Using PageRank for Detecting the Attraction between Participants
and Topics in a Conversation
Costin Chiru, Traian Rebedea and Adriana Erbaru
University Politehnica of Bucharest, Department of Computer Science and Engineering,
313 Splaiul Independetei, Bucharest, Romania
Keywords: CSCL, Natural Language Processing, Participant Assessment, PageRank, Online Conversations.
Abstract: In this paper we present a novel approach that uses the well-known PageRank algorithm for assessing multi-
threaded chat conversations. As online conversations can be modelled as directed graphs, we have
investigated a method for allowing a real-time analysis of the conversation using PageRank by computing
the ranks of the utterances based on the explicit and implicit links available in the discussion. This model
has been also extended to offer a method for computing connections between the debated topics and the chat
participants and between each of the debated topics in the conversation, called the participant-topic and the
topic-topic attraction. The results presented in this paper are promising, but also reflect several important
differences between the existent offline analysis tools for chats and the PageRank method.
1 INTRODUCTION
Chat conversations (instant messaging) represent
nowadays one of the most popular methods of
exchanging ideas online. The easiness in learning
how to use chats and the high efficiency in
transferring the information, promoted chats as one
of the favourite environments for Computer
Supported Collaborative Learning (CSCL) tasks
requiring online and synchronous textual
interactions among participants (Stahl, 2006; Stahl,
2009). Due to this fact, it has been largely adopted in
CSCL activities and it has been enhanced with
functionalities specific to these tasks such as the
explicit referencing mechanism and the whiteboard
facility present in ConcertChat (Muhlpfordt and
Wessner, 2005). Still, in spite of its popularity and
of the huge quantity of data that is exchanged
through chats, there are very few application aimed
at analyzing this type of content (Chiru et. al, 2011;
Rebedea et. al, 2011). More than that, the existing
applications are built starting from a semantic
analysis (Chiru et. al, 2011; Rebedea et. al, 2011)
but the analysis takes far too much time to be used
as a real-time process and can only be applied
offline, at the end of the conversation. Therefore, we
have been searching for a different method to
analyze these conversations faster and got
influenced by the algorithms which are used by
search engines that have to analyze huge quantities
of data in a very short time. Thus, we reached the
conclusion that if the PageRank algorithm (Page et.
al, 1998) could be adapted for chats, this method
could be applied online (displaying the results of the
processing as the conversation unfolds) and
interactive (to signal what threads should be debated
more and involving people who contributed less on
specific threads), this way improving the learning
process and enhancing the participants' innovation.
To achieve this, we started from PolyCAFe
(Rebedea et. al, 2011), a system that is using
innovative methods for analysing CSCL chat
transcripts, helping both computer-assisted learning
and the tutors in evaluating the discussions. This
system analyzes chat logs using Natural Language
Processing (NLP), Latent Semantic Analysis (LSA)
and Social Network Analysis (SNA) techniques in
order to identify the most important utterances from
the conversation (in terms of their content and of the
participants’ involvement in the discussion).
Therefore, this paper presents an extension of
PolyCAFe’s functionality, trying to enhance it with
the ability to analyze CSCL chat sessions in real-
time. The first step of our analysis consists of
detecting the important utterances from the
conversation using the PageRank algorithm. Once
these utterances are identified, we use the PageRank
294
Chiru C., Rebedea T. and Erbaru A..
Using PageRank for Detecting the Attraction between Participants and Topics in a Conversation.
DOI: 10.5220/0004798202940301
In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 294-301
ISBN: 978-989-758-023-9
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
algorithm for detecting what is the attraction of each
participant towards the debated topics and, at the
same time, what is the probability of a topic to
follow another topic within the conversation.
Most of the existing approaches for analyzing
text using social network analysis (SNA) tools are
oriented towards systems that own explicit
referencing tools, such as forums or blogs. The
reason behind this orientation is the ease in
constructing the participant social network based on
the order in which the messages are sent and on the
recipient of the message. Still, there are a few
systems that intended to apply SNA tools to chat
conversations. One such tool was built by
Sundararajan (2010) for analyzing the content
published by the participants to 8 different courses
in order to observe how the respect and influence
earned by each participant influences their efforts to
"collaborate, learn new and conceptual knowledge"
and their satisfaction regarding the courses outcome.
Unfortunately, the author does not mention whether
this analysis is done manually or automatically.
Moreover, a regular SNA method is used for
evaluating the participants from the perspective of
their centrality, betweenness, in-degree, out-degree,
etc. in the network, which represent only
quantitative data. On the other hand, we are rather
interested in what the participants communicate
(what are the topics they know or they are interested
on) and in the interaction patterns between different
concepts that are debated in the conversation, which
is part of a qualitative evaluation of the participants,
topics and the conversation as a whole.
A more similar approach was undertaken by
Tuulos and Tirri (2004). The authors present a semi-
supervised system that uses a combination of topic
modelling and SNA to improve the information
retrieval from chat conversations. For their analysis,
they have used conversations taken from
SearchIRC.com which allowed them to use simple
heuristics in order to identify to whom each
utterance is addressed (and therefore to build the
social network). For this participant network the in-
degree, out-degree and PageRank of each participant
are determined. After that, the authors use some
existing conversations to detect the probabilities of
words to appear in conversations about different
topics, so that when they analyze new conversations
to be able to use these probabilities. Finally, they
evaluate the use of each of the SNA technique in
improving the information retrieval, considering as
baseline the results provided by the topic modelling.
Still, this approach gives them two advantages: first
of all they know both how many and what topics
should be present in the conversation (therefore
knowing what represents off-topic and being able to
discard that part); secondly, they have chosen the
topics from different topics (Bible, C++, Philosophy,
Physics, Politics, Win2000) thus simplifying the task
of identifying to what topic a given concept
corresponds. In our approach neither of these facts
can be exploited: since our system does not have a
learning phase, it gives the possibility to analyze
texts debating about any topics, without being
limited to the ones that were learnt (thus providing
generality in use). At the same time, it can be used to
distinguish between concepts that are from the same
or similar conceptual area. The examples presented
in this paper contain concepts from a single domain
(Human Computer Interaction) especially to prove
that the approach works even at this level, without
requiring that different topics to be debated in the
same conversation.
The paper continues with a short overview of the
PageRank algorithm. Then, we present the
application that has been developed and several
results that have been obtained by employing the
PageRank method adapted for CSCL chats. The
paper ends with an analysis of these results and with
our conclusions regarding the improvement of the
results’ quality.
2 OVERVIEW OF THE
PAGERANK ALGORITHM
Because previous researches have modelled an
online conversation as a graph with implicit and
explicit links between utterances (Rebedea et. al,
2011), we have started to consider that the PageRank
algorithm (Page et. al, 1998) may be a candidate for
the conversation graph analysis. PageRank is an
algorithm that was initially designed for the analysis
of a set of web pages in order to extract the relative
importance of each page from the considered set of
web pages (Page et. al, 1998). The algorithm
expresses the probability that a web surfer will be
able to “find” the considered page within a limited
number of steps (clicking on the links from one page
to another). It is a customization of a “random walk”
in a graph, which in turn is modelled as a Markov
chain in which the states are pages, and the
transitions, which are all equally probable, are the
links between pages.
The formal definition given in the initial paper
describing PageRank (Page et. al, 1998) was: if u is
a web page; Fu (forward links), the pages referred
UsingPageRankforDetectingtheAttractionbetweenParticipantsandTopicsinaConversation
295
by u; || the number of forward links; Bu
(backward links) the ones that refer u, c a
normalization constant and E(u) a source of rank to
make up for the rank sinks (such as cycles) with no
out-edges, than the value of the rank R(u) can be
computed using:
∗
∈
∗
(1)
In order to compute the vector R(u), one starts from
the square matrix (we’ll call it A) having the web
pages on the rows and columns and ,
1/ if there is a link from page to page or 0
otherwise. If R is the a vector of scores over the web
pages, then we can write , which can
be re-written as
1
because the
values of PR are normalized and therefore
‖
‖
1. That means that R is the eigenvector of 1
and the method should also try to maximize the
value of c (Page et. al, 1998). Subsequent research
showed that the optimal value for c should be 0.85
(Brin and Page, 1998).
The value of R can be obtained in an iterative
manner, starting from a vector of values over the
web pages (S) that can have any values (could be the
vector E(u)), using the following iterative algorithm
(Page et. al, 1998):
←
:
←
←
‖
‖
‖
‖
←
←
‖
‖
(2)
, where is a factor for increasing the convergence
rate and for maintaining
‖
‖
1, while10
.
A web page will have a high PageRank if the
sum of the webpages’ PageRank that refers it is
large. This property covers two possible cases: when
a page has many other pages referring it, or when it
is referred by pages with high PageRank.
The PageRank algorithm has proved to be
suitable not only for Google’s rank of web pages,
but also for other tasks in various domains: replacing
the ISI factor with a new formula based on the
PageRank Algorithm (Bollen, Rodriguez and Van de
Sompel, 2006), ranking academic doctoral programs
based on their records of placing their graduates in
faculty positions (Schmidt and Chingos, 2007),
predicting how many people (pedestrians or
vehicles) come to the individual spaces or streets
(Jiang, 2006), performing Word Sense
Disambiguation (Navigli and Lapata, 2010), etc.
Thus, we hoped that it could also work for chat
analysis especially as a conversation can be seen as a
graph of links between utterances where discourse
flows in a similar manner to the importance of the
web pages.
In order to use this approach, we considered that
each utterance from the chat conversation represents
a different document and in order to simulate the
forward and backward links, we used explicit and
implicit links (details will be provided in the next
section). Thus, we managed to develop a method for
very fast identification of the important utterances
from a chat, of the major threads of discussion, and
of the participants’ attraction towards these threads.
3 PAGERANK FOR AUTOMATIC
ASSESSMENT OF CSCL
CHATS
As we have already mentioned, our application starts
from PolyCAFe project and uses some of its
features:
Detection of the areas of high collaborative
discourse from a chat;
Evaluation of the collaboration of each participant
in the discussion (based on multiple criteria);
Graphical representation of the results.
3.1 Pagerank for Conversation
Analysis
We consider that PageRank is appropriate for chat
evaluation as this problem is very similar to the
original problem for which it was initially designed.
The sparsity in the chat is (in our opinion) similar to
the sparsity of relevant content from the web.
Therefore, we consider that the probability of a
person to "land" on a specific page from the web is
similar to the probability of a participant to reply to
a given utterance, while the links between different
pages are well simulated by the semantic
connections represented by the repetitions of the
same word and by the explicit "reply to" links.
Practically, this probability of a participant to reply
to a given utterance can be considered the "rank of
an utterance" (and it highlights the importance of
that specific utterance in the conversation).
The first step in applying the PageRank
Algorithm was the identification of the links that
exist between chat utterances. Since the pre-
processing part of our application was borrowed
from PolyCAFe, we also kept the input format of the
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
296
chats, which allowed the existence of explicit links
(references provided by the chat participants to
specify to which previous utterance their answer is
addressed).
Besides the explicit links, one can also encounter
the situation when two or more utterances contain
concepts that are strongly related and therefore their
authors consider that there is not necessary to
provide an explicit link. We considered that this
situation is a special case of connection (an implicit
link) and tried to identify it in order to augment the
number of explicit connections (that was insufficient
for our purposes). Therefore, we considered that
words repetitions (Chiru et. al, 2011) are example of
such links. If a term appears in an utterance, all lines
that follow and contain that term are considered
implicit links to the initial utterance. Given the
nature of the algorithm, the two types of links that
we consider (explicit and implicit) have equal
weight.
Once we detect all these links, we build utterance
chains (which can be interpreted as discussion
threads since they debate the same concepts) starting
from these links using the DFS (Depth First Search)
Algorithm, thus finding all the existing separate
chains. They are needed for determining the
attraction between two different threads (topics).
The steps that should be followed in order to
determine the threads are:
1. Identify all the utterances that are not referenced
(neither by explicit nor by implicit links) – these
utterances are probably off-topics and therefore
they are ignored;
2. All the remaining utterances are considered to be
roots for the DFS Algorithm;
3. From each of these utterances (considered in the
order they appear in chat) we start a function
(implementing DFS) to detect the threads that
can be built starting from that utterance;
4. Each function will return a thread of utterances.
The next step is to create the transition matrix
corresponding to the chat utterances by considering
the links identified between them. The explicit and
implicit links between two utterances will provide a
value of 1 in the matrix, while the remaining
elements are set to 0 (meaning that there is no
connection between the corresponding two
utterances). Once this matrix is built, it needs to be
normalized with respect to the sum of the elements
from each column.
Finally, the values of the PageRank algorithm for
the given matrix are obtained using the power
method implementation provided by the JAMA
library - A Java Matrix Package (Hicklin et. al, n.d.).
The operations made for the detection of the
eigenvalues and the eigenvectors are:
1. Apply the eig method, which decomposes the
matrix in two other matrices: a matrix D
containing the eigenvalues and a matrix V
containing the eigenvectors;
2. The maximum (dominant) eigenvalue from the
diagonal matrix D is determined and its index is
stored;
3. The dominant vector is the column from the
matrix V having the index identified in the
previous step;
4. The values from this vector are normalized with
respect to the sum of its elements;
5. The final values (the PageRank) are the values
obtained for the normalized eigenvector vd
(utterance[i].rank = vd[i]).
Once these values are determined, one can
evaluate the participant-topic attraction and the
topic-topic attraction as described in the following
sections.
3.2 Participant-topic Attraction
The participant-topic attraction defines the
participants’ drive to get involved in the discussion
of a given concept therefore proving its interest or
knowledge related to that concept. To determine this
factor, we have used the values of the participant’s
utterances containing the words that define the
considered topic.
If p represents a participant and t a topic, then the
attraction between p and t is given by the following
formula:
,
∗
,
∈
∈
(3)
In order to highlight this method, we provide an
example that proves how the above formula works.
For this, we have made the simplifying assumption
that all the utterances belong to the same participant.
Utt: < debated topics > utt value
u1: t t y x 0.5 = rank(u1)
u2: t y z x 0.3 = rank(u2)
u3: t t t z 0.2 = rank(u3)
Using formula (1), the following results will be
returned (2):
,
2∗
1
2
(4)
UsingPageRankforDetectingtheAttractionbetweenParticipantsandTopicsinaConversation
297
3∗
3
,
1
2
,
1
2
,
2
3
In the end, all these values are normalized.
3.3 Topic-Topic Attraction
The topic-topic attraction defines the probability of
having a specific topic following another topic in the
flow of a conversation. To determine it, we use the
utterance chains taking into account both the
frequency of each topic and the case when they co-
occur in the same utterance (topics are very closely
related) or occur separately (more loosely).
Therefore, we extract the threads corresponding
to the two topics and build a matrix for each chain.
This matrix reflects the debating of those topics in
the utterances and their corresponding values within
that chain.
The relationship between the topics and the
utterances is reflected by the matrix that is built as
follows:
1. We build the chain – topic matrix (ct),
represents the value of the topic in the
utterance where
a. 0 if the topic is not debated in
the utterance ;
b.
_
#_
if the -th topic is
debated in the utterance .
2. After filling in the matrix, we apply formula (5)
for each topic t
i
and t
j
.
∑
∈
_
∑
∈
_
∑
∈
_
∈
|
(5)
Below we present an example of matrix (6) for
determining the topic – topic attraction for the
following chain: u5 → u4 → u3 → u2 → u1
utt: <debated topics >
u1 t1 t2 t3
u2 : t1 t5
u3 : t2 t4
u4 : t1 t3
u5 : t1 t2
Then, the attraction between topic t1 and t2 topic is
given by (5) by applying the formula (3), with the
matrix from (4), obtaining formula (7).
3.4 User Interface
The user interface allows the input file selection, and
afterwards the content of this file is analyzed and the
results are displayed in tabular form (see Fig. 1).
The left part of the GUI presents the values for
participant – topic attraction for the selected
participant, while the right part gives the values of
the topic – topic attraction for the selected topic.
Besides the values obtained for the participant –
topic attraction and the topic-topic attraction, the
application also outputs the most important
utterances from the chats computed using their
PageRank.
The results proved to be much stricter comparing
to the results obtained using PolyCafe system or
provided by the human reviewers (Gold-Standard).
This is due to the fact that only very few utterances
have a PageRank greater than 0.
In order to provide an example, we present a part
of the utterances evaluated as being important by
PageRank algorithm. We will use the same chat for
which we presented the examples from the
participant-topic and topic-topic attraction examples.
The automatic analysis performed with PolyCAFe
has identified 132 important utterances (out of 430)
as important. From these, the PageRank algorithm
also identified 17 utterances as (see Table 1). The
values for the ranks computed by PageRank may
seem pretty low, but this is what usually happens
when applying the algorithm on any graph.
Besides these 17 utterances that were considered
important by both PolyCAFe and the PageRank
method presented in this paper, the latter has
identified another 28 turns that were not considered
important by the former. In order to account for
these utterances, we analysed PolyCAFe’s results in
order to discover a possible explanation. At a careful
analysis, we observed that the 28 extra utterances
were marked by PolyCAFe as being continuations
of other utterances. Therefore, it is possible that
PolyCAFe did not consider these utterances to be
1
/3
1
/3
1
/3 0 0
2
/2 0 0 0
2
/2
0
3
/20
3
/2 0
4
/20
4
/2 0 0
5
/2
5
/2 0 0 0
(6)
1,2
1
∗2/32/24/25
2/23/2
(7)
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
298
Figure 1: Application Graphical User Interface.
Table 1: The utterances identified by our algorithm that receive a grade higher than 8 by PolyCAFe.
Utt.
No
PageRank
score
PolyCAFe
score
Utterance Content
169 0.002 10.07
yes, they have wikis that are publicly available, with public information, for the
everyday user that takes an intrest in that company's products
167 0.002 10.01
all major companies have wikis for their technologies. most people like to search
wikis cause they provide accurate and easy to access information. Also, that way
our database servers won't be so used
404 0.202 10.01
Indeed. Our companies image will grow if we have a forum, a blog, a wiki and a
cool web-site that customers or developers can use
310 0.004 9.93
the only problem that still remains is that we need someone to check wiki articles,
blog and forums posts so that classified data does not accidentally reach a "public"
area
348 0.004 9.28
A svn is an open-source revision control system. Users can work on a version of the
application code and commit it. If two users are working on the same thing when
they commit a merge is made with the 2 versions
308 0.001 9.03
we can use a person or a team of people to handle the wiki posts, forums posts,
wave documents and all the other important stuff
331 0.002 8.93
I mean everyone of our employees knows how to use a wiki, forum blog and chat,
and google wave has a extremely friendly interface
314 0.003 8.85 but how can you use a filter for a forum?
399 0.009 8.79
well i think chat is important for our employees, it helps them talk and colaborate,
spare time by not meeting in conferences that much, and be on track with all theit
colleagues are doing
312 0.002 8.78 We can use filters for that firewalls. That can save some money
388 0.047 8.45 not necesarly computers,you can change acounts
333 0.001 8.4 evrybody can use a chat , forum , bog or something like that
423 0.015 8.32 good night everyone, and thank you for your collaboration:)
341 0.004 8.21 We can also use a SVN for our code. What do you think of this?
376 0.042 8.12 this could make them loose time...
373 0.029 8.08 they will use another machine. the restrictions will be only for certain computers
387 0.025 8.07
for every function you don't remember in a programming language, you will have to
move to other computer to find out... but it's ok ... it wouldn't be a big problem i
guess
UsingPageRankforDetectingtheAttractionbetweenParticipantsandTopicsinaConversation
299
important, since the same ideas were present in the
previous utterances, but PageRank, through its
nature, favours this kind of utterances since being
identified as continuations it means that they have
links from other utterances that were considered
important in the past and therefore they receive a
part of these utterances’ rank.
For a better evaluation of our method, we asked
30 students from the Human-Computer Interaction
class to annotate 3 different chat conversations with
the most important utterances. Thus each of the three
chat conversations was evaluated by 10 different
students. We computed the inter-rater agreement
using Fleiss' Kappa for m raters and we have
obtained the values for Kappa 0.133 for the first
chat, 0.142 for the second and 0.177 for the third
one, while the p-values were always 0.000. These
results show how difficult this task is even for
humans. When we computed the results obtained by
our method with the gold standard results provided
by the annotators, we obtained the values of kappa
0.085 for the first chat, 0.0882 for the second and
0.0894 for the last one. These results are below those
of the raters, but we have remarked that if we
discard the last 15% of the utterances in all chat
conversations, where the PageRank accumulated too
much, the results are much better: 0.131 for the first
chat, 0.128 for the second and 0.173 for the last chat
conversation. These results are closer to the inter-
rater agreement and highlight that we should add a
decaying factor for utterances that are closer to the
end of the discussion (as they have fewer out-going
links and thus the rank tends to accumulate in them).
4 INTERPRETATION
OF RESULTS
There are several important observations that can be
drawn up based on the results that we have obtained
and analyzed. First of all, the computed ranks for the
utterances are rarely different from 0, this fact being
generated by several reasons:
Most of the chats contain very few explicit links
(they seem to be ignored quite often by the chat
participants) – we have observed a direct
dependence between the number of explicit links
from the chats and the number of utterances
having non-zero values after applying the
PageRank algorithm.
The PageRank algorithm determines the
utterances’ ranks as a random walk in the graph
of utterances. The significance of these values is
the probability to get to a certain utterance after a
number of steps that goes to infinity. Therefore,
once the algorithm gets to a (relatively small) set
of utterances (lines / columns from the transition
matrix), it will be very difficult to get out of that
set (in the context of random walk) and so the
remaining values will tend to be 0.
The PageRank Algorithm is designed for the
web, where a lot of links exist between different
resources (therefore creating large chains of
links, most of the times having a lot of cycles),
while in the chats the utterance chains are usually
short and rarely having such cycles. Besides, the
topics repetitions might not be synchronized with
the explicit references, therefore not leading to
cycles.
In the current version of the system, we proposed
equal importance for the explicit and implicit
links, which can lead to determining a value too
high / low for some utterances, depending on the
number of words used in that utterance and in the
one to which it is linked.
Secondly, most utterances having values greater
than 0 are positioned at the end of the discussion.
This happens due to lack of explicit links, and
therefore those utterances accumulate a very high
score due to topics repetition, which propagates
through the chat from the beginning until the end. A
solution is needed to link these utterances to other
ones in the chat.
Finally, there are some utterances considered
significant by the PageRank algorithm, but not by
other algorithms or by human evaluators. The reason
is the same as for the previous observation: some
utterances (that are not very significant in terms of
conversation) may receive high values because of
the rank accumulation over time from other
utterances that contain the same topics.
There are a couple of solutions that could be
tried in order to alleviate the presented problems. A
first solution might be the detection of dialog acts
and adjacency pairs, since the main problem of the
proposed method is the lack of explicit links. This
way, one can detect the dialog acts that are present
in the chat (question - answer, agreement –
disagreement, greetings and so on) very quickly and
to use these links as additional explicit links.
Another possibility is to use LSA or lexical chains to
find out more semantic connections between
utterances.
Another solution to avoid reaching too many
zeroes for the computed ranks is to use the Iterative
Method instead of the Power Method for computing
the PageRank values. This way, one is not
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
300
constrained to apply the algorithm until convergence
(after an infinite number of steps), but can stop after
a limited number of steps, so that fewer utterances
reach a zero-value influence.
Finally, in order to be able to discriminate
between the importance of explicit and implicit
links, one can use a generalized algorithm based on
Markov chains having different values for different
link types (explicit or implicit links).
5 CONCLUSIONS
To sum up, our current adaptation of the PageRank
algorithm for online conversations (using only the
explicit and implicit links given by the topics
repetitions) is not powerful enough to provide results
that have the desired accuracy compared with other
solutions that analyse the discussions offline. The
main explanation is that there are not enough
explicit links added by the participants during the
discussion and using only repetitions for detecting
implicit links does not build a graph that is dense
enough. However, the method is much faster and it
can be used online and in real-time for the dynamic
evaluation of multi-threaded discussions involving
multiple participants.
Moreover, in our opinion the assumptions made
in this paper are novel for the analysis of online
discussions and they have not been used to assess
the importance/rank of an utterance in an online
discussion although PageRank follows from
previous work in citation analysis (where the links
between papers are made explicit by authors). The
preliminary results also support the use of PageRank
to compute the most important utterances in a multi-
party online conversation, but several improvements
of this method need to be investigated in order to
achieve similar results to the current state of the art
methods that also employ linguistic analysis.
ACKNOWLEDGEMENTS
This research was supported by project No.264207,
ERRIC-Empowering Romanian Research on
Intelligent Information Technologies/FP7-REGPOT-
2010-1.
REFERENCES
Bollen, J., Rodriguez, M. A. and Van de Sompel, H. 2006.
Journal Status. In: Scientometrics 69 (3), pp. 669-687.
Brin, S.; Page, L. 1998. The anatomy of a large-scale
hypertextual Web search engine. In: Computer
Networks and ISDN Systems 30: 107–117.
Chiru, C., Cojocaru, V. Trausan-Matu, S., Rebedea, T. and
Mihaila, D. 2011. Repetition and Rhythmicity Based
Assessment for Chat Conversation. In: ISMIS 2011,
LNCS 6804, Springer, pp 513-523.
Hicklin, J., Moler, C., Webb, P., Boisvert, R., Miller, B.,
Pozo, R., Remington, K. Jama: a Java matrix package.
(http://math.nist.gov/javanumerics/jama/ - accessed
16/08/2012).
Jiang, B. 2006. Ranking spaces for predicting human
movement in an urban environment. In: International
Journal of Geographical Information Science 23 (7),
pp. 823–837.
Muhlpfordt, M. and Wessner, M. 2005. Explicit
referencing in chat supports collaborative learning.
Paper presented at the Proceedings of CSCL 2005.
Navigli, R. and Lapata, M. 2010. An Experimental Study
of Graph Connectivity for Unsupervised Word Sense
Disambiguation. In: IEEE TPAMI, 32(4), IEEE Press,
pp. 678–692.
Page, L., Brin, S., Motwani, R., Winograd, T. 1998. The
PageRank Citation Ranking: Bringing Order to the
Web, Technical Report. Stanford InfoLab.
Rebedea, T., Dascălu, M., Trausan-Matu, Armitt, G., and
Chiru, C. 2011. Automatic Assessment of
Collaborative Chat Conversations with PolyCAFe. In:
Proceedings of ECTEL 2011, LNCS 6964, Springer,
pp. 299-312.
Schmidt, B. M. and Chingos, M. M. 2007. Ranking
Doctoral Programs by Placement: A New Method. In:
PS: Political Science and Politics 40, pp. 523–529.
Stahl, G. 2006. Group cognition. Computer support for
building collaborative knowledge. Cambridge: MIT
Press.
Stahl, G. 2009. Studying Virtual Math Teams. New York:
Springer.
Tuulos, V. H. and Tirri, H., 2004. Combining topic
models and social networks for chat data mining. In:
Proceedings of WI’04, pp. 206–213.
Sundararajan, B., 2010. Emergence of the Most
Knowledgeable Other (MKO): Social Network
Analysis of Chat and Bulletin Board Conversations in
a CSCL System. Electronic Journal of E-Learning,
8(2), pp. 191-207.
UsingPageRankforDetectingtheAttractionbetweenParticipantsandTopicsinaConversation
301
