A Software Assistant to Provide Notification to Users of Missing
Information in Input Texts
Mandy Goram
Faculty of Mathematics and Computer Science, FernUniversität in Hagen, 58084 Hagen, Germany
Keywords: Software Assistant, Natural Language Processing, Machine Learning, Intelligent Agents, Chatbots.
Abstract: In this paper a software assistant is presented, which supports the users of a small community app in the
creation of ads in the social marketplace by pointing out missing information and useful additions. For this
purpose, questions about potentially missing aspects of the content should create incentives to supplement the
missing information. An insight into the prototypical development of the software assistant shows that
automated support functions can be provided for the users with machine learning procedures and natural
language processing even despite data protection restrictions and less data. The focus of this paper is on the
presentation of text creation support. Its implementation reveals problems with the use of German language
models and their language processing and counteracts these with a rule-based approach. The learning ability
of the system through automated learning procedures enables the software assistants to react and categorize
to linguistic and content-related changes in the input text of the users.
1 INTRODUCTION
The app meinDorf55+ of the research project "Mein
Dorf 55 plus - Trotz Alter bleiben ich!" is a virtual
meeting place for senior citizens in the rural area of
Nassauer Land. The virtual space was created to offer
members a platform for the exchange of information
about projects and cultural activities as well as for
mutual support. This is intended to bring the members
into contact with each other and network with each
other.
The users have different areas at their disposal
through which they can come into contact with each
other and exchange information. One area, for
example, is "Handeln" (a social marketplace), in
which advertisements from the region are presented.
In the "Treffen" (meetings) section, users find
incentives to participate in events on various topics.
In addition, there are the "Mein Kreis" (my social
circle) and "Beachten" (take notice) areas, which can
be used to administer and maintain the welfare
circles.
1.1 Problem Statement
For the users, it is important to be pointed out
interesting and similar contents from the above
mentioned areas "Handeln" and "Treffen". This is
realized by the recommender system SolR. Because
of too few ads (cold-start problem) as well as too short
and incorrect texts, the analysis and evaluation of the
contents is made more difficult or even impossible by
SolR. The cold-start problem is typical for
recommender systems such as SolR (Agarwal and
Chen, 2015), (Schein et al., 2002) and leads to
inappropriate recommendations in the case of the app.
In the app meinDorf55+, not only the problems of
a cold start occur, but also too short and low-content
text (Fig. 1) as well as spelling mistakes.
Figure 1: Example of short content from the app. The too
short advertisement “Wiesbaden Philharmonic Orchestra.
Concert in Wiesbaden.” of Andreas offers SolR only few
information, which provides hardly possibilities for the
content comparison.
In order to support the recommender system in its
functions, especially in the area of "Handeln", the
users are to be supported in the writing of content by
keyword suggestions and questions on potentially
missing aspects of the content. On the one hand, these
Goram, M.
A Software Assistant to Provide Notification to Users of Missing Information in Input Texts.
DOI: 10.5220/0007728304230430
In Proceedings of the 21st International Conference on Enterprise Information Systems (ICEIS 2019), pages 423-430
ISBN: 978-989-758-372-8
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
423
incentives aim to draw the attention of users to
missing and meaningful information in order to
encourage the addition, extension and improvement
of content. On the other hand, enriched keywording
can improve SolR's recommendations on similar ads.
Due to the later integration into the existing
productive system of the app and the associated data
protection law, the solution to be implemented may
only use the user's input in the ad text during the
analysis. The permanent storage of user behaviour is
also not permitted. This takes account of the data
protection requirements of the lawfulness of data
processing (Conrad, 2018), which currently apply to
the use of the app.
1.2 Contribution
In this paper a software assistant is presented, which
supports the users of the app meinDorf55+ in the
writing of ads in the social marketplace by pointing
out missing information and useful additions. In
addition to text input support, the software assistant
also has a recommendation tool. This tool identifies
keywords that match the ad text and returns them to
the user. Keyword suggestions are based either on
similar existing ads or on entities identified from the
language analysis. The tag recommendation
component is not the subject of this paper.
An insight into the prototypical development of
the software assistant shows that automated support
functions can be provided for the users with machine
learning procedures and natural language processing
(NLP) even despite data protection restrictions and
less data. The focus of this paper is on the
presentation of text creation support. Its
implementation reveals problems with the use of
German language models and their language
processing and counteracts these with a rule-based
approach. The learning ability of the system through
automated learning procedures enables the software
assistants to react and categorize to linguistic and
content-related changes in the input text of the users
automatically.
After the description of the methodology for the
development of the software assistant, related work
as well as procedures and tools of the approach are
presented. Then the structure and procedure of the
software assistant are presented and the text and
language processing process are described. Finally,
the limitations and problems of the approach are
described and future research activities are presented.
2 APPROACH
The problems mentioned above are addressed by a
co-learning software assistant as an independent
solution, which supports the user during the capture
of ad texts through content feedback. In a workshop
with users of the app, i.e. representatives from the
senior group, a concept for integrating the software
assistant into the meinDorf55+ app was jointly
developed. The user interface is not the subject of this
work. Based on the problem analysis, the following
requirements arise for the system to be developed.
Requirements for system structure and
integration: R1: The solution to be developed should
be independent and autonomous from the
meinDorf55+ app.
The software assistant should support the user
during data entry. As can be seen from the example
of Andreas' ad (Fig. 1), support should be available
for creating the ad. Texts that are too short and lack
content should be prevented. R2: The user is to be
supported by the software assistant during ad editing.
R3: The software assistant should manage the
data independently from the meinDorf55+
application.
The solution should improve its support
performance and learn from new data. R4: The
software assistant should learn independently from
new, transmitted data.
Requirements for the support functions: The
analysis of the ads contained in the social marketplace
revealed that the user's ad texts, which are too short
and poor in content, play a significant role in SolR's
inappropriate recommendations. The solution
therefore aims to improve the quality of the text by
assisting the user in the ad creation process. R5: The
software assistant should be able to draw the user's
attention to missing and meaningful information.
The processing and analysis of the data must take
place automatically in the software assistant since it
is to be made available as independent solution
without foreign accesses. R6: The software assistant
should process and analyse the inputs of the user
independently.
When processing the data from meinDorf55+, no
user-specific information may be processed or stored.
This results in two requirements: R7: The software
assistant may only use the user inputs of the ad text
during the analysis. R8: The software assistant must
not permanently store the user behaviour.
R9: For the integration of the software assistant into
the existing app the system should be able to be
integrated into the app interface.
ICEIS 2019 - 21st International Conference on Enterprise Information Systems
424
3 RELATED WORK
No system existed at the time of writing that could
address all or part of the aforementioned problems.
The interaction between user and system belongs to
the area of chatbots and agent systems. Chatbots are
divided into simple and intelligent chatbots
(Stanoevska, 2018). The development of intelligent
chatbots is very complex and time-consuming
(Stanoevska, 2018). Since they are based on artificial
intelligence, they require very large amounts of data
(Abdul and Woods, 2015) to simulate a conversation
situation that does not have a predefined sequence.
Simple chatbots have the advantage that they are
easy to develop and have a fast response time. This is
made possible by rule-based processing of voice input
(Stanoevska, 2018). The rule-based approach is easy
to implement and therefore versatile (Stanoevska,
2018). With the Seq2Seq model it is possible to build
chatbots with grammatical, fluent responses, but it is
reported that these chatbots are usually low of
diversity because of the trivial responses. (Jiang and
de Rijke, 2018), (Vinyals and Le, 2015), (Sordoni et
al., 2015). An interesting approach to textual analysis
is described in (Naw and Hlaing, 2013). The paper
deals with a Car Recommendation System, which
identifies suitable offers for users based on
information in an e-mail. (Naw and Hlaing, 2013) use
a content-based filtering approach and use the Jaccard
Coefficient algorithm to identify predefined terms.
The information in the e-mail is pre-processed and
analysed using information retrieval and information
extraction procedures. Words are extracted from the
e-mail and checked against predefined keywords.
Based on the found keywords, relevant vehicles are
determined via a rule-based comparison. If the
information is not sufficient for an exact
determination, the most similar vehicles are
determined with the Jaccard Coefficient. (Naw and
Hlaing, 2013) work without complex and time-
consuming machine learning procedures and instead
use predefined rules for extracting important
information from textual data. This reduces the
processing time and searches for information that is
already known.
Chatbots, intelligent agent systems and
recommender systems are already widely researched
and used in the field of human computer interaction.
The prototypes and solutions, however, were mostly
developed specifically for one or a few application
areas and are not transferable to the questions and
tasks of the present work. Furthermore, there is no
known possibility to integrate single components into
the app.
The software assistant uses a rule-based approach, as
it is common in chatbots, to select the predefined text
modules. The characteristics of a co-learning system
should make it possible to react to changes in
language usage to determine the appropriate text
category for a user text. Unlike simple chatbots,
whose central component is a predefined dialog
(Stäcker and Stanoevska, 2018), no guided sequences
are provided in the software assistant. This was
explicitly not desired by the target group. The
adaptability of the software assistant is a feature in
intelligent agents such as the music information
assistant of the multi-agent framework RASCALLI
(Skowron et al., 2008).
4 COMPONENTS OF THE
SOFTWARE ASSISTANT
The software assistant is part of a complex system and
has many components and functions for processing
user input and determining missing information.
Therefore, a rough overview of the approach is given
first and then the text analysis and the support
function are described.
4.1 Design and Architecture Aspects
The whole assistant system (AS) is an independent
solution in a loosely coupled system according to R1.
Figure 2 shows the call of AS (blue) via a uniform
interface by the clients (grey).
Figure 2: Loosely coupled system structure of AS.
The interface is provided as a web service that
uses the request/response principle of the HTTP
protocol. The web service expects a request in JSON
format. AS does not distinguish the clients, so that the
Loose coupling of the systems
Webbrower
(Client)
Application Server
Software Assistent
AS
AS Database
A Software Assistant to Provide Notification to Users of Missing Information in Input Texts
425
call of the web service from the meinDorf55+
application can be integrated by a controller or by a
script in the web browser. Accordingly to R3 the data
for processing the request is managed by AS in its
own database (Fig. 2 right).
AS is based on the programming language
Python, which makes it possible to use a lightweight
web service and extensive libraries of text analysis
and machine learning. The web service (left) receives
the client's requests and forwards the transmitted data
to the software assistant.
4.2 Workflow Components
The software assistant (blue block) is the central
component in the system flow. It transforms the data
of the request (JSON format) into an internal data
format. Dataframes are used for internal processing,
whereby the data is formatted in tabular form.
Figure 3: Components of AS.
Dataframes can be used flexibly and can be extended
and supplemented during processing. They manage
different data types at the same time and have
corresponding calculation and analysis functions.
As shown in Figure 3, the software assistant
forwards the request data to the main components
(dotted area) of AS. The user inputs are normalized
by a text preprocessing, as usual in information
retrieval. In the text modules component, important
keywords (information extraction) are searched for
and used for the selection of hints. Afterwards an
analysis of the user input for the tag recommendation
takes place. The texts for the document symbols in
Figure 3 show exemplary the results of the step-by-
step processing and analysis of the ad by Andreas
(Fig. 1), which are returned to the client via the
response of the web service (left).
4.3 Webservice
For AS, the information from the title and description
of the ad is relevant without giving either information
a higher weighting, which is why it is processed as an
ad text.
The interfaces of the web service receive a
"content" information containing the combined text
of title and description of the ad. The client must link
these two pieces of information. The web service
provides various calls. Two for identifying hints with
or without returning the identified text category. Two
more for storing new ad texts in the AS database
(training modes), with or without manually
specifying the text category.
5 IDENTIFICATION OF MISSING
INFORMATION AND USEFUL
ADDITIONS
5.1 Text Preparation and Processing
In the text analysis, the text components and
individual words of the same or different texts are
compared. Comparability must be possible. Natural-
language texts, such as Andreas' text, are unstructured
and not directly comparable. Comparability can be
disturbed by upper and lower-case letters, nouns,
suffixes, prefixes and punctuation, spelling mistakes
and synonyms (Thelwall et al., 2010).
In computer-aided language analysis, therefore,
the text to be analyzed is first processed. In a first
step, the text documents are segmented and divided
into linguistic units (Carstensen et al., 2010). For the
content evaluation, irrelevant words (stop words) and
punctuation are removed from the text. Spelling can
be corrected to a certain degree using special
statistical methods, which is not done within the scope
of this paper. AS processes the text input in two steps.
Normalization: During normalization, the texts
are first tokenized and stored as word segments for
further processing. Then existing stop words and
punctuation are removed. Finally, the cleaned user
text and the reference texts are stemmed. The
stemming of texts is only of limited use in German
language processing and only useful if all texts are
stemmed. The NLP analysis also performed
AS
Database
Client
Webservice
Component: Tag
Recommendation
Component:
Text Modules
Request
HTTP/JSON
Response
HTTP/JSON
Software Assistent
Text Processing
Konzertbesuch,
Konzert, Mainz
Hints:
Uhrzeit und Datum
angeben. Anzahl
Teilnehmer?
Tags:
Konzertbesuch,
Konzert, Mainz
Konzertbesuch
Konzert in Mainz
Konzertbesuch
Konzert in Mainz
ICEIS 2019 - 21st International Conference on Enterprise Information Systems
426
insufficient for German language because free
available corpus are too low in quality. The assistant
contains the spacy model corpus.
Vectorizing: After the text data has been
normalized, it is transformed into feature vectors. For
this the TFIDFVectorizer and the BinaryVectorizer of
the Machine Learning Library scikit-learn are used.
Two vector methods (bag of words and TF-IDF) are
provided to allow a later change of the methods
without additional implementation effort.
After the vector transformation the reference data
can be used for the model construction. These are
transferred to the learning procedure together with the
model configuration. In the system the methods
Naive Bayes and the Support Vector Machine of
scikit-learn are available. The training data for the
supervised learning method were created from a
deduction (ads from the meinDorf55+ database) of
the productive system at the beginning. A label of text
category 1, 2, 3, or 4 was assigned to the ad texts,
depending on the content information. The stored
data represent the reference texts for the models of the
learning procedures.
5.2 Selecting Hint Texts
AS has a generic structure and currently four text
categories that can be extended so that users can be
made aware of missing aspects of the content via their
user interface. Each text category contains a list of
keywords associated with it and an associated
questionnaire to encourage input of missing aspects
of the content. The question catalog is the result of an
analysis of the ads in the system at the beginning and
reflects the most important content information in
order to create an informative text for the app users.
To implement the aforementioned support, the
software assistant first determines the text category
belonging to the content. With the help of the
language analysis, the missing information on
keywords is determined. The corresponding
questions are then selected from the question catalog
and displayed to the users.
5.2.1 Question Catalog
Figure 4 shows the text categories. The ads are
generalized into four categories that could be derived
from the ads in the production system at the time the
requirements were collected. Classification is based
on similar questions that should be considered when
create a meaningful ad. The question is reflected in
the type of activity or character of the ad. An example
of an activity is the ad for Andreas' concert visit. This
activity then gives rise to questions about the date,
times of departure and the beginning of the concert,
place of departure, etc.
Figure 4: Overview of the content criteria and the corresponding text classes.
A Software Assistant to Provide Notification to Users of Missing Information in Input Texts
427
As Figure 4 shows, it contains questions about the
place, the time, the participants and the target group,
among other things.
5.2.2 Determine Sequence of Text Modules
This section describes how the user is to be supported
by meaningful notes when entering texts. Figure 5
shows the sequence of the Text modules component.
Figure 5: Overview of the selection process for text
modules.
After receiving the user input, the software assistant
loads the reference data for the learning procedure
from the AS database and forwards the user text and
the reference texts to the module. These are first
normalized via the text preparation. The normalized
data is then made available for model construction
and classification (right). With the help of the
predicted label, the appropriate text modules are
selected (below). To select the appropriate text
modules, the user text is checked for predefined
keywords. The check is based on string search or
entity check from NLP analysis. If the corresponding
information is missing in the user text, the reference
texts are returned via the software assistant. A return
transmission to the client takes place after the
determination of suitable tags.
The text modules are selected according to
defined criteria, since not every hint matches every
type (class or category) of ad. The possible types of
ads have been determined and generalized on the
basis of the properties of the respective category. This
means that an ad can always be assigned to a
category. The assignment is made using a
classification procedure.
5.3 Machine Learning Methods for
Text Categorization and Linguistic
Adaptation
The software assistant system can adapt itself to new
contents and expressions of the users by using an
independent database and machine learning
procedures. For this purpose, the models are trained
online with new ad texts so that they can be used for
future analysis. Due to the modular structure of the
system, different learning methods can be used. The
missing information is determined using text
classification methods, whereby the system can either
configured to use the Naive Bayes or the Support
Vector Machine.
For this purpose, new ad texts with manually set or
automatically determined text categories can be
integrated into the training corpus via service calls.
6 VALIDATION
The verification of the determination of hint texts is
based on a short text with little content.
6.1 Text Module Determination
Figure 6 summarizes the test criteria and text
modules. The column content check indicates which
information is generally searched for. The category
column indicates which text categories use the
analysis and the text module. The hint text column
contains the text of the text module.
Figure 6: Overview of content criteria and hint texts.
Figure 7: AS returns an empty hint list if all useful
information is available in the text.
Content check
Category
Hint text
Date and time
1, 2, 3
Geben Sie eine Zeit und ggf. ein genaues Datum an.
(Specify a time and, if necessary, an exact date.)
Location
1, 2, 3
Geben Sie einen (Ziel-)Ort oder Treffpunkt an.
(Specify a (destination) place or meeting point.)
Participant
details
1, 2
Suchen Sie weitere Teilnehmer und wie viele? (Are
you looking for more participants and how many?)
Target group
information
1
Richtet sich Ihr Angebot an Amateure, Profis etc.?
(Is your offer aimed at amateurs, professionals,
etc.?)
Equipment
information
1, 2, 3
Wird spezielles Equipment benötigt und soll es
mitgebracht werden? (Is special equipment needed
and should it be brought along?)
Accessibility
1, 2
Ist der Ort barrierefrei? (Is the place handicapped
accessible?)
ICEIS 2019 - 21st International Conference on Enterprise Information Systems
428
Table 1 contains the exemplary request text, which
contains all necessary information, so no hints are
returned to the client (Fig. 7).
Table 1: Expected result of the request with a meaningful
text.
Input
Expected value
Wer hat Lust zum Tanzen? Immer
samstags ab 18 Uhr in Mainz.
Anfänger, Profis und Zuschauer
willkommen. Maximal 15
Teilnehmer. Bitte saubere Schuhe
mitbringen. Der Zugang zum
Tanzraum ist barrierefrei. (Who
wants to dance? Always Saturdays
from 18 o'clock in Mainz.
Beginners, professionals and
spectators welcome. Maximum 15
participants. Please bring clean
shoes. Access to the dance room is
handicapped accessible.)
Empty list,
because all
important
information is
available
Table 2 contains a text from the productive system of
the meinDorf55+ app. Due to its content, this text
belongs to category 3 and does not contain all the
required information. The missing information is
indicated accordingly (Fig. 8).
Table 2: Invocation of AS with unknown real data in which
data is missing.
Expected value
Assignment to category 3
Hint texts:
Geben Sie eine Zeit und
ggf. ein genaues Datum
an. (Specify a time and, if
necessary, an exact date.)
Wird spezielles
Equipment benötigt und
soll es mitgebracht
werden? (Is special
equipment needed and
should it be brought
along?)
6.2 Verification of the Extension of the
Supervised Model
By saving unknown ad texts with a suitable text
category, unknown texts can be classified correctly
later. For validation, unknown, short ad texts from the
meinDorf55+ app is tested. Fig. 9(a) shows an ad
from the app whose content is not included in the
training data. For this reason, the category 3 is
wrongly assigned instead of 4.
The still unknown ad (Fig. 9(a)) is transmitted to AS
via the training mode with indication of category 4.
Figure 8: Unknown ad text from the app meinDorf55+.
If a similar ad text like "TV Tipp - Das Urteil ?
Great film, which I can only recommend. Also
available in the ZDF media library," sent to AS, the
appropriate category is assigned to this ad text (Fig.
9(b)). AS has learned this by correcting it via the
training mode.
Figure 9: (a) Displays the wrong category for the ad text.
(b) Displays the correct category after a similar text has
been saved in that category.
7 CRITICAL ANALYSIS
The feasibility of the solution approach is
demonstrated by the prototypical implementation. It
leads to the content support of a user when entering
A Software Assistant to Provide Notification to Users of Missing Information in Input Texts
429
ad texts by hints to missing information in the user
text. An evaluation of the support functionality in the
field could not be carried out within the scope of the
present work.
The validation examined the determination of text
modules based on changing input texts, which are
returned to the caller (client). It could be shown that
AS recognizes keywords in the input texts and
correctly recognizes the content to be checked. Hints,
which are assigned as text modules to a text category,
were selected by AS and transmitted to the client if
there were no information in the input text. The hint
texts do not always fit into the context of the input
text, which was already clear when defining the text
categories.
The validation showed that the prototype meets
the requirements of an independent solution (R1),
which can process and analyse the input texts
independently (R6) based on a self-administered
database (R3). Missing information in the input text
is recognized and corresponding information texts
(R5) are returned to the client. AS learns by storing
new ad text data (R4) in its own database. No user
data is analysed (R7) or stored (R8) for processing.
With an average response time of 1.6 seconds, AS
can promptly generate a response, enabling the client
to provide the user with the information during ad text
capturing (R2). By providing the services as a web
service, the assistant can be easily integrated into
existing interfaces (R9).
The support function of AS is limited by its rough
rasterization into four text categories. This is partly
due to an insufficient named entity recognition of the
German language model. Currently, this model can
only detect locations relatively reliably. Other
information must be searched with predefined
keywords in the input text.
8 CONCLUSIONS
The solution presented in this paper will enable users
in the future (i.e. after integration into the app) to be
supported in entering ad texts in the meinDorf55+
app. For this purpose, questions about potentially
missing aspects of the content should create
incentives to supplement the missing information.
This should improve both the quality of the ads and
the recommendations generated by the app in the
"Handeln" area. A detailed evaluation of the assistant
must be carried out.
The object of further research activities is the
acquisition of additional external data and the
integration of user-specific information from the
system in order to be able to open up the context of
the individual user and to enable adequate support.
Future work will also focus on improving German
language processing.
REFERENCES
Abdul-Kader, S. A., Woods, J. C., 2015. Survey on chatbot
design techniques in speech conversation
systems. International Journal of Advanced Computer
Science and Applications, 6(7).
Agarwal, D. K., Chen, B. C., 2015. Statistical methods for
recommender systems. Cambridge University Press.
Carstensen, K. U., Ebert, C., Ebert, C., Jekat, S., Langer,
H., Klabunde, R. (Eds.)., 2010. Computerlinguistik und
Sprachtechnologie: Eine Einführung. Springer-Verlag.
Conrad, C. S. 2018. Kann die Künstliche Intelligenz den
Menschen entschlüsseln? - Neue Forderungen zum
Datenschutz. Datenschutz und Datensicherheit-DuD,
42(9), 541-546.
Jiang, S., de Rijke, M., 2018. Why are Sequence-to-
Sequence Models So Dull? Understanding the Low-
Diversity Problem of Chatbots. arXiv preprint
arXiv:1809.01941.
Naw, N., Hlaing, E. E., 2013. Relevant words extraction
method for recommendation system. Bulletin of
Electrical Engineering and Informatics, 2(3), 169-176.
Schein, A. I., Popescul, A., Ungar, L. H., Pennock, D. M.,
2002. Methods and metrics for cold-start
recommendations. In Proceedings of the 25th annual
international ACM SIGIR conference on Research and
development in information retrieval (pp. 253-260).
ACM.
Skowron, M., Irran, J., Krenn, B., 2008. Computational
framework for and the realization of cognitive agents
providing intelligent assistance capabilities. In 18th
European Conference on Artificial Intelligence,
Cognitive Robotics Workshop (pp. 88-96).
Sordoni, A., Galley, M., Auli, M., Brockett, C., Ji, Y.,
Mitchell, M., Dolan, B., 2015. A neural network
approach to context-sensitive generation of
conversational responses. arXiv preprint
arXiv:1506.06714.
Stanoevska-Slabeva, K., 2018. Conversational Interfaces –
die Benutzerschnittstelle der Zukunft?
Wirtschaftsinformatik & Management, 10(6), 26-37.
Stäcker, O., Stanoevska-Slabeva, K., 2018. Quo vadis
Chatbots? Wirtschaftsinformatik &
Management, 10(6), 38-46.
Thelwall, M., Buckley, K., Paltoglou, G., Cai, D., Kappas,
A., 2010. Sentiment strength detection in short informal
text. Journal of the American Society for Information
Science and Technology, 61(12), 2544-2558.
Vinyals, O., Le, Q., 2015. A neural conversational model.
arXiv preprint arXiv:1506.05869.
ICEIS 2019 - 21st International Conference on Enterprise Information Systems
430