in development and since
we have not processed large volumes of Open Data files we have yet to
determine the extent to which Open Data files can be readily made available as
services. We have proposed to leverage SADI Semantics Web services given that
registries of SADI services, along with associated query tools, can support the
target functionality where complex workflows of combined data retrieval and
data analytics services can be run. Moreover we can point to recent work where
researchers [
We have presented a prototype to annotate Open Data files with subject specific tags on agriculture. The target objective is to make Open Data in ODPs more discoverable and intelligible for potential data reuse purposes. We have proposed to do this using a multi-phase approach involving crawling and indexing of Open Datasets, a custom tagging approach leveraging lexical term matching and a scoring algorithm. Files enriched with tags in this way are then made available to a recommendation engine to support alerting of end users. Subsequent to this we proposed the provisioning of data as services with semantic descriptions to support ad hoc federation of data in response to complex user queries.