We present RealFoodTrade (RFT), a system that allows farmers and fishermen to sell their products directly to the end-buyer. RFT makes use of Linked Data sets, together with a domain ontology designed by experts, to perform semantic search over products on sale. RFT employs geo-location technology on mobile devices to match demand and supply according to the location. We sketch the semantic search techniques in RFT and illustrate a prototype tailored to the fishing industry.
In this paper we introduce the system Real Food Trade (RFT), which provides
a marketplace for any food producer to sell their produce directly, freeing the
producer both from the wholesaler and from the effort of retail. The main idea
is that the producer uses the system to advertise flash stands, each of which
consists of (
In RFT, buyers search for flash stands selling a certain product within a certain area. The wanted product is specified by keywords, and the system identifies the stands that sell products which are semantically close to the input keywords. To do this, we aim at employing novel techniques for keyword search, employing as data sets both the ones available in the Linked Open Data cloud1 and “handcrafted” ontologies. We intend to empower the ontologies of the latter type with the less-polished, but richer, information available at Linked Data stores.
Related work. Several techniques have been proposed to automatically calculate the semantic relatedness between words, texts or concepts in a way corresponding closely to that of human subjects. Some of the commonly used methods derive statistical information from text corpora and combine that information with lexical sources [8,2]. The lack of domain-specific coverage of the resources used by these measures makes them ineffective for use in domain specific tasks where the context plays an important role. Most classical methods to compute semantic measures exploit particular lexical sources: corpus, dictionaries, or well structured taxonomies such as WordNet [7]. Some of these methods explore path lengths among nodes in taxonomies [10]; others exploit textual descriptions of concepts in dictionaries [8], while a last group relies on annotated corpora to compute information content [3,9]. Some recent research efforts has focused on using Wikipedia to improve coverage with respect to traditional thesauri-based methods [11,5]. In particular, the work in [5] proposes a method to represent the meaning of texts or words as weighted vectors of Wikipedia-based concepts, using machine learning techniques. Wikipedia seems to be unable to effectively discover and evaluate implicit relationships [6]. In order to guarantee maximum coverage, the work in [6] focuses on methods that exploit the Web as source of knowledge. In [1] the authors propose a similarity measure that combines various similarity scores based on page counts, with another one based on lexico-syntactic patterns extracted from text snippets.
In the rest of the paper we will describe the Real Food Trade system, and how it deals (and plans to deal) with the problem of keyword search on ontologies represented as Linked Data. 2
In this section we sketch how RFT works and overview its architecture, as shown in Fig. 1.
The system exploits the Google data cloud driven architecture2. In particular the essential components in the proposed mobile solution architecture are: – Android mobile clients; – Google Cloud Endpoints used for communications between the clients and the backend over REST API with OAuth2 authentication; – A backend application code running on Google App Engine3 and responsible for serving requests from the clients.
A typical requirement for a mobile solution with a backend is to store data outside of client devices. These data can be categorized into two groups: (i)
large, and typically binary objects (e.g. images) and (ii) fine-grained properties and entities, which can also include a reference, for example, object name and optionally bucket name or URL, to objects stored in Google Cloud Storage.
Both groups of data are managed by two modules in RFT: (i) App Engine Authentication and (ii) the App Engine Query Analysis. The first application is responsible to store and manage all user profiles. In particular we embedded OpenID4 technologies to exploit existing email or social network accounting (e.g. Gmail, Yahoo, Facebook, twitter and so on). Once the user is authenticated, the App Engine Query Analysis is responsible to analyze the user request and to invoke several instances of the application to retrieve all the data best fitting the user query. To this aim, the App Engine Ontology Manager is responsible to implement semantic matchmaking techniques exploiting the Linked Open Data cloud.
In particular, RFT employ different RDF taxonomies that characterize the domains of interest. Such taxonomies are indexed in the Google cloud and linked to the LOD network (i.e. DBPedia). Then by using different heuristics (i.e. user rating, usage statistics, and lexicographic analysis) we rank the similarities between concepts in the taxonomies. In this way RFT can exploit a weighted similarity search based on the rank. Intuitively depending on how many products the user would view, the similarity distance from a concept matching the
required product can increase or decrease. In the next section we will discuss a case study for our system.
Finally, a natural place to store this kind of data is App Engine Datastore. It provides a NoSQL schemaless object data store, with a query engine and atomic transactions. These entities will often map to the Resources exposed over Google Cloud Endpoints API. 3
As case study for RFT, we analyzed Fish & Seafood Markets in Chile. In particular we have two specific users: the producer (i.e. the fisherman) and the buyer (private or business). The producer is interested to spot the available produce by exploiting e-commerce technology (e.g. eBay model); in this case RFT provides real time points of interest (POIs) through Google Maps representing market announcements for all people interested in buying fish. Such POIs are invoked by producers and generated through RFT.
For instance, Fig. 2 shows our system at work: (i) the user compiles a keyword search query, (ii) the system retrieves all POIs on the map and (iii) the user selects and reserves items.
In this scenario the keyword search query processing involves different semantic tasks. To this aim, we adopted the Network of Fisheries Ontology5 provided 5 http://aims.fao.org/network-fisheries-ontologies by the Food and Agriculture Organization of the United Nations (FAO). Such ontology is a fine-grain classification of species: the classification is made by both biological nature and regional proximity. Whenever a seller (fisherman) enters a product, typing its name, a concept (corresponding to a particular type of fish) in the ontology is associated to the product; such a node is the semantically closest to the entered one. The association between the input and the concept in the ontology is established in a fashion similar to that of [4], where ontological information extracted off-line from Linked Data sets is also employed. Similarly, whenever a buyer searches for a particular kind of fish, the semantically closest one in the ontology is returned, as well as other ones within the same species (alternative names for the same fish, or sub-species). The match is also performed during the input (when only a partial input is entered), so that the user can receive suggestions from the interface in order to save input time — this is especially useful when entering input on mobile devices. 4
We have sketched the main features of the RFT system, which provides a location-based infrastructure for the sale of food. After having briefly illustrated the architecture of the system and the technologies employed in it, we have presented a case study, where we employ RFT to fisheries in Chile, in particular small fishing boats. We have shown how we integrate the FAO’s Network of Fisheries Ontology with linked data sources in order to match entities in the ontology with fish names entered by users (fishermen or buyers).
We plan to extend the field of application of RFT to other markets, in
particular that of agriculture. This will require the automated extraction of high-quality
ontologies from Linked Data sets. Moreover, we intend to incorporate learning
algorithms that incorporate knowledge from users’ behaviour in order to enhance
the knowledge base used in the system. Last but not least, we intend to carry
out extensive experiments on real fish markets and report on the results, both
from the system point of view and from the socio-economic point of view.
Acknowledgments. Andrea Cal`ı acknowledges support by the EPSRC project
“Logic-based Integration and Querying of Unindexed Data” (EP/E010865/1).
3. Budanitsky, A., Hirst, G.: Evaluating wordnet-based measures of
lexical semantic relatedness. Comput. Linguist. 32(