FoodOntoMap: Linking Food Concepts across Different Food Ontologies

Gorjan Popovski

1,2 a

, Barbara Korou

c Seljak

3 b

and Tome Eftimov

3,4,5 c

Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, 1000 Skopje, North Macedonia

zef Stefan International Postgraduate School, 1000 Ljubljana, Slovenia

Computer Systems Department, Jo

zef Stefan Institute, 1000 Ljubljana, Slovenia

Center for Population Health Sciences, Stanford University, 94305 California, U.S.A.

Department of Biomedical Data Science, Stanford University, 94305 California, U.S.A.

Keywords:

Food Data Normalization, Food Data Linking, Food Ontology, Food Semantics.

Abstract:

In the last decade, a great amount of work has been done in predictive modelling in healthcare. All this work is

made possible by the existence of several available biomedical vocabularies and standards, which play a crucial

role in understanding health information. Moreover, there are available systems, such as the Uniﬁed Medical

Language System, that bring and link together all these biomedical vocabularies to enable interoperability

between computer systems. However, in 2019, Lancet Planetary Health published that the year 2019 is going

to be the year of nutrition, where the focus will be on the links between food systems, human health, and the

environment. While there is a large number of available resources for the biomedical domain, only a limited

number of resources can be utilized in the food domain. There is still no annotated corpus with food concepts,

and there are only a few rule-based food named-entity recognition systems for food concepts extraction. There

are also several food ontologies that exist, each developed for a speciﬁc application scenario. However there

are no links between these ontologies. For this reason, we have created a FoodOntoMap resource that consists

of food concepts extracted from recipes. For each food concept, semantic tags from four food ontologies are

assigned. With this, we have created a resource that provides a link between different food ontologies that can

be further reused to develop applications for understanding the relation between food systems, human health,

and the environment.

1 INTRODUCTION AND

MOTIVATION

Nowadays, the use of predictive modeling in health-

care increases with the large amount of data that is

becoming available. One example of such data are

the electronic health records (EHRs) (Gligic et al.,

2019; Wang et al., 2019), which represent the largest

source of medical data. Analyzing them, the medi-

cal information is presented as natural language text

(i.e. unstructured data) and the key challenge is to ex-

tract terms that are different medical concepts (e.g.,

drugs, diseases, procedures, treatments, etc.). For

this reason, a lot of named-entity recognition meth-

ods (NERs) have been developed (Boag et al., 2015),

which are further used to extract this information for

https://orcid.org/0000-0001-9091-4735

https://orcid.org/0000-0001-7597-2590

each patient and then trying to ﬁnd a patient’s rep-

resentation for some predictive study. Besides the

unstructured data, there are also resources that con-

sists of structured patient medical information. One

such example is the MIMIC-III data (Johnson et al.,

2016), which consists of data relating to patients who

stayed within the intensive care units at Beth Israel

Deaconess Medical Center. The common thing about

the medical data, no matter from where it comes

(unstructured or structured data), is that it is further

used to ﬁnd a patient’s representation by projecting

the data into a continuous vector space (Beam et al.,

2018; Choi et al., 2016; Miotto et al., 2016). In

this way, medical embeddings are learned in order

to capture the non-linear relationships that exist be-

tween the medical concepts. These representations

are further used with some advanced machine learn-

ing or deep learning methods to perform predictive

studies in healthcare. However, all this happens as

a result of the availability of biomedical vocabularies

Popovski, G., Seljak, B. and Eftimov, T.

FoodOntoMap: Linking Food Concepts across Different Food Ontologies.

DOI: 10.5220/0008353201950202

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 195-202

ISBN: 978-989-758-382-7

195

and standards that can be used to normalize the med-

ical concepts before learning the embedding space.

One such example is the Uniﬁed Medical Language

System (UMLS) that brings and links together several

biomedical vocabularies to enable interoperability be-

tween computer systems (Bodenreider, 2004).

However, in 2019, the Lancet Planetary Health

published that the year 2019 is going to be the year

of nutrition, where the focus will be on the links be-

tween food systems, human health, and the environ-

ment. Contrary to the large number of available re-

sources for the biomedical domain, in the food do-

main there is a limited number of resources that can

be used. There is still no annotated corpus with food

concepts, and there are few rule-based food named-

entity recognition systems that can be used for food

concepts extraction (Eftimov et al., 2017b; Popovski

et al., 2019). Additionally, a number of food ontolo-

gies exist, each developed for a speciﬁc application

scenario, but there are no links between them (Bou-

los et al., 2015). In order to move ahead the work for

ﬁnding relationships between the human health, food

systems and environment, we should have resources

that will be used for food concepts normalization. For

this reason, we have created a FoodOntoMap resource

that consists of food concepts extracted from recipes

and for each one the semantic tags from four food on-

tologies are assigned. With this, we have created a

resource that provides a link between different food

ontologies that can be further reused to develop em-

bedding space for food concepts and applications for

understanding the relation between food system, hu-

man health, and the environment.

2 RELATED WORK

In this section we provide an overview of the avail-

able food semantic resources that can be used for food

data normalization, followed by the methodologies

that can be used for food information extraction and

normalization.

2.1 Food Semantic Resources

In the domain of food, several food ontologies already

exist, such as:

FoodWiki provides a model of different types of

foods, together with their nutritional information, and

the re-commended daily intake (C¸ elik, 2015).

AGROVOC is a large multilingual thesaurus,

whose terminology is widely used in practice for sub-

ject ﬁelds in agriculture,ﬁsheries, forestry, food and

related domains (Caracciolo et al., 2012).

Open Food Facts is an open source global food

database that allows users to learn about a food’s

nutritional information and compare products from

around the world (Boulos et al., 2015). It is also ben-

eﬁcial for the food industry, where it can be used to

track, monitor, and strategically plan food production.

Food Product Ontology describes food products

using common representation, vocabulary and lan-

guage for the food product domain. It is an extended

version of a widely used standardized ontology for

product, price, store, and company data (Kolchin and

Zamula, 2013).

FOODS (Diabetics Edition) is an ontology-driven

system that delivers a web-based food-menu recom-

mendation system for patients with diabetes in Thai-

land (Snae and Br

uckner, 2008).

FoodOn focuses on the human-centric categoriza-

tion and handling of food (Grifﬁths et al., 2016). Its

main goal is to develop semantics for food safety,

food security, agricultural and animal husbandry prac-

tices linked to food production, culinary, nutritional

and chemical ingredients and processes. It uses parts

from several ontologies covering anatomy, taxonomy,

geography and cultural heritage. Its usage is related

to research and clinical data sets in academia and gov-

ernment.

A detailed review of the aforementioned food on-

tologies was provided by Boulos et. al. (Boulos et al.,

2015).

OntoFood is an ontology with SWRL rules of nu-

trition for diabetic patients and is available in the Bio-

Portal.

SNOMED CT (Systematized Nomenclature of

Medicine - Clinical Terms) is a standardized, multi-

lingual vocabulary of clinical terminology that is used

by physicians and other health care providers for the

electronic health records (Donnelly, 2006). Beside

the medical concepts that are the main focus of this

ontology, there is also a Food concept that can be fur-

ther used for food concept normalization.

The Hansard corpus is a collection of text and

concepts created as a part of the SAMUELS project

(2014-2016). It consists of nearly every speech given

in the British Parliament from 1803-2005. The main

beneﬁt is that it allows semantically-based searches of

these speeches. More details about semantic tags can

be found in (Alexander and Anderson, 2012; Rayson

et al., 2004). The words are organized in 37 higher

level semantic groups, in which one of them is also

Food and Drink (i.e. AG).

Table 1 summarizes the availability of food se-

mantic resources.

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

196

Table 1: Availability of food semantic resources.

Resource name Availability

FoodWiki https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4496660/

AGROVOC http://agroportal.lirmm.fr/ontologies/AGROVOC

Open Food Facts https://vest.agrisemantics.org/content/open-food-facts-food-ontology

Food Product Ontology https://vest.agrisemantics.org/content/food-product-ontology

FoodOn https://foodontology.github.io/foodon/

OntoFood https://bioportal.bioontology.org/ontologies/OF/?p=summary

SNOMED CT https://conﬂuence.ihtsdotools.org/display/DOC/Technical+Resources

Hansard corpus https://www.hansard-corpus.org/

2.2 Food Named-entity Recognition

System

Contrary to the extensive work in named-entity recog-

nition methods for biomedical tasks, the situation in

the food and nutrition science is completely different.

The UCREL Semantic Analysis System (USAS)

is a framework for automatic semantic analysis of

text, which distinguishes between 21 major cate-

gories, one of which is “food and farming” (Rayson

et al., 2004). The USAS can provide additional infor-

mation about the food entity, but the limitation is that

it works on a token level. For example, if in the text

two words (i.e. tokens), like “grilled chicken”, denote

one food entity that needs to be extracted and ana-

lyzed, the semantic tagger would actually parse the

words “grilled” and “chicken” as separate entities and

obtain separate semantic tags.

In (Eftimov et al., 2017b), a rule-based NER used

for information extraction (IE) from evidence-based

dietary recommendation, called drNER, is presented,

where among other entities, food entities were also of

interest. This work was extended into a rule-based

named-entity recognition method for food informa-

tion extraction, called FoodIE (Popovski et al., 2019).

It is a rule engine, where the rules are based on com-

putational linguistics and semantic information that

describe the food entities. Evaluation showed that

FoodIE behaves consistently using different indepen-

dent evaluation data sets and very promising results

have been achieved.

The NCBO Annotator is a Web service that anno-

tates text provided by the user by using relevant ontol-

ogy concepts (Jonquet et al., 2009). It is available as

a part of the BioPortal software services (Noy et al.,

2009). The annotation workﬂow is based on a highly

efﬁcient syntactic concept recognition (using concept

names and synonyms) engine and on a set of seman-

tic expansion algorithms that leverage the semantics

in ontologies. The methodology leverages ontologies

to create annotations of raw text and returns them us-

ing semantic web standards.

2.3 Food Concepts Normalization

In the last few years, food concepts normalization is

an open research question that is highly researched by

the food and nutrition science community, calling it

food matching. For this reason, StandFood (Eftimov

et al., 2017a) is recently introduced, which a semi-

automatic system for classifying and describing foods

according to a description and classiﬁctaion system,

such as FoodEx2, proposed by the European Food

Safety Agency (EFSA) ((EFSA), 2015). It consists of

three parts. The ﬁrst involves a machine learning ap-

proach and classiﬁes foods into four categories, with

two for single foods: raw (r) and derivatives (d), and

two for composite foods: simple (s) and aggregated

(c). The second uses a natural language processing

approach and probability theory to perform food con-

cepts normalization. The third combines the result

from the ﬁrst and the second part by deﬁning post-

processing rules in order to improve the result for the

classiﬁcation part. However, the food normalization

process was based only on lexical similarity between

the food concepts names, avoiding the semantic simi-

larity between them.

3 FoodOntoMap DESIGN

To provide a data set in which food concepts are

normalized by semantic tags from different food on-

tologies, we selected 22,000 recipes from Allrecipes

(Groves, 2013), which is the largest food-focused so-

cial network where people share their recipes and pro-

vide information about recipes in a non-standardized

manner. The recipes were selected from ﬁve recipe

categories: Appetizers and snacks, Breakfast and

Lunch, Dessert, Dinner, and Drinks.

To extract and annotate food concepts using dif-

ferent food ontologies we used two approaches.

FoodOntoMap: Linking Food Concepts across Different Food Ontologies

197

First, we used the recently proposed rule-based food-

named entity recognition method called FoodIE. Us-

ing FoodIE, the extracted food concepts are addi-

tionally annotated using the Hansard corpus seman-

tic tags. Then, we also used the NCBO annota-

tor together with the food ontologies that are avail-

able in the BioPortal (i.e. FoodOn, OntoFood, and

SNOMED CT). The semantic tags that are assigned

to each food concept are the semantic tags that belong

to the tokens from which the concept is constructed,

so it is a multi-label annotation process. Further in

the paper, we are going to explain each step in more

detail.

After collecting the recipes, FoodIE is used to

extract the food concepts. Further, we addition-

ally annotated the food concepts by assigning the

semantic tags from the Hansard corpus. For ex-

ample, if the food concept is ”grilled chicken”, it

obtains the semantic tags AG.01.t.07[Cooking] and

AG.01.d.06[Fowls]. After assigning the semantic tags

to each food concept, the results were organized in the

BioC format, which is a simple format to share text

data and annotations, with the goals of simplicity, in-

teroperability, and broad use and reuse (Comeau et al.,

2013). The BioC format for one recipe and its anno-

tations are presented in Figure 1. From it, it is evident

that each recipe is presented as a document for which

the category, description (text), and food annotations

are included. Each annotation consists of the food

concept that is extracted, the semantic tags from the

Hansard corpus that are assigned to it, and the offset

that points the token position from the beginning of

the text where the food concept starts and its charac-

ter length. The FoodIE NER method is publicly avail-

able on GitHub (https://github.com/GorjanP/foodie).

The data set organized in the BioC format is also pub-

licly available at http://cs.ijs.si/repository/FoodBase/

foodbase.zip.

Furthermore, the same recipe data set was pro-

cessed using the NCBO annotator with the FoodOn,

OntoFood, and SNOMED CT ontology. For this

reason, we used an R client that uses the An-

notator API available at http://data.bioontology.org/

documentation. The annotator API was used with the

recipe text as input and ﬁlters provided by the ontol-

ogy id (i.e FoodOn, OntoFood and SNOMED CT).

When SNOMED CT was used, additionally another

ﬁlter was applied based on a semantic type for which

food was speciﬁed (i.e. Food (T168)). An example

of an annotation is presented in Figure 2. Using this

ﬁgure, it is evident that for each annotation we have

the semantic tag (i.e. id), the food concept text, and

two numbers indicating the starting and ending posi-

tion of the food concept mention, which are referred

as from and to. It is important to note that the BioC

recipe format uses different metrics to deﬁne the po-

sition of the mentioned food concept. The NCBO

annotator uses character based offsets (from and to),

while the BioC format uses two metrics referred to as

offest and length. The offset indicates at which token

from the beginning of the recipe text the mentioned

food concept occurs, while the length indicates the

character length of the food concept. Due to this dif-

ference, these numbers were converted between the

two types of location description while the mapping

was being performed. Using these location metrics it

is determined which food concept mentions extracted

by the NCBO annotator refer to the same food con-

cept. More speciﬁcally, if some food concept mention

is a subset of another mention, their information is ag-

gregated and represented as the superset food concept

mention. With this duplicates, synonyms and multiple

labels are resolved into a more complete food entity.

The FoodOntoMap data set consists of food con-

cepts extracted from the recipes and normalized to

different food ontologies. It also provides a link be-

tween the food ontologies. For this reason, the food

concepts are matched and for each food concept the

semantic information from each dataset is assigned.

The concept matching is done by iterating through

each food concept that is extracted by the NER

method FoodIE. If the concept is also recognized

wholly or partially by the NCBO annotator in com-

bination with any of the selected ontologies, the se-

mantic tags from that ontology are also assigned to

the food concept. However, it is not uncommon while

using the NCBO annotator for it to provide semantic

tags on a token level instead of on a concept level.

Such an example would be when an ontology returns

two outputs for the food concept “salad dressing”, in-

stead of classifying it as a single food concept con-

sisting of two tokens. In these cases, each incom-

plete food concept extracted by the NCBO needs to

be matched to its corresponding superset food con-

cept concept extracted by FoodIE. This was done by

checking if the location metrics from the NCBO an-

notator are in accordance (more speciﬁcally, whether

they are a subrange) with the location metrics of a

food concept provided by FoodIE. If such a match ex-

ists, the NCBO food concept is added to the mapping

set of the FoodIE concept. By doing this, the map-

ping sets aggregate all the corresponding food con-

cept mentions that map to a food concept, along with

their semantic information, even if the NCBO anno-

tator does not classify some food concepts wholly.

To illustrate this with an example, on Figure 1

it is evident that we have multiple food concept

mentions that overlap. Such mentions are “SALAD

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

198

Figure 1: BioC format for a single recipe.

Figure 2: NCBO tagger output for a single recipe, using the ontology FoodOn.

DRESSING”, starting at 19 and ending at 32 and

appearing twice with different semantic information;

“SALAD”, starting at 19 and ending at 23; and

“DRESSING”, starting at 25 and ending at 32. Dur-

ing the matching step that is mentioned above, all

these are aggregated into a single food concept men-

tion that holds all the semantic information of its com-

ponent food concepts. The sole food concept mention

becomes “SALAD DRESSING”, starting at 19 and

ending at 32 along with four different semantic tags,

one from each component mention. Then, after con-

verting the location metrics to be compatible with the

BioC recipe format, the food concept is matched to

its corresponding food concept from the BioC recipe

dataset. This means that the resulting mapping is “dry

ranch salad dressing mix” to “SALAD DRESSING”.

Notice that not all tokens are caught by the NCBO

annotator using FoodOn. With this, the normaliza-

tion is complete, the code mapping being A000046 to

B000027.

In instances where one code from a dataset maps

to two separate codes from another dataset, the NCBO

tagger failed to produce one food concept mention

which contains the full semantic information. Instead,

it produced only a token-level mentions, and as such

they were not aggregated into one larger food concept

mention. Such an example is “SALTED WATER”,

which maps to “WATER” and “SALTED” separately,

the codes being A000123, B000012 and B000065, re-

spectively.

The ﬂowchart of the FoodOntoMap design is pre-

sented in Figure 3.

The availability of the resources that are used to

create the FoodOntoMap is presented in Table 2.

The results from the FoodOntoMap are four dif-

ferent datasets and one mapping. Each dataset con-

sists of an artiﬁcial id for each unique food concept

that is extracted using each approach, the name of the

FoodOntoMap: Linking Food Concepts across Different Food Ontologies

199

Recipes

FoodIE NCBO annotator

FoodON

OntoFood

SNOMEDCT

Food concepts

annotated using the

Hansard corpus

Food concepts

annotated using the

selected

ontologies

FoodOntoMap

Figure 3: FoodOntoMap design.

Table 2: Availability of resources used to create FoodOntoMap.

Resource name Availability

FoodIE https://github.com/GorjanP/foodie

BioC format recipe set http://cs.ijs.si/repository/FoodBase/foodbase.zip

Mapper client and NCBO annotator client https://github.com/GorjanP/FOM mapper client

Hansard corpus https://www.hansard-corpus.org/

NCBO annotator http://bioportal.bioontology.org/annotator

NCBO annotator REST API http://data.bioontology.org/documentation

FoodOn https://foodontology.github.io/foodon/

OntoFood https://bioportal.bioontology.org/ontologies/OF/?p=summary

SNOMED CT https://conﬂuence.ihtsdotools.org/display/DOC/Technical+Resources

extracted food concept, and the semantic information

assigned to it. Each dataset corresponds to one of the

four semantic resources: Hansard corpus, FoodOn,

OntoFood, and SNOMED CT. At the end there is one

data set mapping, called FoodOntoMap, which, for

each concept that appears at least in two datasets, pro-

vides the links between them by listing the artiﬁcial id

of the concepts from each of the datasets in which it

is mentioned.

An example for one instance from FoodOntoMap

is A000016, B000011, C000012, D000002. This

means that this food concept is mapped from the

Hansard corpus to the respective ontologies, i.e.

FoodOn, SNOMED CT and OntoFood. The provided

codes are artiﬁcial unique identiﬁers that are assigned

to the food concept using each of the aforementioned

food semantic resources, respectively. If we look at

the separate datasets, we can see that A000016 cor-

responds to “garlic” with semantic tag AG.01.h.02.e

[Onion/leek/garlic] from the Hansard corpus,

B000011 corresponds to “garlic” with semantic

tag http://purl.obolibrary.org/obo/NCBITaxon 4682

from the FoodOn, C000012 is for “garlic” with the

semantic tag http://purl.bioontology.org/ontology/

SNOMEDCT/735030001 from the SNOMED

CT, and D000002 corresponds to enquotegarlic

with the semantic tag http://www.owl-ontologies.

com/Ontology1435740495.owl#Garlic from the

OntoFood.

The datasets consist of 13,205; 1,069; 111; and

582 unique food concepts, obtained using Hansard

corpus, FoodOn, OntoFood, and SNOMED CT, re-

spectively. The FoodOntoMap data set consists of

1,459 food concepts that are found in at least two food

semantic resources.

From the results, we can conclude that FoodIE

with the Hansard corpus gives the most promising re-

sults because it can extracted larger number of food

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

200

concepts compared with the NCBO annotator in com-

bination with some of the selected ontologies. More-

over, the results prove that the three food ontologies

(i.e. SNOMED CT, FoodOn, OntoFood) do not rep-

resent the food domain well, as many food concepts

cannot be extracted using the NCBO annotator, which

indicates that they do not exist in the food ontologies

themselves.

4 DISCUSSION

4.1 Impact

To the best of our knowledge, FoodOntoMap is the

ﬁrst resource that provides normalization of food con-

cepts to different food ontologies, additionally pro-

viding a link between them. The motivation for build-

ing such a resource in the food domain comes from

the existence of the UMLS, which is extensively used

in the biomedical domain. For example, the MR-

CONSO.RRF table that is a part of the UMLS is

used in a lot of semantic web applications because

it can map the medical concepts to different biomed-

ical standards and vocabularies. To make progress in

analysing the large amount of data that is available in

order to ﬁnd these relations, resources for food con-

cepts normalization are extremely valuable and wel-

come.

The main beneﬁt of using FoodOntoMap is that

the food concepts can be normalzied by mapping

them to a uniﬁed system. Furthermore, the seman-

tic tags can be reused to ﬁnd the non-linear relations

that exist between the concepts in the vector space,

by learning the embedding space. This can also be

done together with some medical and environment

concepts. Once the embedding space is learned, the

embeddings can be used for predictive studies in order

to explain the relations between human health, food

systems, and the environment.

4.2 Reusability

FoodOntoMap can be used as a resource that repre-

sents a normalized dataset of food concepts. Addi-

tionally, users can also follow the described pipeline

of steps used to create the FoodOntoMap mapping

in order to create their own new resource where the

food concepts will be normalized. Both approaches,

FoodIE and NCBO, used for food concepts extrac-

tion have already been well documented and evalu-

ated. The ontologies that are used by the NCBO an-

notator have also been well documented and are easy

to utilize. Furthermore, FoodOntoMap can be easily

extended on additional recipes, as well as a wide vari-

ety of different ontologies. With this it can provide an

ever wider coverage of food concepts. Additionally,

the FoodOntoMap pipeline can be used to normalize

food concepts that exist in food consumption and food

composition databases.

4.3 Availability

The resource FoodOntoMap is published and pub-

licly available for download at https://doi.org/10.

5281/zenodo.2635437. under a Creative Commons

Attribution-NonCommercial-ShareAlike 4.0 Interna-

tional (CC BY-NC-SA 4.0) licence. With this, we

encourage users to further contribute to this resource

and modify it as need be. Zenodo was the platform of

choice, as it provided all the framework tools needed

for such a dataset. Hosted along with the resource’s

datasets is a DCAT speciﬁcation ﬁle, which brieﬂy de-

scribes the structure of the resource. Furthermore, the

resource will be actively maintained and extended as

new ontologies, annotators, NER methods and NLP

methods become available. The goal is to keep the re-

source relevant and contemporary while also improv-

ing its domain coverage with the ever improving NLP

tools.

5 CONCLUSION

The resource that is presented in this paper represents

a data normalization pipeline. FoodOntoMap targets

the domain of food and nutrition science, normaliz-

ing food concepts across three different ontologies

and one semantic dataset. Additionally, it provides

a set with semantic information for each food concept

present in the dataset. This was made possible with

the use of NER methods for information extraction

from unstructured textual data. FoodOntoMap rep-

resents a ﬁrst of its kind in the domain of Food and

Nutrition, with no other such resources being readily

available to the best of our knowledge. As such, it

is crucial in building tools that improve knowledge in

this domains and the public health effects it implies,

as well as building data models that can be utilized

to further discover relations and links in the food and

nutrition domain.

FoodOntoMap: Linking Food Concepts across Different Food Ontologies

201

ACKNOWLEDGEMENT

This work was supported by the Slovenian Research

Agency Program P2-0098 and the European Union’s

Horizon 2020 research and innovation programme

[grant agreement No 769661].

REFERENCES

Alexander, M. and Anderson, J. (2012). The hansard cor-

pus, 1803-2003.

Beam, A. L., Kompa, B., Fried, I., Palmer, N. P., Shi, X.,

Cai, T., and Kohane, I. S. (2018). Clinical concept

embeddings learned from massive sources of medical

data. arXiv preprint arXiv:1804.01486.

Boag, W., Wacome, K., Naumann, T., and Rumshisky, A.

(2015). Cliner: A lightweight tool for clinical named

entity recognition. AMIA Joint Summits on Clinical

Research Informatics (poster).

Bodenreider, O. (2004). The uniﬁed medical language sys-

tem (umls): integrating biomedical terminology. Nu-

cleic acids research, 32(suppl 1):D267–D270.

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

hoot, 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.

Caracciolo, C., Stellato, A., Rajbahndari, S., Morshed,

A., Johannsen, G., Jaques, Y., and Keizer, J. (2012).

Thesaurus maintenance, alignment and publication as

linked data: the agrovoc use case. International Jour-

nal of Metadata, Semantics and Ontologies, 7(1):65–

75.

C¸ elik, D. (2015). Foodwiki: Ontology-driven mobile safe

food consumption system. The Scientiﬁc World Jour-

nal, 2015.

Choi, E., Bahadori, M. T., Searles, E., Coffey, C., Thomp-

son, M., Bost, J., Tejedor-Sojo, J., and Sun, J. (2016).

Multi-layer representation learning for medical con-

cepts. In Proceedings of the 22nd ACM SIGKDD In-

ternational Conference on Knowledge Discovery and

Data Mining, pages 1495–1504. ACM.

Comeau, D. C., Do

gan, R. I., Ciccarese, P., Cohen, K. B.,

Krallinger, M., Leitner, F., Lu, Z., Peng, Y., Rinaldi,

F., Torii, M., et al. (2013). Bioc: a minimalist ap-

proach to interoperability for biomedical text process-

ing. Database, 2013:bat064.

Donnelly, K. (2006). Snomed-ct: The advanced terminol-

ogy and coding system for ehealth. Studies in health

technology and informatics, 121:279.

(EFSA), E. F. S. A. (2015). The food classiﬁcation and

description system foodex 2 (revision 2). EFSA Sup-

porting Publications, 12(5):804E.

Eftimov, T., Koro

sec, P., and Korou

c Seljak, B. (2017a).

Standfood: Standardization of foods using a semi-

automatic system for classifying and describing foods

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

Eftimov, T., Korou

c Seljak, B., and Koro

sec, P. (2017b).

A rule-based named-entity recognition method for

knowledge extraction of evidence-based dietary rec-

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

Gligic, L., Kormilitzin, A., Goldberg, P., and Nevado-

Holgado, A. (2019). Named entity recognition

in electronic health records using transfer learn-

ing bootstrapped neural networks. arXiv preprint

arXiv:1901.01592.

Grifﬁths, E. J., Dooley, D. M., Buttigieg, P. L., Hoehn-

dorf, R., Brinkman, F. S., and Hsiao, W. W. (2016).

Foodon: A global farm-to-fork food ontology. In

ICBO/BioCreative.

Groves, S. (2013). How allrecipes. com became the worlds

largest food/recipe site. roi of social media (blog).

Johnson, A. E., Pollard, T. J., Shen, L., Li-wei, H. L.,

Feng, M., Ghassemi, M., Moody, B., Szolovits, P.,

Celi, L. A., and Mark, R. G. (2016). Mimic-iii, a

freely accessible critical care database. Scientiﬁc data,

3:160035.

Jonquet, C., Shah, N., Youn, C., Callendar, C., Storey, M.-

A., and Musen, M. (2009). Ncbo annotator: seman-

tic annotation of biomedical data. In International

Semantic Web Conference, Poster and Demo session,

volume 110.

Kolchin, M. and Zamula, D. (2013). Food product ontol-

ogy: Initial implementation of a vocabulary for de-

scribing food products. In Proceeding of the 14th

Conference of Open Innovations Association FRUCT,

Helsinki, Finland, pages 11–15.

Miotto, R., Li, L., Kidd, B. A., and Dudley, J. T. (2016).

Deep patient: an unsupervised representation to pre-

dict the future of patients from the electronic health

records. Scientiﬁc reports, 6:26094.

Noy, N. F., Shah, N. H., Whetzel, P. L., Dai, B., Dorf, M.,

Grifﬁth, N., Jonquet, C., Rubin, D. L., Storey, M.-A.,

Chute, C. G., et al. (2009). Bioportal: ontologies and

integrated data resources at the click of a mouse. Nu-

cleic acids research, 37(suppl 2):W170–W173.

Popovski, G., Kochev, S., Korou

c Seljak, B., and Efti-

mov, T. (2019). Foodie: A rule-based named-entity

recognition method for food information extraction.

In Proceedings of the 8th International Conference

on Pattern Recognition Applications and Methods,

(ICPRAM 2019), pages 915–922.

Rayson, P., Archer, D., Piao, S., and McEnery, A. (2004).

The ucrel semantic analysis system.

Snae, C. and Br

uckner, M. (2008). Foods: a food-oriented

ontology-driven system. In Digital Ecosystems and

Technologies, 2008. DEST 2008. 2nd IEEE Interna-

tional Conference on, pages 168–176. IEEE.

Wang, Q., Zhou, Y., Ruan, T., Gao, D., Xia, Y., and He,

P. (2019). Incorporating dictionaries into deep neu-

ral networks for the chinese clinical named entity

recognition. Journal of biomedical informatics, page

103133.

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

202