A Method to Discover Digital Collaborative Conversations in

Business Collaborations

Antoine Flepp

1,2

, Julie Dugdale

, Fabrice Bourge

and Tiphaine Marie-Cardot

Research Area of Digital Enterprise, Orange Labs, 14000, Caen, France

CNRS – LIG, Univ. Grenoble Alpes, F-38000, Grenoble, France

Keywords: Conversation Threading, Communication, Collaboration, Knowledge, Knowledge Worker, Explicit

Knowledge, Tacit Knowledge, Digital Tools, Messaging Service, Digital Communication and Collaboration

Tools, Collaborative Conversation of Document Production.

Abstract: Many companies have a suite of digital tools, such as Enterprise Social Networks, conferencing and

document sharing software, and email, to facilitate collaboration among employees. During, or at the end of

a collaboration, documents are often produced. People who were not involved in the initial collaboration

often have difficulties understanding parts of its content because they are lacking the overall context. We

argue there is valuable contextual and collaborative knowledge contained in these tools (content and use)

that can be used to understand the document. Our goal is to rebuild the conversations that took place over a

messaging service and their links with a digital conferencing tool during document production. The novelty

in our approach is to combine several conversation-threading methods to identify interesting links between

distinct conversations. Specifically we combine header-field information with social, temporal and semantic

proximities. Our findings suggest the messaging service and conferencing tool are used in a complementary

way. The primary results confirm that combining different conversation threading approaches is efficient to

detect and construct conversation threads from distinct digital conversations concerning the same document.

1 INTRODUCTION

Recent years have seen a huge increase in the

number of digital applications to promote

communication and collaboration. Many companies

now have a large suite of digital tools, such as

Enterprise Social Networks (ESN), conferencing and

document sharing software, as well as standard

email and messaging clients; all of which are used to

facilitate communication and collaboration among

their workers. One of the main outputs of the

collaborative activity is a document. The

collaborative process leading up to the final version

of this document may have entailed in-depth

discussions drawing upon the deep and tacit

knowledge of the collaborators. However the final

document rarely contains any underlying rationale

and the rich context in which decisions have been

made could be buried in the digital exchanges

between collaborators. For a new collaborator, these

hidden elements may be the key to quickly

understanding the decisions that were made without

having to contact the original collaborators. We

argue that there is valuable knowledge within these

tools, which we refer to as Digital Communication

and Collaboration Tools (DCCT), can be used to

facilitate future collaboration. We focus on the

content and the use of DCCT to extract and rebuild

the Collaborative Conversation of Document

Production (CCDP). The following section discusses

related work on knowledge in business

collaborations, the use of digital tools in such a

setting, and current works on conversation

threading: the tree classification of Internet

messages. The larger goal of our work is to create a

knowledge base that will be used to store valuable

reusable knowledge about collaborations; this will

make future collaborations more efficient and

effective. The work described in this paper concerns

enriching the database with meta-data from

analysing email exchanges during collaborations and

understanding better the complementarity of using

email with a conferencing tool. The study focuses on

collaborations within the Orange Company.

However, we believe that the results will be

applicable to other companies or other institutions

Flepp, A., Dugdale, J., Bourge, F. and Marie-Cardot, T.

A Method to Discover Digital Collaborative Conversations in Business Collaborations.

DOI: 10.5220/0006899401630170

In Proceedings of the 10th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2018) - Volume 1: KDIR, pages 163-170

ISBN: 978-989-758-330-8

163

since the need to enhance collaborations is a

common goal and our method of tracing Internet

messages can be applied to all situations where

email is used. Section 3 presents this proposal and

section 4 describes the adopted methodology. The

results concerning conversation threads in

collaborative business messaging are presented and

discussed in Section 5.

2 RELATED WORK

2.1 Knowledge in Collaborations

Knowledge Worker is a concept coined by Peter

Drucker to describe a person whose main capital is

developing or using knowledge in the workplace

(Drucker, 1969). Unfortunately its application to

workers is limited by the difficulty in measuring and

making knowledge explicit (OECD, 1996).

Nevertheless since 1996 the Organisation for

Economic Co-operation and Development (OECD)

has recommended basing the European economy on

Knowledge which they consider as a driver of

productivity and economic growth, thus showing its

importance for businesses and workers.

Tacit knowledge has been described by the idea

that “we can know more than we can tell” (Polanyi,

1966). Nowadays knowledge is considered as a

dynamic concept, moving along a spectrum from

explicitness to tacitness (Flepp et al., 2017).

Although a document’s content may be

considered as explicit knowledge, the process of

understanding the document may be considered as

uncovering tacit knowledge (Flepp et al., 2017). We

hypothesize that this understanding could be greatly

aided if we can capture valuable and reusable

information embedded in DCCT that were used by

the original CCDP. The evaluation of the relevance

and a faster exploitation of extracted information

will be the foundation of our knowledge base.

2.2 Digital Communication and

Collaboration Tools (DCCT)

With the advent of new DCCT many researchers

believed that older tools, such as email, would be

replaced by newer tools such as Enterprise Social

Network (ESN). However this has not been the case

(Leclercq-Vandelannoitte et al., 2017) and a

superposition effect, called the “Napoleon effect”,

has been observed (Isaac et al., 2008). This effect

refers to stacking digital tools in an organisation

without replacing or substituting the old digital

tools. ESNs are not replacing the messaging service

but are used in a complementary way. For example,

(Alimam et al., 2016) showed that there was no

correlation between an email graph representing 37

workers’ professional relations and an ESN graph of

the same workers. Indeed ESNs can be used “to

expand a worker’s scope of relations for future

collaboration or communication”. Although it is

difficult for a knowledge worker to use an ESN to

generate new collaborations (Pralong, 2017),

engagement and networking can increase by 15%

per year (Lecko, 2018). At Orange, the ESN has

been available to approximately 150,000, employees

for over two years (Orange, 2015) with 41% active

collaborators at the end of 2017. The Orange ESN,

called Plazza, allows employees to create a group

around business or extra-professional issues. Plazza

has the typical features of a social network

(comment, "like", messaging), but it also has

features to facilitate daily work (project management

tool, co-editing of documents, enriched directory).

A frequently used DCCT at Orange is CoopNet,

a conferencing tool that is used by 58% of

collaborators. An organizer can create Virtual

meeting rooms and participants can talk via a

conference call and view shared documents or

applications. CoopNet has an average of 4,800

teleconferences per day (around 1,200,000 in 2016).

This shows the complementarity between different

DCCT, e.g. by using email to send an invite or

report of a teleconference. (Flepp et al., 2017) also

showed that collaborations could depend on the type

of the collaboration, the collaborators involved, and

the used media.

We believe that each service should not be

considered individually in the CCDP, but as one in a

suite of DCCT available to the knowledge worker.

To extract knowledge of a CCDP we focus on email

and its link with a CoopNet; two DCCTs most

frequently used at Orange.

2.3 Conversation Threading

Conversation threading is a feature provided by

DCCT to easily group Internet messages and their

replies. Linked Internet messages constitute a

conversation thread. However there exist different

approaches for conversation threading.

First, messages sent via the Internet adhere to an

Internet Message Format that is defined by a set of

RFC specifications. RFC 5322 (Resnick, 2008)

specifies the syntax for text messages sent via email.

The header text in emails provides a way for

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

164

conversation threading. The header fields

“References:” and “In-Reply-To” can be used to

identify a conversation thread i.e. storing IDs to

retrieve the origin of an email conversation, its

replies, forwarding, or CCs. Email servers often use

these two header fields to build their own

conversation threads, called a Thread-Index

(Microsoft, 2013). However a Thread-Index is not

always available for email clients (Yeh and Harnly,

2006; Palus and Kazienko, 2010). For example if a

new email is sent there will be no associated Thread-

Index until the email has been answered, because the

Thread-Index is generated by the email server.

Therefore, some properties of an Internet message

are only available for email clients and not for the

server messaging service. To thread conversations

from a messaging service it is necessary to use other

properties than just the “References:” header field.

Therefore, the amount of research on conversation

threading, not just on header fields, is increasing

(Domeniconi et al. 2016). To thread conversations

from a DCCT, three other properties in the email are

often used: textual content, collaborators, and the

date and time of sending an email (time proximity).

(Palus and Kazienko, 2010) use only two

properties (collaborators, and date and time) for

email conversation threading. They compare their

results with those of (Klimt and Yang, 2004) that

also used two properties (textual content and,

collaborators). Palus and Kazienko failed to rebuild

as many conversation threads as Klimt and Yang.

However Palus and Kazienko’s results differ by only

2% to 4% depending on the time window. This

supports the idea of using properties other than the

“References:” header field, such as time proximity.

(Yeh and Harnly, 2006) compare two methods

for conversation threading. The first, a header-based

method, uses the Thread-Index generated by the

messaging service server. The second is a similarity-

matching method that combines the three properties

noted above: textual content, collaborators and time

proximity. The header-based method is easier to

implement and it performs reasonably well in most

cases. However, the method fails when the Thread-

Index is encoded by different server messaging

services or when there is no header information.

Thus, Yeh and Harnly suggest two advancements.

The first is the combination of their two methods

described above. The second is the extension of their

similarity-matching method with any natural

language processing (NLP) tool, because their

similarity-matching method does not compare the

meaning of a sequence of words but only identical

words.

Another work (Domeniconi et al., 2016) is closer

to ours. Their approach identifies conversation

threads from a heterogeneous pool of internet

messages from different DCCTs, such as social

networks, emails, forums, etc. Although we did not

implement deep learning, as did Domeniconi et al.,

their work shows the complementary use of different

DCCTs. The complementarity is detected on

properties of textual content, collaborators and time

proximity from different DCCTs. Finally, they

suggested a variation of their approach, which is:

“the reconstruction of conversational trees, where

the issue is to identify the reply structure of the

conversations inside a thread”.

The innovation we propose is to combine these

two approaches (header-field information and social,

temporal and semantic aspects) not only to

reconstruct conversation threads, but also to identify

logical links between distinct conversations.

3 PROPOSITION

First, we define a Collaborative Conversation of

Document Production (CCDP). We then present a

use case of conversation threading at Orange from

two DCCTs. We introduce the concept of a

Collaborative Conversation Thread as being: a

conversation thread based on email header

information and similarity matching. We suggest

that this complementary use of different approaches

of conversation threading will lead to a richer

knowledge base than if we used only one

conversation threading approach with only one

DCCT. Our main contribution is conversation

threading in a CCDP, which we believe will be

useful to detect conversation threads in different

conversations from one or several DCCT.

3.1 Collaborative Conversation of

Document Production (CCDP)

We define a CCDP on two criteria: information

exchanges about the production and sharing of the

document, and the complementarity between the

DCCTs. For the first criterion we retain three

properties to identify conversation threads: a group

of collaborators, working on the same topic over a

limited time period. For the second criterion, we

focus on the complementarity between the two most

commonly used DCCTs at Orange: messaging

service and its link with a conferencing tool.

A Method to Discover Digital Collaborative Conversations in Business Collaborations

165

3.2 Existing Conversation Thread

(ECT)

At Orange, the conversation threads of the

messaging service and the conferencing tool are

independent; consequently they have no common

Thread-Index. However, in practice a knowledge

worker will often use their messaging service to

send invites to a teleconference or to report on a

teleconference. Therefore to identify the

complementarity of the messaging service and a

conferencing tool, we distinguish three types of

messages of a knowledge worker messaging service:

emails, meetings (as they appear in the messaging

service calendar) and meeting notifications (e.g.

invites, acceptances). Note that we only focus on

meetings containing a link to a teleconference. The

final goal is to rebuild ECTs of these three types of

messages and find new relevant links between them.

3.3 Collaborative Conversation Thread

A collaborative conversation thread is a weighted

link made by proximities between two messages

from different ECTs. A collaborative conversation

thread logically links together different ECTs to

constitute a CCDP (Figure 1). For example, if the

subject field in an ECT of emails is changed, the

messaging service will create a new conversation

thread. This happens even if there is a minor

modification. We presume that we can find

proximities between different ECTs by detecting

relevant links between two messages from different

ECTs.

Figure 1: Collaborative conversation thread in a CCDP.

4 METHODOLOGY

First we define an equation to link different ECTs.

Then we manually apply this equation for a first

qualitative evaluation.

4.1 Global Proximity (GP)

We calculate a Global Proximity (GP) between any

two messages of the type email, invite to a meeting,

or meeting. If a GP value is above a certain threshold

and the two associated messages belong to different

ECTs, we consider that we have detected a

collaborative conversation thread and we assign

these ECTs to the same CCDP.

We define GP as the weighted average of three sub-

proximities: Interlocutors Proximity (IP), Time

Proximity (TP), and Semantic Proximity (SP), the

value of which are all bound in the interval [0,1].

GP = (a.IP + b.TP + c.SP) / (a + b + c) (1)

Coefficients a, b and c may be used to ajust the

impact of any sub-proximity.

4.1.1 Interlocutors Proximity (IP)

IP aims to identify how closely the interlocutors are

linked through collaboration. A messaging service

can identify three different roles of interlocutors in

an Internet message. Using the RFC 5322 field

names these are: From, To and CC. However we

define a fourth role, which is used when an

interlocutor is absent from a specific message

(Absent) while still being part of the conversation.

Note that we did not take into account the blind

carbon copy role (BCC) because it does not directly

concern an active collaboration and it is rarely used

at Orange to collaborate. Also, we did not consider

the Internet messages sent to oneself since this does

not constitute a collaboration. To calculate the IP we

attribute an IP coefficient to each combination of

roles involved in a conversation (Table 1). Between

messages M

and M

a coefficient W

with a value

in the interval [0,1] is defined for each pair of

messages to reflect how collaborator’s roles change.

Table 1: Coefficients of Interlocutors Proximity.

From To Cc Absent

From W

To W

Cc W

Absent W

The combinations involving the two main roles

(From and To) are likely to have higher values than

those involving the secondary role CC. The

combinations involving the Absent role have null

values (by default) or low values if necessary. After

identifying the different interlocutor roles in a

conversation and assigning the coefficients we

calculate the IP; this is the sum of the IP coefficients

divided by the total number of interlocutors

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

166

(InterlocNbr) between M

and M

(

Ei,E

=  W

)/ InterlocNbr

(2)

4.1.2 Time Proximity (TP)

We define TP as the difference between the date and

time of sending an Internet message (t

) and another

one (t

). To bound this proximity in the interval [0,1]

we use an exponential inverse function:

TP = e

–(t2-t1)/k

(3)

Note that k may be modified to have different

temporal granularities (e.g. week, day, hour),

provided that t

and k have the same unit.

4.1.3 Semantic Proximity (SP)

We use textual semantic proximities between

subject-fields, between names of attachments, and

between subject-fields and names of attachments.

We currently do not take into account email body

contents or attachment contents to calculate

semantic proximity. Thus we define SP as the result

of three semantic sub-proximities: Subject Semantic

Proximity (SSP), Attachment Semantic Proximity

(ASP) and Cross-Semantic Proximity (CSP).

To calculate these semantic proximities, we use a

NLP tool developed at Orange, called SimBow

(Charlet and Damnati, 2017). SimBow not only

generates a semantic proximity but detects and

stores semantic and lexical relations between

different bags-of-words. This is useful in our

approach where different subject fields in Internet

messages may refer to the same topic even if the

subject fields have no words in common. For

example, “launching” and “closure” are easily

recognizable by a human as relating to a project, but

are not as easily identifiable to a computer without

tools such as SimBow. SimBow supplies a

semantical proximity bounded in the interval [0,1].

If two texts are semantically close the calculated

proximity tends to 1 (i.e. identical texts have a

calculated proximity of 1) and tends to 0 if the two

texts are semantically distant. Finally, SP is the

maximum proximity found between the three sub-

proximities:

SP = max(SSP, ASP, CSP) (4)

Subject Semantic Proximity (SSP)

To calculate the SSP, we delete the possible prefixes

“RE” or “FWD” of subject texts and we use

SimBow to give us a semantic proximity between

two texts. If at least one of the two messages that are

compared has an empty subject text field, the use of

SimBow is not necessary and SSP = 0.

Attachment Semantic Proximity (ASP)

ASP is calculated in a similar way to SSP. When a

message has more than one attachment, we currently

concatenate their names into a single string. This

simple approach was chosen since it provides us

with a first step in semantically comparing

attachments. If at least one of the two internet

messages to compare does not have any attachment,

SimBow needs not be used and ASP = 0.

Cross-Semantic Proximity (CSP)

CSP is useful to identify the semantical proximity

between a subject text (subj) of a message and the

name of the attachment (att) of another message. For

two Internet messages M

and M

we define CSP as:

CSP

(Mi,Mj)

= max[SimBow(subj M

att M

SimBow(subj M

att M

)]

(5)

As with SSP and ASP, if at least one of the two

text fields is empty, CSP = 0.

4.2 Computation of Proximities

The equations were first tried on a small data set of

message exchanges that were selected by a

knowledge worker who considered them as a CCDP

in his own work.

This CCDP is composed of 3 conversations. A

first conversation of 7 exchanges starts with an

administrative person requesting the head of a

research program to close a project. The latter then

transfers the information to the relevant people

asking them to prepare a Powerpoint document. The

second conversation is composed of two exchanges

requesting a teleconference to close the project

“officially”. The third conversation concerns the

invitation to the requested teleconference and the

distribution of the meeting report.

Although these three conversations are distinct

and have different subjects, they are logically linked

as they relate to the same objective (“project

closure”) and share versions of a Powerpoint

document that evolves over time. After the data set

was built, four collaborators each subjectively

attributed a global proximity measurement to each

pair of messages. Note that the collaborators knew

A Method to Discover Digital Collaborative Conversations in Business Collaborations

167

about the body content of the messages, whereas our

equation does not yet take this into account. We

refer to this small sample of subjective global

proximities as the Gold Standard. This was

calculated by averaging the four subjective global

proximities. To evaluate and validate our equation,

we compare the Gold Standard to the Calculated

Proximities.

5 RESULTS

A value of 360 hours for k was found to give the

largest standard deviation in time proximities; this

best discriminates time proximity values. We then

subjectively defined different weighted coefficients

to calculate the Interlocutors Proximity (Table 2).

Table 2: Weighted Coefficients for Interlocutors

Proximity.

From To Cc Absent

From

1 1 0.25 0

To 1 1 0.5 0

Cc 0.25 0.5 1 0

Absent 0 0 0

For the weighted coefficients of the Global

Proximity, we currently use an equal distribution (a

= b = c = 1/3).

5.1 Data Dispersion Analysis

Although the data set is small we analyse the data

dispersion in order to compare the difference

between the Gold Standard and the Calculated

Proximities. We look at the proximities between any

two messages globally, between messages that

belong to the same conversation thread, or between

messages that belong to different conversation

threads. Table 3 shows the results at the global level

regardless of whether two messages belong or not to

the same ECT.

Table 3: Data dispersion between the Gold Standard (GS)

and Calculated Proximities (CP).

GS CP

Max = 0.950 0.979

Min = 0.350 0.097

Average = 0.640 0.493

Average Absolute Deviation = N/A 0.186

The average absolute deviation may vary

between 0 and 1. Therefore, any value represents an

error rate between the Gold Standard and Calculated

Proximities, which can be translated in percentage.

Hence, with a global level analysis we obtain an

average absolute deviation of 0.186, i.e. an 18.6%

error rate. Thus, the Calculated Proximities are

significantly different from the Gold Standard and so

they are not very satisfactory. If we analyse the data

further and look at two Internet messages in the

same and different ECTs we obtain Table 4 and 5.

Table 4: Data Dispersion between the GS and CP for

messages belonging to the same ECT.

CP in a same

ECT

Max = 0.950 0.979

Min = 0.350 0.387

Average = 0.617 0.603

Average Absolute Deviation = N/A 0.110

Table 5: Data Dispersion between the GS and CP for

messages belonging to different ECTs.

CP in different

ECTs

Max = 0.875 0.719

Min = 0.500 0.097

Average = 0.655 0.415

Average Absolute Deviation = N/A 0.241

The average absolute deviation of 0.186 in Table

3 is largely due to the average absolute deviation

between messages of different ECTs, that has a

value of 0.241 (Table 5), an 24.1% error rate.

Contrarily the fairly low average absolute deviation

in Table 3 (11%) indicates that the proximities

calculation is efficient between two messages in the

same ECT. Nevertheless, the average absolute

deviation of 0.241 in Table 5 means that we should

improve the calculation of proximities when

applying the equation to pairs of messages from

different ECTs.

5.2 Proximities Balancing

We also analysed the Sub-Proximities. Applying a

threshold to the Gold Standard values identified

pairs of messages that have the highest subjective

proximities. However the data dispersion analysis in

Table 3 indicates that we cannot directly compare

the Gold Standard with the calculated proximities

since the difference between the two sets of data, in

particular regarding average values, is too great.

Hence we apply a compensatory coefficient to the

Gold Standard so that its average equals that of the

Calculated Proximities (Table 6).

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

168

We should apply the compensatory coefficient to

the computed proximities rather than to the Gold

Standard (which is the reference). However, some of

the CPs would have a value greater than 1, which is

meaningless.

Table 6: Data dispersion between the GS and CP after

applying a compensatory coefficient.

GS CP

Max = 0.733 0.979

Min = 0.270 0.097

Average = 0.493 0.493

Average Absolute Deviation = 0.124 N/A

After applying this compensatory coefficient, the

average absolute deviation between the Gold

Standard and calculated proximities changes from

0.186 to 0.124, giving a more reasonable

comparison between with fewer differences between

the dispersions of the datasets.

5.3 Proximities Analysis

Figure 2: Collaborative Conversation Thread detection.

Figure 2 shows the calculated proximities of the

collaborative conversation threads. We focus on the

Gold Standard proximities with the highest values

between the three conversation threads (A, B and C).

An interesting calculated proximity is the one

between messages A7 and B1, which belong to

different ECTs and share an attachment name

identical to the subject of the conversation A. The

calculated proximity is 0.719 and the Gold Standard

value is 0.636 which is close (less than 0.1).

From the sub-proximities in table 7 the semantic and

temporal proximities have a value of 1 and 0.977

respectively, whereas the interlocutors proximity is

only 0.179. This means that a collaborative thread

has been identified by the calculation of proximities

(and verified by the Gold standard) and that this

collaboration is due to the messages being

semantically and temporally close, and not due to

interlocutors’ closeness.

Table 7: Example of a temporal and semantic proximities.

Global

Proximity

Interlocutors

Proximity

Temporal

Proximity

Semantic

Proximity

0.719 0.179 0.977 1.000

Moreover, this calculated proximity corresponds

to a pair of messages that has the second highest

proximity value in the Gold Standard, signifiying

that the calculation of proximities is humanly

relevant. More precisely, we can see that the

calculated proximity between A7 and B1 is greater

than calculated proximities between B1 and other

messages of conversation A (Figure 2), as

knowledge workers done it in the Gold Standard.

A second calculated proximity of interest is the

highest proximity value in the Gold Standard

between B2 and C1. This calculated proximity is

close to the Gold Standard value (0.590 and 0.675,

respectively). This is encouraging because it

demonstrates that our approach is efficient

significantly linking distinct conversations involving

collaborations in document production.

Some calculated proximities differ significantly

from the Gold Standard as for example with

messages A2 and B1 (Table 8):

Table 8: Example of a lack of proximity.

Messages Gold Standard Calculated Proximity

A2 - B1 0.540 0.097

The calculated proximity differs significantly

from the Gold Standard because the equation does

not take into account the body content. Therefore it

is unable to make any semantic proximity between

A2 and B1 (Table 9).

Table 9: Example of a lack of global proximity by

analysing the sub-proximities.

Global

Proximity

Interlocutors

Proximity

Temporal

Proximity

Semantic

Proximity

0.097 0.250 0.040 0.000

6 DISCUSSION - FUTURE WORK

An interesting point is that we are able to identify

the most relevant proximities. However, we can

alsoidentify less relevant proximities between two

messages within the same conversation. This is

useful since we will be able to delete those messages

A Method to Discover Digital Collaborative Conversations in Business Collaborations

169

that may be irrelevant to collaborations. This helpd

to highlight more clearly the elements of the future

knowledge base to aid collaboration. A first

validation of the equation was done by comparing

the calculated proximities with the subjective global

proximities. A next step of validation will be to

assess if the knowledge base is of practical use to a

new member of the collaboration who has not been

involved in the final document production. In

addition we will extend our approach by: applying

the calculation of proximities to more data; using

data from different digital communication and

collaboration tools, and taking into account the

message body content to improve the semantic

proximity and thus the global proximity.

7 CONCLUSIONS

In this paper, a new approach for conversation

threading is introduced that combines a header-field

method and a proximity method. Since the header-

field method is insufficient for conversation

threading on a client messaging service, it is

completed by the proximity method using social,

temporal and semantic proximities between

messages. The results show that our equation can

detect that the messaging service and conferencing

tool are used in a complementary way during

document production. In addition, we confirm that

the combination of different approaches of

conversation threading is useful to link distinct

digital conversations when they concern the same

document. These primary results suggest our

approach is useful to construct a knowledge base

that will aid document understanding.

REFERENCES

Alimam, M., Bertin, E., & Crespi, N., 2016. Enterprise

2.0–Literature Taxonomy and Usage Evaluation. IFIP

Working Conference on The Practice of Enterprise

Modeling. Springer, Cham, pp. 26-40.

Charlet, D., Damnati, G., 2017. Simbow at semeval-2017

task 3: Soft-cosine semantic similarity between

questions for community question answering. In:

Proceedings of the 11th International Workshop on

Semantic Evaluation (SemEval-2017), pp. 315-319.

Domeniconi, G., Semertzidis, K. Moro, G., Lopez, V.,

Kotoulos, S., M. Daly, E., 2016 Identifying

Conversational Message Threads by Integrating

Classification and Data Clustering. In: International

Conference on Data Management Technologies and

Applications. Springer, Cham, pp. 25-46.

Drucker, P. F., 1959. The Landmarks of Tomorrow. New

York: Harper and Row.

Flepp, A., Bourge, F., Dugdale, J., Marie-Cardot, T., 2017.

Extraction de connaissances à partir de l’usage des

outils professionnels de communication et de

collaboration. In: Conférence Nationale d’Intelligence

Artificielle, pp. 55-60. Éditors: Berger Levrault.

Isaac, H., Kalika, M., Charki, N.B., 2008. An empirical

investigation of e-mail use versus face-to-face

meetings: Integrating the Napoleon effect perspective.

Communications of the Association for Information

Systems, vol. 22(1), article 27.

Klimt, B., & Yang, Y., 2004. Introducing the Enron

Corpus. Conference on Email and Anti-Spam (CEAS).

Lecko, 2018. État de l’art de la transformation interne des

organisations. [online] Available at: http://referentiel.

lecko.fr/etatdelart_tome10/ [Accessed April 2018].

Leclercq-Vandelannoitte, A., Bertin, E., Colléaux, A.,

2017. EMAIL: SURIVOR OR WALKING DEAD?

An exploratory study of the potential replacement of

email by enterprise social networks. Association

Information et Management, AIM2017.

Microsoft.com, 2013. Finding related messages in a

mailbox by using EWS in exchange 2010. [online]

Available at: https://msdn.microsoft.com/en-us/

library/office/ff406135(v=exchg.140).aspx [Accessed

April 2018].

Orange.com, 2015. Orange lance son réseau social interne

de 2

ème

génération à l’échelle mondiale. [online]

Available at: https://www.orange.com/fr/Press-Room/

communiques/communiques-2015/Orange-lance-son-

reseau-social-interne-de-2eme-generation-a-l-echelle-

mondiale [Accessed April 2018].

Organisation for Economic Co-operation and

Development (OECD), 1996. The Knowledge-based

Economy. [online] Available at: http://www.oecd.org/

sti/sci-tech/1913021.pdf [Accessed April 2018]

Palus, S., Kazienko, P. (2010). Conversation Threads

Hidden within Email Server Logs. World Summit on

Knowledge Society. Springer, Berlin, Heidelberg , pp.

407-413.

Pralong, J., 2017. Mode collaboratif ou collaboratif à la

mode ? Pourquoi les réseaux sociaux d’entreprise

peinent encore à développer des comportements

collaboratifs. [online] Available at: http://www.igs-

ecoles.com/wp-content/uploads/pdf/rh-presse.pdf

[Accessed April 2018].

Polanyi, M., 1966. The tacit dimension. 1

ed. New-York:

Doubleday.

Resnick, P., 2008. Internet Message Format, RFC 5322,

DOI 10.17487/RFC5322. [online] Request for

Comments. Available at: https://www.rfc-editor.org/

info/rfc5322 [Accessed April 2018].

Yeh, J.-Y. and Harnly, A., 2006. Email Thread

Reassembly Using Similarity Matching. In:

Proceedings of Conference on E-mail and Anti-Spam

(CEAS’06).

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

170