The RICHFIELDS Framework for Semantic Interoperability of Food

Information Across Heterogenous Information Systems

Tome Eftimov

, Gordana Ispirova

1, 2

, Peter Koro

sec

1, 3

and Barbara Korou

c Seljak

Computer Systems Department, Jo

zef Stefan Institute, Jamova cesta 39, 1000 Ljubljana, Slovenia

zef Stefan International Postgraduate School, Jamova cesta 39, 1000 Ljubljana, Slovenia

Faculty of Mathematics, Natural Sciences and Information Technologies, Glagolja

ska ulica 8, 6000 Koper, Slovenia

Keywords:

Semantic Interoperability, RICHFIELDS Ontology, Food Information, Ontology Population, Semantic

Annotation.

Abstract:

In an EU-funded project RICHFIELDS, a data platform was designed with the aim to collect, link and harmo-

nize, analyze, store, and deliver food- and nutrition-related data and information to various stakeholders. To

integrate heterogenous food data sets, we propose a RICHFIELDS framework for semantic interoperability

of food information, which is a combination of already developed NLP approaches for the food domain. The

framework includes i) a food ontology to which foods are linked, ii) a part that explains how the relevant foods

can be extracted and represented in a structured way, and iii) a similarity measure that is used to link the foods

to the ontology. To evaluate the RICHFIELDS framework, we selected two distinct data sets from different

food information systems. The experimental results provided promising results,i.e., 81.5% and 87.5% of the

foods from the ﬁrst and the second data set, respectively, obtained a tag from the ontology (i.e., semantic

annotation was performed). The annotations provided by the framework allow automatic integration of food

information provided in both data sets.

1 INTRODUCTION

Creating a healthy diet requires a lot of informa-

tion and knowledge from food science. Nowa-

days, there are many information systems that pro-

vide food- and nutrition- related data. These sys-

tems can be either: a scientiﬁc cloud (e.g., European

Open Science Cloud (EOSC, 2018), Zenodo (Zenodo,

2018), and FigShare (FigShare, 2018)), a server (e.g.,

Quisper (QuaLiFY, 2018), EuroFIR (EuroFIR, 2018),

and GS1 GDSN (Global Data Synchronisation Net-

work) (GDSN, 2018)), or application (e.g., PRECI-

OUS (PRECIOUS, 2018), FitBit, Twitter, and Face-

book). Each system uses its own way of describing

information. In order to exchange data with unambi-

guous, shared meaning, semantic interoperability is a

requirement to enable machine computable logic, in-

ferencing, knowledge discovery, and data federation

between information systems. This is made by ad-

ding metadata about the data, linking each data ele-

ment to an ontology (i.e., semantic data model). For

this reason, in autumn 2015, the H2020 project RI-

CHFIELDS started with the aim is to collect, link

and harmonize, analyze, store and deliver food- and

nutrition-related data and information to various sta-

keholders. Data may be of any type, i.e., structured,

semi-structured or unstructured; small or big; open

or linked, raw or aggregated. To make this possi-

ble, semantic enrichment should be applied to solve

some of the most common problems, allowing for:

effective search in databases, integration of hetero-

geneous data sets, faster information retrieval, regu-

larly updated domain knowledge, etc. Since seman-

tic enrichment involves adding metadata to the data,

or linking speciﬁc data to an ontology, in the case of

RICHFIELDS, a domain ontology that covers food-

and nutrition-related domain should be speciﬁed. Ha-

ving such a representation of the domain, there are

several questions that must be addressed: what type

of data (e.g., structured, semi-structured, or unstruc-

tured) needs to be harmonized; if the data is unstruc-

tured, how we can extract the relevant data that should

be linked to the domain ontology; what is the simila-

rity measure that will be used for linking data to the

ontology; etc. To address such questions in the case of

RICHFIELDS, we propose a framework for semantic

interoperability of food information across heteroge-

nous information systems.

Eftimov, T., Ispirova, G., Korošec, P. and Seljak, B.

The RICHFIELDS Framework for Semantic Interoperability of Food Information Across Heterogenous Information Systems.

DOI: 10.5220/0006951703150322

In Proceedings of the 10th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2018) - Volume 1: KDIR, pages 315-322

ISBN: 978-989-758-330-8

315

The paper is organized as follows: Section 2 gives

an overview of the related work. Section 3 introdu-

ces the RICHFIELDS framework for semantic inter-

operability of food information. Section 4 presents a

RICHFIELDS case study, in which discussion of re-

sults is presented. The conclusions are presented in

Section 5.

2 RELATED WORK

To make ambiguous and heterogenous content inter-

linked, semantic interoperability plays an important

role. It involves two steps: selecting or developing

an ontology that describes the domain, and enriching

relevant data with metadata, which are machine-

processable data pieces (i.e., tags from the ontology).

However there are subquestions related to each step

that need to be answered.

Having a domain-ontology that describes the

domain in general is a challenging task because each

ontology is developed for a speciﬁc application sce-

nario. In the domain of food, several food ontolo-

gies already exist, such as: FoodWiki, AGROVOC,

Open Food Facts, Food Product Ontology, Foods, and

FoodOn. A detailed review of the aforementioned

food ontologies is provided in (Boulos et al., 2015).

The problem of generating a general food domain on-

tology has been partly solved by the QuaLiFY Euro-

pean project (http://quisper.eu/), where existing food

information systems were explored by scientiﬁc bo-

dies like EuroFIR (European Food Information Re-

source Network) and NuGO (http://www.nugo.org/).

After a domain ontology is selected, the relevant

data should be linked to the ontology. The questi-

ons that appear here are: how can we extract the rele-

vant data (e.g., especially if the data is unstructured),

and how the extracted data can be linked to the onto-

logy in an automatic way. If the data is represented

as structured or semi-structured, different rule-based

approaches could be applied in order to extract the

relevant data that further will be linked to the onto-

logy. On the other hand, when the data is unstructu-

red (i.e., represented as text), a more complex scena-

rio is presented, in which information extraction (IE)

methods should be applied in order to extract the re-

levant data. Nowadays, the IE from the biomedical

literature is a very important task in order to improve

public health. IE is a task of automatically extracting

information from unstructured data and in most cases

concerns processing of human language texts by me-

ans of natural language processing (NLP) (Aggarwal

and Zhai, 2012; Piskorski and Yangarber, 2013). The

information to be extracted is predeﬁned by users, and

consists of predeﬁned concepts of interest (entities),

relationships between them and events. One of the

classic IE tasks is named-entity recognition (NER),

which addresses the problem of the identiﬁcation and

classiﬁcation of predeﬁned concepts (entities). Va-

rious NER methods exist: terminology-driven NER

methods (Miller et al., 1992; Aronson, 2001; Zhou

et al., 2006), rule-based NER methods (Farmakiotou

et al., 2000; Petasis et al., 2001; Hanisch et al., 2005),

corpus-based NER methods (Rindﬂesch et al., 2000;

Rockt

aschel et al., 2012; Alnazzawi et al., 2015; Le-

aman et al., 2015), NERs based on active learning

(Settles, 2010), and NERs that use deep neural net-

works (Collobert and Weston, 2008; Collobert et al.,

2011; Chiu and Nichols, 2015; Huang et al., 2015;

Santos and Guimaraes, 2015; Lample et al., 2016;

Habibi et al., 2017; Lopez and Kalita, 2017). Be-

cause NER methods with best performances are usu-

ally corpus-based NER methods, there is a need for

annotated corpus from biomedical literature that will

include the entities of interest. For this purpose, diffe-

rent annotated corpora are produced by shared tasks,

where the main aim is to challenge and encourage re-

search teams on NLP problems.

To allow automatic integration of extracted infor-

mation from a NER task, the information needs to be

further processed. The problem that appears is that

the same entity can be mentioned in different ways

in the same or different documents, using different

phrases regarding the text variability. To collect the

information for a given entity or even more to com-

bine the information for the entity from different do-

cuments, it is crucial to map the entity to a concept

that exists in a terminological resource (i.e., an onto-

logy). By mapping it to a concept from a terminolo-

gical resource, the extracted entity receives a unique

identiﬁer which is the identiﬁer for that entity in the

terminological resource. Having unique identiﬁers

helps the process of collecting and combing the in-

formation for some entity, even if it has different tex-

tual representations. The process of automatic map-

ping between an entity in text and a concept in a ter-

minological resource is known as text normalization.

Many normalization methods are based on string si-

milarity measures. String similarity measures give us

a metric for similarity (or dissimilarity) between two

text strings (Metzler et al., 2007; Gomaa and Fahmy,

2013). They can be performed on a character, term

level, or a mix of both. Also, there are methods that

use post-processing rules applied after the concept is

extracted or regular expressions to ﬁnd some matches

for which a speciﬁc form may not occur in the ter-

minological resource (Ramanan et al., 2013). Some

normalization methods are based on ranking techni-

KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval

316

que in order to rank the candidate matches and then

to ﬁnd the most relevant one (Collier et al., 2015). An

example of automatic normalization method focussed

on phenotypic information, which integrates a num-

ber of different similarity measures, is presented in

(Alnazzawi et al., 2016). Normalization methods can

also use ML algorithms to improve results, which was

shown in the gene normalization task as part of Bio-

Creative II (Morgan et al., 2008) and BioCreative III

(Lu et al., 2011). To the best of our knowledge, no

previously reported automatic normalization method

has focussed on the food domain.

3 THE RICHFIELDS

FRAMEWORK FOR SEMANTIC

INTEROPERABILITY

Let us assume that food information systems share

data sets that should be integrated. To make them

understandable in machine-readable format, the RI-

CHFIELDS framework for semantic interoperability

is proposed. The framework is presented in Figure 1

and consists of four steps:

1. Select the domain ontology;

2. Apply pre-processing of the data with regard to its

type to obtain structured data;

3. Link the structured data to concepts that exist in

the ontology using a similarity measure;

4. Perform semantic annotation or ontology popula-

tion.

RICHFIELDS

ontology

Food Information Systems

Is the information

unstructured?

. . .

Pre-processing

to select only

food information

No Yes

Food Information

Extraction

(i.e. NERs)

. . .

Semantic

Annotation

Ontology

Population

Figure 1: The RICHFIELDS framework for semantic inter-

operability.

The ﬁrst step is to select the domain ontology that

is a model to which the data will be linked and from

which the metadata will be used to make the data un-

derstandable in machine readable format.

Next, the type of the data should be deﬁned i.e.

structured, semi-structured, or unstructured. If we are

dealing with structured or semi-structured data, pre-

processing can be made by applying some heuristics

(e.g., rules based on regular expressions) in order to

select the data that will be linked. If the data is un-

structured, ﬁrst information extraction methods (i.e.

named-entity recognition methods, NERs) should be

applied to extract and structure the relevant informa-

tion that will be linked to the ontology.

The third step deﬁnes a similarity measure used

for linking the data to concepts that already exist in

the ontology. Since most of the data is presented as

text, different text normalization methods that involve

text similarity measures can be applied.

Finally, according to the value of the similarity

measures, semantic annotation or ontology popula-

tion should be performed. A threshold value for the

similarity measure needs to be deﬁned. If the value

of the similarity measure is greater or equal than the

selected threshold, a semantic annotation should be

performed. This means that this matching is a good

one, the concept of searching already exists in the on-

tology, so it should be annotated with the metadata.

In this case, the data set is changed by including tags

from the ontology. In the other case, when the value

of the similarity measure is lower than the selected

threshold, the data cannot be linked to the ontology

because such concept does not exist in the ontology,

so ontology population should be performed. This in-

volves adding an instance for such concept in the on-

tology.

In general, the above mentioned steps are familiar

for each framework used for semantic interoperabi-

lity, the questions that appear are related to informa-

tion extraction methods and the deﬁnition of the simi-

larity measure used for linking, since each domain is

speciﬁc and it follows that if some methods are good

for a speciﬁc domain it does not follow that they will

also be good for other domains.

4 RICHFIELDS CASE STUDY

We continue by explaining in detail each part of the

RICHFIELDS framework for semantic interoperabi-

lity. First, the food ontology is explained, followed by

the methods that can be used for food information ex-

traction from unstructured data. Then, the similarity

measure used for linking foods to the RICHFIELDS

The RICHFIELDS Framework for Semantic Interoperability of Food Information Across Heterogenous Information Systems

317

ontology is reintroduced. Finally, the results of lin-

king two food-related data to ontology are explained

and discussed.

4.1 The RICHFIELDS Ontology

The development of the ontology that is used by RI-

CHFIELDS started from the Quisper ontology, which

was previously developed by JSI as part of the EU-

funded project QuaLiFY (QuaLiFY, 2018). The onto-

logy consists of six super classes: Component, Food,

FoodGroup, Personal, Single Nucleotide Polymor-

phism, and Unit, which were further described with

data and object properties. RICHFIELDS covers a

wider domain than the QuaLiFY project and for this

reason, the Unit concept was replaced with a concept

with the same name from a widely-used ontology, cal-

led Units of Measurements Ontology (UO) (Gkoutos

et al., 2012), which is currently being used in many

scientiﬁc resources for the standardized description

of measurements units. Also, a new concept, Ma-

trixUnit, was added with the corresponding subcon-

cepts for matrix units that can be found in the Euro-

FIR Thesauri. Because this study was focused on se-

mantic interoperability of food-related data sets, the

RICHFIELDS ontology was updated by populating

the Food concept with 5,416 food individuals from

FoodEx2 data (EFSA, 2017), which were also descri-

bed with two data properties FoodName and FoodEx2

code that are speciﬁc for the FoodEx2 representation.

4.2 drNER

If we are working with unstructured data (i.e., repre-

sented as text) the ﬁrst step is to extract the relevant

food data that should be linked to the ontology. For

this reason, NERs should be applied. From an over-

view of the existing IE methods from the biomedi-

cal literature, a lot of NER methods that exist in the

domain of biomedical literature are focused on diffe-

rent biomedical domains. The commonly used NER

methods are the corpus-based NER methods that rely

on annotated corpus for the domain of interest, which

is produced by the domain experts. Several studies

are conducted in the dietary domain, but with dif-

ferent goals. For example, Xia et al. (Xia et al.,

2013) presented an approach to identify rice protein

resistant to Xanthomonas oryzae pv. oryzae, which is

an approach to enhance gene prioritization by combi-

ning text mining technologies with a sequence-based

approach. Co-occurrence methods were also used

to identify ingredients mentioned in food labels and

extracting food-chemical and food-disease relations-

hip (M

uller et al., 2004; do Nascimento et al., 2013;

Jensen et al., 2014). We did not ﬁnd any research

that focuses on extracting dietary information from

evidence-based dietary recommendations and for this

reason we recently proposed a rule-based NER met-

hod, known as drNER (Eftimov et al., 2016; Eftimov

et al., 2017c). It is a combination of a terminological-

driven NER and rule-based NER. The difference with

purely terminological-driven NERs is that we do not

only use dictionaries with concepts and synonyms (as

terminological resources), but we allow the reuse of

some corpus-based NERs that exist for some entities.

If corpus-based NERs exist for some entities we are

interested in, we use them to annotate text data and

then to see if some tokens have labels that correspond

to entities of interest. We also combine corpus-based

NERs that exist for some entities in which we are in-

terested, following the idea of ensemble learning in

order to achieve better performance than the perfor-

mance obtained from any corpus-based NER alone.

The difference with the rule-based NERs is that we do

not use rules associated with the characteristics of the

entities. This is because having rules for each of the

entities we are interested in requires too much time

and effort to produce them. We only used a small

number of Boolean algebra rules that are not related

to the characteristics of the entities, but help us deﬁne

the phrases that are the entities mentions. Evaluation

of the method showed that the method gives promi-

sing results and can be used for information extraction

of evidence-based dietary recommendations.

4.3 StandFood

When the foods are represented in structured way, the

next step is to link each of them to a concept that has

already existed in the ontology. To do this, text nor-

malization methods that are based on string simila-

rity measures should be applied. In (Eftimov et al.,

2017a; Eftimov et al., 2017b), we presented a met-

hod, known as StandFood, which is used for stan-

dardization of foods according to FoodEx2 that is a

comprehensive food classiﬁcation and description sy-

stem for exposure assessment introduced by EFSA

(EFSA, 2017). StandFood is a semi-automatic system

for classifying and describing foods and consists of

three parts: the ﬁrst classiﬁes the food concept into

one from four FoodEx2 categories (i.e., raw, deriva-

tives, simple composite, and aggregated composite)

using ensemble of classiﬁers, the second describes the

food concept using the FoodEx2 code using natural

language processing approach, and the third combi-

nes the results from the ﬁrst and the second part to

improve the result for the classiﬁcation part by deﬁ-

ning post-processing rules. As a similarity measure in

KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval

318

the RICHFIELDS framework, we use the description

part of StandFood. For this reason, we are going to

reintroduce this part that uses the POS-tagging proba-

bility weighted method (Eftimov and Korou

c Seljak,

2015).

For the similarity measure, let D

and D

are the

names of two foods that are linked. Let us deﬁne

= {nouns extracted f rom D

= {ad jectives extracted f rom D

= {verbs extracted f rom D

}, (1)

where i = 1, 2. To ﬁnd the similarity between these

two food names, an event is deﬁned as a product of

two other events

X = N(A +V), (2)

where N is the similarity between the nouns found

in N

and N

, and A + V is the similarity between

the two sets of adjectives and verbs handled toget-

her as A

and A

. The adjectives and verbs

are handled together to avoid different forms with the

same meaning. For example, if adjectives and verbs

are handled separately, the match ”apple dry” and

”dried apples” will not be a perfect match. To avoid

this, lemmatization is applied for each extracted noun,

verb and adjective, and the similarity event uses their

lemmas. Because these two events are independent,

the probability of the event X can be calculated as

P(X) = P(N)P(A +V ). (3)

For this, the probabilities of each of the two events

need to be deﬁned. Because the problem looks for the

similarity between the two sets, it is logical to use the

Jaccard index, J, which is used in statistics for com-

paring similarity and diversity of sample sets (Real

and J.M.Vargas, 1996). For the similarity between

the nouns, the Jaccard index is used, while for the si-

milarity between the adjectives and verbs the Jaccard

index is used in combination with Laplace probabi-

lity estimate (Cestnik et al., 1990), this is because in

some food names the additional information provided

by the adjectives or verbs can be missed, but the re-

levant match can be found, so there will be no zero

probabilities. The probabilities are calculated as

P(N) =

∩ N

∪ N

P(A +V) =

|(A

∪V

) ∩ (A

∪ N

)| + 1

|(A

∪V

) ∪ (A

∪ N

)| + 2

. (4)

By substituting Equation 4 into Equation 3, we obtain

a weight for each matching pair. Finally, the pair

with the highest weight is the most relevant found ma-

tch. More details about the POS-tagging probability

weighted method can be found in (Eftimov and Ko-

rou

c Seljak, 2015).

In the case of RICHFIELDS, each preproces-

sed data set that contains foods is linked to the RI-

CHFIELDS ontology in a way that each food con-

cept is linked to food individuals that exist in RI-

CHFIELDS ontology using the POS-tagging proba-

bility weighted method. At the end the pair with the

highest value of the similarity measure is selected as

the real one. However, it can happen that non returned

match is true. One reason for this could be that such

food concept does not exist as a food individual in the

RICHFIELDS ontology. For this reason the simila-

rity measure value that is returned as a match is furt-

her checked with a threshold value given as a priori

information, which in the case of RICHFIELDS is set

at 0.125 and it comes from experimental evaluations

performed on food matching problem. If the value is

greater or equal than 0.125 then the food concept in

the data set is annotated using the tag for the food in-

dividual from the ontology, otherwise we cannot ﬁnd

a match, the concept does not exist as individual in

the ontology, so the RCIHFIELDS ontology must be

populated with this concept.

4.4 Food Information Systems

To show how the RICHFIELDS framework for se-

mantic interoperability works, we used two food-

related data sets that are provided from two food in-

formation systems (i.e., PRECIOUS and GS1 GDSN)

that rely on different standards, which are related to

the same concepts but use different terminology and

classiﬁcation. PRECIOUS and GS1 GDSN provide

data in semi-structured form (i.e., JSON format and

GS1 XML format, respectively).

4.4.1 PRECIOUS

PRECIOUS is a mobile app for preventive health and

wellbeing care that was developed in the FP7 project

PREventive Care Infrastructure based On Ubiquitous

Sensing (PRECIOUS, 2018). It was decided to col-

lect different kinds of biometric data (e.g. nutrition,

physical activity, sleep, etc.). Our PRECIOUS data

set consists of 437 foods, some of them are described

in English and some in Spanish. An example of one

food concept from the PRECIOUS data set is given in

Figure 2.

4.4.2 GS1 GDSN

The GS1 Global Data Synchronisation Network is a

network of interoperable data pools enabling colla-

borating users to securely synchronise master data

The RICHFIELDS Framework for Semantic Interoperability of Food Information Across Heterogenous Information Systems

319

Figure 2: An example of food concept from the PRECIOUS

data set.

based on GS1 standards. GDSN supports accu-

rate, real-time data sharing and trade item updates

among subscribed trading partners. Currently avai-

lable GDSN standards for nutrition and health are

available at https://www.gs1.org/gdsn-standards. The

data provided by GS1 consists of 25 foods provided

by the GS1 Slovenia. All foods are available with

their Slovenian and English names.

4.5 Results

Since the data from PRECIOUS data set is semi-

structured, ﬁrst, by using regular expressions we par-

sed the document to structure it. Then, we split the

data set into two parts: Spanish names and English

names. The Spanish names are translated in English

using a simulation that involves scraping with Sele-

nium and Google Translate web site. After that, the

translated Spanish food names are merged with the

existing English food names, which results in one data

set that will be linked to the RICHFIELDS ontology.

For the GS1 GDSN data we used the English names

that are provided when we linked it to the ontology.

In the process of linking 81.5% of the foods from

the PRECIOUS data set obtained tags from the RI-

CHFIELDS ontology, while we need to populated the

ontology for the other 18.5% when the match does not

exist. In the case of GS1 GDSN, 87.5% of the foods

obtained their tags from the ontology, while 12.5%

were included as new food individuals.

An annotated example from the PRECIOUS data

set is presented in Figure 3, in which the food concept

is described by an additional RICHFIELDS tag that

is the tag for the same food concept that exists in the

ontology.

Annotated examples from the GS1 GDSN are pre-

sented in Table 1. The Global Trade Item Number

(GTIN) can be used by a company to uniquely iden-

tify all of its trade items. In our study, we presented

it as a string (e.g., GTIN1), in order not to invade the

privacy of the real data.

Figure 3: An example of annotated food concept from the

PRECIOUS data set.

Though each food individual is described with a

FoodName and a FoodEx2 code, in the case of onto-

logy population we cannot provide the FoodEx2 code

for the new individual. Ontology population hap-

pens when the food individual does not exist in the

ontology, which means that it does not exist in the

FoodEx2 data set since the ontology is populated with

all existing foods from the FoodEx2 data set. From

here, a new problem arises, which is how to generate

a FoodEx2 code for new food individual. This is also

one direction for our future work.

5 CONCLUSIONS

To allow integration of heterogenous food data sets,

faster information retrieval, and regularly updated

food knowledge, we propose a RICHFIEDLS frame-

work for semantic interoperability of food informa-

tion. The framework includes a food ontology that is

the resource to which data sets, which used different

standards to describe foods, are linked. Depending

on the data type (i.e., structured, semi-structured, or

unstructured), pre-processing should be applied to se-

lect and represent only food information in a structure

way. Then, each food concept is linked to the onto-

logy using a similarity measure. Depending on the

similarity measure value, semantic annotation or on-

tology population should be applied.

To show how the proposed framework works, we

used two food-related data sets that are provided from

two different food information systems, PRECIOUS

and GS1 GDSN. The experiment results provided

promising results, where 81.5% and 87.5% of the

foods from PRECIOUS and GS1 GDSN obtained a

tag from the ontology (i.e., semantic annotation was

performed), respectively. Further, the RICHFIELDS

ontology annotations allow automatic integration of

food information provided in these two data sets.

Food items, for which the linking does not give

good results (i.e., the food item does not exist in the

KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval

320

Table 1: Annotated examples of foods from the GS1 data set.

GTIN Food name (in Slovenian) English food name RICHFIELDS tag

GTIN1 Liker z limono Liqueur with lemonn http://www.semanticweb.org/tome/ontologies/2018/2/Richﬁelds#A03NS

GTIN2 Keﬁr z borovnicami Keﬁr with blueberries http://www.semanticweb.org/tome/ontologies/2018/2/Richﬁelds#A02NV

GTIN3 Sirup z malinami Juice with raspberries http://www.semanticweb.org/tome/ontologies/2018/2/Richﬁelds#A03CD

ontology), were used for ontology population. Howe-

ver, there is an additional challenge that arises while

performing this process, which is also one direction of

our future work. The problem is how to generate the

FoodEx2 code for a food that does not exist in the on-

tology, which can often happen when we are working

with composite foods (i.e., recipes).

ACKNOWLEDGEMENTS

This work was supported by the project from the

Slovenian Research Agency (research core funding

No. P2-0098), from the European Union’s Horizon

2020 research and innovation program under grant

agreement No. 654280 (RICHFIELDS), and from the

European Union’s Seventh Framework Programme

for research, technological development and demon-

stration under grant agreement No. 621329 (ISO-

FOOD). We would also like to thank the PRECIOUS

team from Aalto University and GS1 Slovenia for pro-

viding the data sets that are used in this case study.

REFERENCES

Aggarwal, C. C. and Zhai, C. (2012). Mining text data.

Springer Science & Business Media.

Alnazzawi, N., Thompson, P., and Ananiadou, S. (2016).

Mapping phenotypic information in heterogeneous

textual sources to a domain-speciﬁc terminological re-

source. PloS one, 11(9):e0162287.

Alnazzawi, N., Thompson, P., Batista-Navarro, R., and

Ananiadou, S. (2015). Using text mining techniques

to extract phenotypic information from the phenochf

corpus. BMC medical informatics and decision ma-

king, 15(2):1.

Aronson, A. R. (2001). Effective mapping of biomedical

text to the umls metathesaurus: the metamap program.

In Proceedings of the AMIA Symposium, page 17.

American Medical Informatics Association.

Boulos, M. N. K., Yassine, A., Shirmohammadi, S., Na-

mahoot, C. S., and Br

uckner, M. (2015). Towards an

internet of food: Food ontologies for the internet of

things. Future Internet, 7(4):372–392.

Cestnik, B. et al. (1990). Estimating probabilities: a crucial

task in machine learning. In ECAI, volume 90, pages

147–149.

Chiu, J. P. and Nichols, E. (2015). Named entity recog-

nition with bidirectional lstm-cnns. arXiv preprint

arXiv:1511.08308.

Collier, N., Oellrich, A., and Groza, T. (2015). Concept

selection for phenotypes and diseases using learn to

rank. Journal of biomedical semantics, 6(1):24.

Collobert, R. and Weston, J. (2008). A uniﬁed architec-

ture for natural language processing: Deep neural net-

works with multitask learning. In Proceedings of the

25th international conference on Machine learning,

pages 160–167. ACM.

Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavuk-

cuoglu, K., and Kuksa, P. (2011). Natural language

processing (almost) from scratch. Journal of Machine

Learning Research, 12(Aug):2493–2537.

do Nascimento, A. B., Fiates, G. M. R., dos Anjos, A.,

and Teixeira, E. (2013). Analysis of ingredient

lists of commercially available gluten-free and gluten-

containing food products using the text mining techni-

que. International journal of food sciences and nutri-

tion, 64(2):217–222.

EFSA ((accessed February 17, 2017)). The food classiﬁ-

cation and description system FoodEx2 (revision 2).

https://www.efsa.europa.eu/.

Eftimov, T., Ispirova, G., Koro

sec, P., and Korou

c Seljak,

B. (2017a). A semi-automatic system for classifying

and describing foods according to FoodEx2. In 3rd

IMEKO FOODS, Metrology promoting Standardiza-

tion and Harmonization in Food and Nutrition, pages

56–59.

Eftimov, T., Koro

sec, P., and Korou

c Seljak, B. (2017b).

Standfood: Standardization of foods using a semi-

automatic system for classifying and describing foods

according to FoodEx2. Nutrients, 9(6):542.

Eftimov, T. and Korou

c Seljak, B. (2015). POS tagging-

probability weighted method for matching the inter-

net recipe ingredients with food composition data. In

Knowledge Discovery, Knowledge Engineering and

Knowledge Management (IC3K), 2015 7th Internati-

onal Joint Conference on, volume 1, pages 330–336.

IEEE.

Eftimov, T., Korou

c Seljak, B., and Koro

sec, P. (2017c). A

rule-based named-entity recognition method for kno-

wledge extraction of evidence-based dietary recom-

mendations. PloS One, 12(6):e0179488.

Eftimov, T., Korou

c Seljak, B., and Koro

sec, P. (2016).

Grammar and dictionary based named-entity linking

for knowledge extraction of evidence-based dietary

recommendations. In Proceedings of the 8th in-

ternational Joint Conference on Knowledge Disco-

very, Knowledge Engineering and Knowledge Mana-

gement, (IC3K 2016), volume 1:KDIR, pages 150–

157.

EOSC (2018). European Open Science Cloud. accessed

June 12, 2018.

EuroFIR (2018). European Food Information Resource.

accessed September 18, 2016.

The RICHFIELDS Framework for Semantic Interoperability of Food Information Across Heterogenous Information Systems

321

Farmakiotou, D., Karkaletsis, V., Koutsias, J., Sigletos, G.,

Spyropoulos, C. D., and Stamatopoulos, P. (2000).

Rule-based named entity recognition for greek ﬁnan-

cial texts. In Proceedings of the Workshop on Com-

putational lexicography and Multimedia Dictionaries

(COMLEX 2000), pages 75–78. Citeseer.

FigShare (2018). Simplifying your research workﬂow.

accessed June 12, 2018.

GDSN, G. (2018). The Global Data Synchronisation Net-

work. accessed June 12, 2018.

Gkoutos, G. V., Schoﬁeld, P. N., and Hoehndorf, R. (2012).

The units ontology: a tool for integrating units of me-

asurement in science. Database, 2012:bas033.

Gomaa, W. H. and Fahmy, A. A. (2013). A survey of text

similarity approaches. International Journal of Com-

puter Applications, 68(13).

Habibi, M., Weber, L., Neves, M., Wiegandt, D. L., and Le-

ser, U. (2017). Deep learning with word embeddings

improves biomedical named entity recognition. Bioin-

formatics, 33(14):i37–i48.

Hanisch, D., Fundel, K., Mevissen, H.-T., Zimmer, R.,

and Fluck, J. (2005). Prominer: rule-based protein

and gene entity recognition. BMC bioinformatics,

6(1):S14.

Huang, Z., Xu, W., and Yu, K. (2015). Bidirectional

lstm-crf models for sequence tagging. arXiv preprint

arXiv:1508.01991.

Jensen, K., Panagiotou, G., and Kouskoumvekaki, I. (2014).

Integrated text mining and chemoinformatics analy-

sis associates diet to health beneﬁt at molecular level.

PLoS computational biology, 10(1):e1003432.

Lample, G., Ballesteros, M., Subramanian, S., Kawa-

kami, K., and Dyer, C. (2016). Neural architec-

tures for named entity recognition. arXiv preprint

arXiv:1603.01360.

Leaman, R., Wei, C.-H., Zou, C., and Lu, Z. (2015). Mining

patents with tmchem, gnormplus and an ensemble of

open systems. In Proce. The ﬁfth BioCreative chal-

lenge evaluation workshop, pages 140–146.

Lopez, M. M. and Kalita, J. (2017). Deep learning applied

to nlp. arXiv preprint arXiv:1703.03091.

Lu, Z., Kao, H.-Y., Wei, C.-H., Huang, M., Liu, J., Kuo,

C.-J., Hsu, C.-N., Tsai, R. T.-H., Dai, H.-J., Okazaki,

N., et al. (2011). The gene normalization task in bio-

creative iii. BMC bioinformatics, 12(8):S2.

Metzler, D., Dumais, S., and Meek, C. (2007). Similarity

measures for short segments of text. In European Con-

ference on Information Retrieval, pages 16–27. Sprin-

ger.

Miller, R. A., Gieszczykiewicz, F. M., Vries, J. K., and

Cooper, G. F. (1992). Chartline: providing biblio-

graphic references relevant to patient charts using the

umls metathesaurus knowledge sources. In Procee-

dings of the Annual Symposium on Computer Appli-

cation in Medical Care, page 86. American Medical

Informatics Association.

Morgan, A. A., Lu, Z., Wang, X., Cohen, A. M., Fluck, J.,

Ruch, P., Divoli, A., Fundel, K., Leaman, R., Haken-

berg, J., et al. (2008). Overview of biocreative ii gene

normalization. Genome biology, 9(2):S3.

uller, H.-M., Kenny, E. E., and Sternberg, P. W. (2004).

Textpresso: an ontology-based information retrieval

and extraction system for biological literature. PLoS

biology, 2(11):e309.

Petasis, G., Vichot, F., Wolinski, F., Paliouras, G., Karkalet-

sis, V., and Spyropoulos, C. D. (2001). Using machine

learning to maintain rule-based named-entity recogni-

tion and classiﬁcation systems. In Proceedings of the

39th Annual Meeting on Association for Computatio-

nal Linguistics, pages 426–433. Association for Com-

putational Linguistics.

Piskorski, J. and Yangarber, R. (2013). Information ex-

traction: past, present and future. In Multi-source,

multilingual information extraction and summariza-

tion, pages 23–49. Springer.

PRECIOUS (2018). Preventive Care Infrastructure based

On Ubiquitous Sensing. accessed June 12, 2018.

QuaLiFY (2018). Information service for personalised nu-

trition and lifestyle advice. accessed June 12, 2018.

Ramanan, S., Broido, S., and Nathan, P. S. (2013). Perfor-

mance of a multi-class biomedical tagger on clinical

records. In CLEF (Working Notes).

Real, R. and J.M.Vargas (1996). The probabilistic basis of

jaccard’s index of similarity. Systematic biology, pa-

ges 380–385.

Rindﬂesch, T. C., Tanabe, L., Weinstein, J. N., and Hunter,

L. (2000). Edgar: extraction of drugs, genes and re-

lations from the biomedical literature. In Paciﬁc Sym-

posium on Biocomputing. Paciﬁc Symposium on Bio-

computing, page 517. NIH Public Access.

Rockt

aschel, T., Weidlich, M., and Leser, U. (2012). Chem-

spot: a hybrid system for chemical named entity re-

cognition. Bioinformatics, 28(12):1633–1640.

Santos, C. N. d. and Guimaraes, V. (2015). Boosting named

entity recognition with neural character embeddings.

arXiv preprint arXiv:1505.05008.

Settles, B. (2010). Active learning literature survey. Uni-

versity of Wisconsin, Madison, 52(55-66):11.

Xia, J., Zhang, X., Yuan, D., Chen, L., Webster, J., and

Fang, A. C. (2013). Gene prioritization of resis-

tant rice gene against xanthomas oryzae pv. oryzae by

using text mining technologies. BioMed research in-

ternational, 2013.

Zenodo (2018). Zenodo. accessed June 12, 2018.

Zhou, X., Zhang, X., and Hu, X. (2006). Maxmatcher: Bi-

ological concept extraction using approximate dictio-

nary lookup. In Paciﬁc Rim International Conference

on Artiﬁcial Intelligence, pages 1145–1149. Springer.

KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval

322