Adding Temporal Dimension to Ontology Learning Models for

Depression Signs Detection from Social Media Texts

Patricia Martin-Rodilla

Information Retrieval Lab (IRLab), Facultade de Informática, University of A Coruña, Spain

Keywords: Ontology Learning, Time, Ontology Evolution, Text Mining, Social Media, Depression, Early Risk Prediction.

Abstract: Approaches to early detection of depression based on individual’s language are receiving increasing attention,

with detection software systems based on lexical, grammatical or discursive components applied to medical

corpus or social media texts. However, these first detection systems are defragmented, each attending to a

specific feature or linguistic level, and not addressing a more conceptual level. Existing ontology learning

(OL) methods extract the ontology referred in the text. In addition, existing systems perform language analysis

for the detection of depression as a snapshot of each individual, regardless of their temporal dimension. Is it

possible that suitable linguistic features to detect early signs of depression vary over time? And the underlying

ontology? This paper presents a model that adds the temporal component to current ontology learning models

to perform evolutionary analysis of both linguistic and ontological features to texts from social networks. The

model has been applied to an external corpus of depression from social media texts, with a two-fold goal: 1)

validating the model by contrasting it with OL models without temporal component 2) producing a corpus of

evolutionary OL results applied to the depression detection from social media texts.

1 INTRODUCTION

Depression is considered one of the most common

mental illness and one of the main diseases in recent

decades. Due to its complex diagnosis and variability

of presentations in different people, interdisciplinary

research approaches have become the most successful

methods for early risk prediction (Losada et al.,

2018a) and depression detection, combining

information about patient’s profile, doctor’s

experience, well-tested medical questionnaires and

semiautomatic analysis for assisting doctors to

analyze indicators. Within this last category, patient

language is a powerful indicator of personality traits

and emotions, and provides valuable clues about

mental health, presenting good results as auxiliary

indicators for helping doctors in an early risk

prediction of depression (Al-Mosaiwi & Johnstone,

2018, 2019; Pennebaker, Mehl, & Niederhoffer,

2003).

This connection between depression and

distinctive linguistic patterns have serve to create

promising software systems for assisting depression

early prediction. These systems have been tested

http://orcid.org/0000-0002-1540-883X

using textual corpus from clinical sessions or from

social networks, and classified them (Brewster, 2006;

Hazman, El-Beltagy, & Rafea, 2011; Wimalasuriya

& Dejing, 2010) regarding the linguistic parameters

analyzed for each system. Most of them are focused

on lexical or grammatical features (propositions or

pronouns) or ontological, discursive or topic-based

features. The ontological level, relating the

underlined ontology with the language patterns used

by the patient, offers important information about the

universe of discourse of the patient (Gruber, 1995), as

well as how the patient refers to its discourse

universe.

However, as far as we know, the temporal

component of this language pattern analysis,

especially at ontological or discursive levels, is not

considering as a feature in the current software

systems. This means that the information

corresponding to changes in patient language patterns

is not tracked or analyzed. The historical information

of the patients is one of the most valuable inputs for

doctors in the detection, diagnosis and evaluation of

mental illnesses. In addition, temporal dimension is

crucial in order to reveal early signs of psychological

Martin-Rodilla, P.

Adding Temporal Dimension to Ontology Learning Models for Depression Signs Detection from Social Media Texts.

DOI: 10.5220/0009351903230330

In Proceedings of the 15th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2020), pages 323-330

ISBN: 978-989-758-421-3

323

disorders trying to detect some indicators of

appearance of initial signs of depression and

understanding the evolution of an individual from the

early stages (e.g. mood changes, lack of sleep) to

severe stages (e.g. suicidal thoughts).

Focusing only in the ontological component of the

language patterns, this paper proposes a model that

adds the temporal component to current ontology

learning models, allowing us to perform evolutionary

analysis of both linguistic and ontological features to

texts from social networks. The model has been

applied to an external corpus of depression from

social media texts. The application shows how we can

add the temporal dimension to existing ontology

learning models in a real case, as well as produces a

valuable corpus of ontological and linguistic pattern

results over time in depression contexts.

2 BACKGROUND

Two main areas are related with our proposal: 1)

ontology learning methods from English unstructured

text from social networks, and 2) existing works

specifically focused on depression detection

software, contextualizing the application of our

proposal to this field.

2.1 Ontology Learning

Ontology Learning is defined as the discovering of

the underlined ontology from textual sources

(Hazman et al., 2011). As an ontology, we understand

here “an explicit, formal specification of a shared

conceptualization of a domain of interest” (Gruber,

1995). Thus, the underlined ontology of a given text

allow us to extract information about a) concepts and

relations referred in the source texts and b) linguistic

patterns used for referring to these concepts and

relations. This information conforms a relevant input

in the language studies, including applications of

ontology learning in biomedical or legal domain

(Morales, Scherer, & Levitan, 2017).

Firstly, we can find in literature initial approaches

trying to extract in a semiautomatic or automatic way

some ontological information from linguistic

patterns, such as processes relations or event mining

(Reuter & Cimiano, 2012). Regarding these studies,

most of them present high scores on recognition in a

limited functional environment or limited to a specific

domain or tasks.

Secondly, there are existing attempts for enriching

ontology learning with text mining techniques from

2000, e.g. some workshops in ECAI conference

(Staab, Maedche, Nedellec, & Wiemer-Hastings,

2000), to present. Main concerns here includes topical

concepts and concept definitions agreement within

the corresponding community, learning associations

from texts, Named Entity and Terminology

extraction, Acquisition of selected restrictions from

texts, Word Sense disambiguation or computation of

concept lattices from texts. We can also classify all

these text mining works in function of the kind of

technological technique employed: supervised (based

on previous annotations) vs. unsupervised.

Wimalasuriya survey (Wimalasuriya & Dejing, 2010)

presents the most common software architecture for

this kind of techniques, as well as some examples of

classical applications domains. In addition, some

authors (Asim, Wasim, Khan, Mahmood, & Abbasi,

2018; Brewster, 2006; Hazman et al., 2011;

Shamsfard & Barforoush, 2003; Somodevilla, Ayala,

& Pineda, 2018; Wimalasuriya & Dejing, 2010)

recently perform exhaustive reviews of the current

software methods for ontology learning. All these

methods have been successfully applied to a wide

variety of domains, which makes ontology learning a

solid area to consider when we want to extract

complex information (linguistically and conceptually

based) from unstructured sources.

Regarding ontology learning in social media

contexts, most of the approaches focused on

extracting parts of the ontology (Asim et al., 2018;

Breslin, 2012; Reuter & Cimiano, 2012), such as

concepts or events, particularizing approaches for

texts with shorter length and interaction

characteristics similar to dialogue (posts-based

interactions).

In summary, ontology learning is a promising area

with successful applications both at the level of

manual analysis and semi-automation. However,

none of the current methods recently reviewed (Asim

et al., 2018; Wimalasuriya & Dejing, 2010), even in

social media contexts, have specific temporality

support. This means that current methods extract the

underlying ontology as a snapshot of the text at a

specific time. In reported applications, this snapshot

condenses enough information. However, it does not

allow the study of the evolution in the underlying

ontology of the text or its linguistic patterns over

time. Because of our needs in the domain of mental

illness, we think that the inclusion of a temporary

layer to the ontology learning methods will facilitate

this evolutionary analysis and allow us a better

investigation of the connection and evolution of

linguistic and ontological patterns in depression

contexts.

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

324

2.2 Depression Signals and Early Risk

Prediction from Social Media

Language and depression studies are recently gaining

importance in research (Al-Mosaiwi & Johnstone,

2018; Kiss & Vicsi, 2017; Kokanovic et al., 2013;

Morales et al., 2017). Until recently, it was very

difficult to obtain reliable data on depression from

any source (from medical reports, due to their

classified character; from social networks, due to

their confidentiality, reliability and true diagnosis

problems). From 2016, The Early Risk Prediction on

the Internet (eRisk) (D. Losada, Crestani, & Parapar,

2017; David E. Losada et al., 2018) workshop

explores the interaction between language and mental

disorders in online social media. In particular, the

workshop proposed to address the early detection of

depression in an automatic way and released a corpus

of social media users who suffered from depression.

The results of the workshop showed that there is a

large spectrum of techniques that can be used to

detect this mental illness. In this paper, we use the

eRisk corpus to validate our model.

Specific works on depression detection from

social media texts are a field relatively new, with

promising approaches. Most of them works with

depression lexicons from Twitter (De Choudhury,

Gamon, Counts, & Horvitz, 2013), or for Reddit and

micro-blogs platforms (Coppersmith, Dredze, &

Harman, 2014; David E Losada & Gamallo, 2018). It

is also necessary to highlight that the Computational

Linguistics and Clinical Psychology Workshop

(clpsych.org) has recently organized "shared tasks" of

depression and post-traumatic disorders, performing

content analysis in support forums for people with

disorders (Resnik, Resnik, & Mitchell, 2014). These

tasks were oriented to automatic classification (do not

focus on early prediction or temporal analysis) but,

still, they will be valuable references for our research.

In addition, there are a few initiatives related to them,

such as personality and health mining competitions as

CLEF eHealth (http://sites.google.com/view/clef-

ehealth-2018/home) or PAN (http://pan.webis.de).

Al these works indicate the possibilities of

depression prediction from social media. However,

ontology learning applications with temporal

component to this field are still barely unexplored.

The method proposed here tries to solve some of these

shortcomings, contributing with ontological

information and their temporal evolutionary analysis

to all this set of technologies, methods and

approaches.

3 PROPOSAL: A

TEMPORAL-BASED

ONTOLOGY LEARNING

METHOD

As we previously detailed, ontological information

extracted from texts can be a very valuable input in

the early detection of signs of depression from social

networks. However, the application of any of the

current ontology learning methods lacks the

necessary temporal approach that allows an

evolutionary analysis.

Employing the current methods of ontology

learning it is possible to make two approaches: 1)

treat the entire corpus as a large text, so we would

obtain an aggregated underlined ontology (result of

evaluating all the texts at once) or 2) treat each corpus

document as a separate text, so we would obtain an

underlined ontology for each of the analysed texts.

Note here two important aspects of both approaches:

neither deals with a temporary component, so, in the

first case, we only have one ontology at a given time

(without evolutionary analysis), and in the second

case, we have several ontologies but no connection

between them, so we cannot perform evolutionary

analysis either. This absence of a temporary

component also severely penalizes applications to

short and disconnected texts such as those derived

from social networks and, as we have explained

before, temporal analysis is a crucial analysis in

mental illness application domains.

In order to fill this gap, we present a proposal to

add a temporary layer to existing ontology learning

methods. Using ontology learning principles as a

basis, we developed the proposal from scratch but in

highly modular schema, as the pipeline shows in

Figure 1.

The initial input is a corpus or collection of free-

style textual documents. Following the ontology

methods existing methods, we decide to apply the

second approach studied above for the discovery of

ontological information: extract an ontology for each

existing text in the corpus or collection. Thus, each

corpus analysed produced a set of underlined

ontologies (not only one) that allow us to perform

evolutionary analysis of the information. Phase I of

the pipeline runs the ontology learning (OL) model

chosen for extracting candidates for concepts and

their relationships for each existing text. The result of

Phase I is a ranked set of candidates for concepts and

relationships for each document. Then, the pipeline

consults the temporal information available in the

collection. Phase II is responsible for resolving, stor-

Adding Temporal Dimension to Ontology Learning Models for Depression Signs Detection from Social Media Texts

325

Figure 1: BPMN model (OMG, 2011) of the proposal pipeline, adding temporal dimension to OL methods.

ing and keeping the temporary information of the

ontological extraction process updated. In the phase

II the pipeline evaluates the given corpus regarding

temporal information available:

If the corpus has temporary information about the

production of texts (that is, when a due text is written

and / or added to the corpus), the pipeline creates the

necessary entities to maintain that information in each

of the candidates for concepts and relationships, as a

label of them (PATH I).

If the corpus does not have this information, the

pipeline creates the necessary entities to maintain

temporary information in real time of the analysis

performed, that is, it will assign a timestamp to each

of the candidates for concepts and relationships, as a

label of them (PATH II). This will allow us to

perform ontology learning evolutionary analysis also

in collections of texts or corpus without prior

temporal information. Thus, we can order our texts

and perform an evolutionary analysis as well, but with

temporal labels relative to the analysis itself and not

to the date of production of the texts.

Each of the paths of the pipeline produces the

ontological information with the temporal layer

completed, one result by each analysed text presented

in the original corpus/collection.

As a result, we can maintain the temporal

traceability of the production of the texts in the corpus

analysed in the case of the PATH I, performing

evolutionary analysis of the ontological information

and the linguistic patterns contained on it. For

example, we could now know what concepts appear

and disappear from the ontology over time or in a

range of dates, or if a linguistic pattern (the use of a

term for a concept, or the use of a particular verb for

a relationship between concepts) is maintained or

varies over time. In our specific case in depression

detection, this is especially useful given the studies

that relate this type of linguistic-ontological patterns

and specific disease states, allowing its analysis as a

differential factor in much greater depth. In the case

of PATH II, in corpus that do not have temporary

information on the production of the texts, the module

for ordering the documents by the user is necessary to

perform a relative evolutionary analysis.

This prior arrangement serves as a reference

timestamp for the evolutionary analysis (the first

document will be analysed before and its associated

ontology and information will carry a timestamp prior

to the last document) so that an analysis similar to that

of PATH I can be performed but only relative in the

corpus itself. Thus, we can also analyse what

concepts appear and disappear from the ontology over

time or if a linguistic pattern (the use of a term for a

concept, or the use of a particular verb for a

relationship between concepts) is maintained or

varies, but we cannot place it in its real time of

production. Next paragraphs detail the proposed

definitions and calculus for the PHASE II paths.

Taking C

as the input collection of n documents

written in free style that we want to analyse, we

formalized for PATH I the following results:

Vector of ontology learning candidates extracted:

for one document d, p pieces of text (posts in social

media) and k ontological entities extracted:

= 







(1)

CValue

d =









(2)

CValue

, Vector of ontology learning score values (c-

value): for one document d and k ontological entities

extracted and C-value score calculated for these k

entities.









, Set of ranked aggregated vectors of

candidates for the C

collection, with k ontological

entities extracted and their C-values scores. For each

d document in C

, PATH I searches, temporal

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

326

information available in the corpus (although this

aspect of the pipeline does not correspond to the

scope of the paper, it possible to see a possible

temporal information search implementations here

(Rust, 2018).



























(3)

For each d document in C

with temporal information,

PATH I calculates:

T (O

) = 











tprod

(4a)

T (O

): Set of resultant values of assigning to each

ontological entity (extracted from the n document)

their corresponding timestamp referred to the input

corpus information t

prod

Regarding PATH II, we define D

u,d

as the set of d

documents of the collection ordered by the user.

Then, using this order as a temporal reference, PATH

II calculates V

, CValue

and









applying same

PATH I formulae. Then, for each d document in C

PATH II stamps current temporal information of the

system. For each d document in C

with temporal

information, PATH I calculates:

T’ (O

) = 











tcurrent

(4b)

T’ (O

): Set of resultant values of assigning to each

ontological entity (extracted from the n document)

their corresponding current relative timestamp t

current

Note that the proposed pipeline has been defined

independently of the ontology learning method

selected, the domain of application or the corpus used

as a source. Thus, our proposal could serve as a

general solution for the need of a temporal layer in

current ontology learning methods. In the next

section, we particularize the application by selecting

all these aspects to illustrate the entire pipeline

proposal.

4 VALIDATION: APPLICATION

TO A SOCIAL MEDIA

DEPRESSION CORPUS

Within the scope of our proposal, it was necessary to

apply and evaluate the proposed pipeline with respect

to its ability to resolve, store and keep updated the

temporary information in a real corpus. Performing a

specific application of the pipeline requires defining

a) the ontology learning method chosen b) the

application domain and c) the corpus or collection

used as a source (the pipeline input).

Regarding the ontology method employed, the

review detailed in section 2 offer us a relevant pool of

methods and their current reliability, strengths and

weaknesses. Very few methods reviewed above

presents open-source implementations that allow us

to use them as initial method for our pipeline

implementation. From them, the C-value (Frantzi,

Ananiadou, & Mima, 2000) method presents better

behaviour in ontology learning, especially in concepts

extraction (Asim et al., 2018). C-value domain-

independent method is a well-known ontology

learning method that calculates a C-value score for

each text analysed, in function of linguistic and

statistical parameters, and giving as a result a vector

of ontological entities candidates ranked by the C-

value score. It presents especially good behaviour in

the semi-automatic extraction of multi-word and

nested ontological concepts from English corpora

(Asim et al., 2018). In addition, C-value have been

previously used in the medical domain with good

results (Lossio-Ventura, Jonquet, Roche, & Teisseire,

2013). C-value have been tested and previously

implemented in several platforms. We have used an

updated open source implementation of the algorithm

for English language (Conde, 2018) as a base for

adding the temporal dimension and implementing our

pipeline. All these reasons made us choose C-value as

a starting learning ontology method to illustrate our

pipeline and apply it to a real corpus.

The application domain (mental illness) and the

corpus chosen is due to our interest on depression

detection based on language patterns through

software systems. The original eRisk corpus (D.

Losada & F. Crestani, 2016; D. E. Losada & F.

Crestani, 2016) contains 2-year textual interactions

on Reddit from 892 users, divided into two groups:

137 subjects have explicitly declared that they have

been diagnosed with depression by medical

professionals, and the remaining 755 subjects are a

control group. All details regarding data acquisition,

initial annotations, depression diagnosis criteria, data

cleaning and legal treatment etc. are reported by the

original authors (D. E. Losada & F. Crestani, 2016).

We have applied the proposed pipeline to the

depression corpus:

1. Each eRisk document presents Reddit posts from

a subject (that it could be in the depression group

or in the control group) in JSON format. In Phase

I, the pipeline implements C-value algorithm,

obtaining a ranked vector of candidates to

Adding Temporal Dimension to Ontology Learning Models for Depression Signs Detection from Social Media Texts

327

ontological entities with their corresponding

ontological C-value score.

2. Then, the pipeline checks again the eRisk corpus,

searching for temporal information (Rust, 2018).

Due to eRisk corpus contains temporal

information of each post of the Reddit production

timestamp, the pipeline decides executing PATH

I algorithm (see

Figure 1).

3. In PATH I, the pipeline storages the timestamp of

each Reddit post for each JSON document

(corresponding to each subject in the collection).

This information is aligned to their corresponding

vector of ontological entities candidates for each

post. As a result, the pipeline obtains a vector of

ontological concepts aligned with their timestamp

of production. This allow us to perform

evolutionary analysis on the ontology, answering

questions regarding the ontological evolution over

the 2 years of data for each subject.

The execution of the pipeline would be similar in case

of the pipeline would execute PATH II but ordering

the corpus documents (See Figure 1). The proposed

method in form of a pipeline have allowed us to track

and maintain temporal information for each

ontological vector of the C-value algorithm output

vector. Figure 2 shows a screenshot of the final output

of the pipeline (PATH I) for our depression corpus

application. We selected a few p posts of the one

subject, showing the original texts and their

corresponding ontological entities extracted,

including temporal information of the production of

the texts.

5 DISCUSSION

The proposed method represents the first known

approach to the combination of ontology learning

methods with temporal analysis to allow ontological

evolutionary analysis. Due to the innovative nature of

the proposal, it is not possible to evaluate the results

obtained with existing benchmarks, beyond the good

results already reported (Asim et al., 2018) of the

ontology learning methods from unstructured

sources.

Said that, evaluating at a qualitative level, the

results obtained add value compared to traditional

methods of ontology learning: we will have, in

addition to the underlying ontology, evolutionary

information, as seen in our case of depression corpus

application. In the case of relative evolutionary

analysis (PATH II), the pipeline presents an

alternative solution to have a chronological reference

in form of timestamp for the ontological information

extracted during the analysis itself. More work is

needed to validate its usefulness in specific

application contexts.

Other known restrictions are 1) results

dependence on the original language of the texts from

corpus or collection and 2) results dependence on the

performance of the chosen ontology learning method.

In the first case, because most of them only support

English, we value as future work exploring the

possibilities of ontology learning methods for other

languages and studying the possibilities of language

generalization of the method.

In the second case, ontology learning (as emergent

research area) present some needs in terms of

methods and tools for evaluating and comparing

results. This means that the evaluation of the

information extracted is based mainly on a principle

of utility, satisfaction and quality perceived by the

user who will use the information extracted to make

decisions. In fact, “not so much consensus about a

delimit task of automatic extraction of ontologies”

(Buitelaar, Cimiano, & Magnini, 2005). Said that, this

lack of evaluation possibilities difficult the

comparison between approaches and the evaluation

of the results, with very few raw results published and

in non-standard formats, that could serve as a gold

standard for improving the approaches. For this

reason, we make available both the pipeline

implementation as an adaptation of the selected

ontology learning method to our pipeline in a public

repository (Martin-Rodilla, 2019). In addition, we

offer our results of extracted ontology for depression

with temporal component with the source corpus (D.

Losada & F. Crestani, 2016; D. E. Losada & F.

Crestani, 2016) under petition for research purposes,

producing a corpus of evolutionary OL results applied

to the depression detection from social media texts.

These contributions try to collaborate to the

proliferation of research resources for comparison,

dissemination and evaluation in this area.

6 CONCLUSIONS

This paper presents a double contribution:

1) a pipeline model that adds the temporal component

to current ontology learning models, performing

evolutionary analysis of both linguistic and

ontological features extracted form unstructured texts

and 2) an application case of the method to an external

corpus on depression prediction from social media

texts.

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

328

Figure 2: Original texts for a depressed subject from the depression corpus and the corresponding pipeline output: ontological

entities vector and their temporal information results.

We are at a favourable time to investigate

depression signs, due to the existence of validated

collections of reliable diagnosis’s data.

In the first case, the method adds value to the

existing ontology learning method and approaches,

allowing the evolutionary analysis of their current

outputs. However, more work is needed in order to

test the pipeline method in different application

scenarios and input corpus. In the second case, the

application showed here constitutes a valuable dataset

itself for continuing researching on the software

prediction and detection of first signs of the

depression phenomenon, with complex and varied

features that requires interdisciplinary perspectives.

ACKNOWLEDGEMENTS

Funding by “Ministerio de Ciencia, Innovación y

Universidades” of the Government of Spain (research

grant RTI2018-093336-B-C21, co-funded by the

European Regional Development Fund,

ERDF/FEDER program).

REFERENCES

Al-Mosaiwi, M., & Johnstone, T. (2018). In an absolute

state: Elevated use of absolutist words is a marker

specific to anxiety, depression, and suicidal ideation.

Clinical Psychological Science, 6(4), 529-542.

Al-Mosaiwi, M., & Johnstone, T. (2019). In an absolute

state: Elevated use of absolutist words is a marker

specific to anxiety, depression, and suicidal ideation:

Corrigendum.

Asim, M. N., Wasim, M., Khan, M. U. G., Mahmood, W.,

& Abbasi, H. M. (2018). A survey of ontology learning

techniques and applications. Database, 2018.

Breslin, J., Ellison, N., Shanahan, J., and Tufekci, Z.

(2012). Proceedings of the 6th International

Conference on Weblogs and Social Media (ICWSM -

12).

Brewster, C. (2006). Ontology Learning from Text:

Methods, Evaluation and Applications. Computational

Linguistics - COLI, 32, 569-572.

doi:10.1162/coli.2006.32.4.569

Buitelaar, P., Cimiano, P., & Magnini, B. (2005). Ontology

learning from text: methods, evaluation and

applications (Vol. 123): IOS press.

Conde, A. (2018). Neuw84 Github account: c-value

algortihm implementation. Retrieved from

https://github.com/Neuw84/CValue-TermExtraction

Coppersmith, G., Dredze, M., & Harman, C. (2014).

Quantifying Mental Health Signals in Twitter. Paper

presented at the Workshop on Computational

Linguistics and Clinical Psychology: From Linguistic

Signal to Clinical Reality, Baltimore, Maryland, USA.

De Choudhury, M., Gamon, M., Counts, S., & Horvitz, E.

(2013). Predicting depression via social media. Paper

presented at the Seventh international AAAI

conference on weblogs and social media.

Frantzi, K., Ananiadou, S., & Mima, H. (2000). Automatic

recognition of multi-word terms:. the C-value/NC-

value method. International Journal on Digital

Libraries, 3(2), 115-130. doi:10.1007/s007999900023

Gruber, T. R. (1995). Toward principles for the design of

ontologies used for knowledge sharing? International

Journal of Human-Computer Studies, 43(5), 907-928.

doi:https://doi.org/10.1006/ijhc.1995.1081

Hazman, M., El-Beltagy, S. R., & Rafea, A. (2011). A

survey of ontology learning approaches. International

Journal of Computer Applications, 22(9), 36-43.

Kiss, G., & Vicsi, K. (2017). Comparison of read and

spontaneous speech in case of automatic detection of

depression. Paper presented at the 2017 8th IEEE

International Conference on Cognitive

Infocommunications (CogInfoCom).

Kokanovic, R., Butler, E., Halilovich, H., Palmer, V.,

Griffiths, F., Dowrick, C., & Gunn, J. (2013). Maps,

models, and narratives: the ways people talk about

depression. Qualitative Health Research, 23(1), 114-

125.

Losada, D., & Crestani, F. (2016). A Test Collection for

Research on Depression and Language use. Retrieved

Adding Temporal Dimension to Ontology Learning Models for Depression Signs Detection from Social Media Texts

329

from: https://tec.citius.usc.es/ir/code/dc.html

Losada, D., Crestani, F., & Parapar, J. (2017). eRISK 2017:

CLEF Lab on Early Risk Prediction on the Internet:

Experimental Foundations.

Losada, D. E., & Crestani, F. (2016). A test collection for

research on depression and language use. Paper

presented at the International Conference of the Cross-

Language Evaluation Forum for European Languages.

Losada, D. E., Crestani, F., & Parapar, J. (2018a). Overview

of eRisk: Early Risk Prediction on the Internet.

https://doi.org/10.1007/978-3-319-98932-7_30

Losada, D. E., & Gamallo, P. (2018). Evaluating and

improving lexical resources for detecting signs of

depression in text. Language Resources and

Evaluation, 1-24.

Lossio-Ventura, J. A., Jonquet, C., Roche, M., & Teisseire,

M. (2013). Combining c-value and keyword extraction

methods for biomedical terms extraction. Paper

presented at the LBM: Languages in Biology and

Medicine.

Martin-Rodilla, P. (2019). OL Temporal Layer Depression

Project. Retrieved from

https://github.com/patrimrodilla/OLTemporalLayer_D

epression

Morales, M., Scherer, S., & Levitan, R. (2017). A cross-

modal review of indicators for depression detection

systems. Paper presented at the Proceedings of the

Fourth Workshop on Computational Linguistics and

Clinical Psychology—From Linguistic Signal to

Clinical Reality.

OMG. (2011). Business process model and notation

(BPMN 2.0), formal/2011-01-03, OMG,

http://www.omg.org/spec/BPMN/2.0 (May 2011).

Pennebaker, J., Mehl, M., & Niederhoffer, K. (2003).

Psychological Aspects of Natural Language Use: Our

Words, Our Selves. Annual review of psychology, 54,

547-577.

doi:10.1146/annurev.psych.54.101601.145041

Resnik, P., Resnik, R., & Mitchell, M. (2014). Proceedings

of the Workshop on Computational Linguistics and

Clinical Psychology: From Linguistic Signal to

Clinical Reality.

Reuter, T., & Cimiano, P. (2012). Event-based

classification of social media streams. Paper presented

at the Proceedings of the 2nd ACM International

Conference on Multimedia Retrieval, Hong Kong,

China.

Rust. (2018). Rust Project: Crate regex. Retrieved from

https://docs.rs/regex/1.3.1/regex/

Shamsfard, M., & Barforoush, A. A. (2003). The state of

the art in ontology learning: a framework for

comparison. Knowl. Eng. Rev., 18(4), 293-316.

doi:10.1017/s0269888903000687

Somodevilla, M., Ayala, D., & Pineda, I. (2018). An

Overview on Ontology Learning Tasks. Computación y

Sistemas, 22. doi:10.13053/cys-22-1-2790

Staab, S., Maedche, A., Nedellec, C., & Wiemer-Hastings,

P. M. (2000). Proceedings of the First Workshop on

Ontology Learning OL'2000, Berlin, Germany, August

25, 2000.

Wimalasuriya, D. C., & Dejing, D. (2010). Ontology-based

information extraction: An introduction and a survey of

current approaches. Journal of Information Science,

36(3), 306-323. Retrieved from

https://doi.org/10.1177/0165551509360123

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

330