Twitter is both a micro-blogging service and a platform for public conversation. Direct conversation is facilitated in Twitter through the use of @'s (mentions) and replies. While the conversational element of Twitter is of particular interest to the marketing sector, relatively few data-mining studies have focused on this area. We analyse conversations associated with reciprocated mentions that take place in a data-set consisting of approximately 4 million tweets collected over a period of 28 days that contain at least one mention. We ignore tweet content and instead use the mention network structure and its dynamical properties to identify and characterise Twitter conversations between pairs of users and within larger groups. We consider conversational balance, meaning the fraction of content contributed by each party. The goal of this work is to draw out some of the mechanisms driving conversation in Twitter, with the potential aim of developing conversational models.
The rapid uptake of online social media, combined with consumer behavioural
changes around television and news broadcasting, has instigated a sea change
in attitudes within the advertising and marketing sectors. A frequently
encountered adage is that “everything is about conversation and not about
broadcasting” [
Although it may not always be clear, even from message content, what
intention a user had in mind when posting—information seeking or information
sharing, broadcasting or conversation—we have tried to specifically extract
conversations by focusing our data-analysis on reciprocated tweets. Moreover, we
have completely ignored the content of conversations and concentrated on
structural and dynamic properties of the underlying mentions network. Our main
objective was to mine actionable insights that could inform our knowledge of
conversational mechanisms and the frequency/timings of tweets. Our hope is
that empirical observations and quantifiable insights from this analysis could
inform a simple, data driven model of the timing and structure of Twitter
conversations. One possible application would be for automated recommendations
of conversation trends, as discussed in [
A large number of registered Twitter accounts are operated by automated
software scripts, known as bots [
In Section 2, we give an account of previous work in this space. Our results of pairwise and multiple conversations and the Twitter dataset we used are presented in Section 3. Finally, in Section 4 we summarise and describe possible directions of future work. 2
The phenomenal uptake of Twitter over the last few years has resulted in a
rapidly growing interest in mining Twitter data and particularly sentiment
analysis of tweets. A recent study analyzing a large amount of Twitter and
Facebook data [
A study of conversations within a sample of 8.5k tweets collected over an hour
long period [
“...Tweets with @ signs are more focused on an addressee, more likely
to provide information for others, and more likely to exhort others to do
something—in short, their content is more interactive. ”
Twitter conversations also contain both momentarily salient or ‘peaky’ topics,
signified by increased word-use frequency of specific terms, as well as more
‘persistent conversations’, in which less salient terms recur over longer periods [
In [
In our work we completely ignore content, instead focusing on timing, structure and balance of conversation between pairs of individuals as well as multi-user conversations. Our contribution is an attempt to map the structure of Twitter exchanges over a relatively large dataset, while o↵ering some new methods to mine conversation data and improve statistical models of dialogue. 3 3.1
Data The Twitter data-set investigated in this paper was collected on our behalf by Datasift, a certified Twitter partner, allowing us to access the full Twitter 3 Tie-strength is an increasing function of the number of exchanged messages between two people and the number of messages exchanged between them and their mutual friends. firehose rather than being rate-limited by the API. The data-set consists of all UK based4 Twitter users that sent tweets with at least one mention between 8 Dec 2011 and 4 Jan 2012 (28 days in total). In the remainder of the paper, use of the word ‘tweet’ will specifically mean tweets containing at least one mention. Mentions are messages that include an @ followed by a username. Thus if person a puts “@b”, it designates that a is addressing the tweet to b specifically. Mentions are not private messages and can be read by anyone who searches for them. A tweet can be addressed to several users simultaneously using @ repetitively. Any Twitter user can mention any other Twitter user, they don’t have to be related in any way. Since conversational characteristics are influenced by many factors, including language, culture, community membership etc., one has to keep in mind the natural limitations of the results of our analysis.
We preprocessed the data, removing empty mentions and self-addressing5 and created a directed multigraph, or mentions network, containing 3, 614, 705 timestamped arcs (individual mentions) from a total of 819, 081 distinct usernames, or nodes. Of these distinct usernames, 732, 043 were “receivers”, i.e. to whom a message was addressed, and 137, 184 were “tweeters”, i.e. people who tweeted a message with a mention. There were approximately 50k nodes that appeared both as tweeters and receivers. Note that our graph is a multigraph, meaning that multiple arcs are allowed between pairs of nodes, each having a direction and timestamp. 3.2
Conversations
An important feature of both face-to-face conversation [
To establish if this is the case, we assume the null hypotheses that contributions are independent events with probability PA that party a contributes to a conversation and thus probability PB = 1 PA that party b contributes. For a given interaction sequence of length N between parties a and b, we are interested in the number of occurrences of B following A and vice-versa. We call these transitions, thus the sequence ABAABBA of length N = 7, has 4 transitions. Note that we focus on reciprocated interactions, meaning that each party makes at least one contribution and consequently that there is by default at least one transition in all interactions that we consider. We call the remaining transitions the excess transitions. For any sequence of length N , the maximum possible number of excess transitions is clearly N 2. Under the null hypotheses, excess 4 All Twitter users appearing in our data-set had selected the UK as their location. 5 Self-mentioning was surprisingly common in the data-set: 12,680 die↵rent users created a total of 44,319 self-mentions, with the maximum being 5,586 from an automated service that advertises itself at the end of each tweet. transitions occur with probability PT = 2PA(1 PA). Since we assume that transitions are independent, the probability distribution of a given number of excess transitions is binomial, and thus the expected number is ET = (N 2)PT with variance VT = (N 2)PT (1 PT ).
To test the null hypothesis, we consider all reciprocated pairwise interaction sequences in our Twitter data-set. For each sequence having nX contributions from party X 2 { A, B}, we assume that the probability of party a contributing is simply nA/(nA + nB). This does not yield any problematic probabilities (i.e. 0 or 1) since both parties always make at least one contribution.
Each sequence may have a di↵erent number of interactions and a di↵erent transition probability, but assuming that the pairwise interactions are independent, the expectation and variance of the ensemble is simply equal to the sum of the interaction expectations and variances respectively. Doing this, we find that the expected number of transitions is 85,390 with a standard deviation of 226.3, but we observe 88,758 transitions in practice, more than 15 standard deviations above the expected value. We take this as strong evidence that we can reject the null hypothesis and thus infer that the data contains a significant level of turn-taking and hence conversation.
Each sequence of pairwise interactions may constitute a number of di↵erent conversations, but ascertaining when one conversation ends and another begins may be an extremely dicult task, especially when the goal is to apply an automated processes to a large data-set. Instead of using a time-intensive lexical analysis, we investigate whether we can detect conversations by applying a simple threshold rule to the time gap between responses, where we assume that a time gap that is larger than the threshold indicates the start of a new conversation.
This method requires that we can identify a suitable threshold. To achieve this, we divide each sequence of pairwise interactions up according to a given threshold, then define distinct conversations to be reciprocated sub-sequences, i.e. sequences containing a contribution from both parties. Thus the number of sub-sequences nI is always larger than the number of distinct conversations nC. In Fig. 1(a) and (b) we plot the mean number of sub-sequences and the mean number of distinct conversations respectively over a range of threshold values. The number of distinct conversations nC has a peak value at approximately 9hrs. This peak is expected, since we only count reciprocated interactions as distinct conversations. Thus small threshold values, which split an interaction sequence up into a large number of short sub-sequences (see Fig. 1(a)), result in relatively few distinct conversations because many of the sub-sequences feature contributions from only one party. High threshold values also result in a small number of conversations, but this is simply because they do not split the sequence up into many sub-sequences. Thus the maximum at 9hrs is a natural choice of threshold and corresponds to one’s intuition that conversations may reflect diurnal patterns.
The mean and median number of tweets during conversations were 13.09 and 4 respectively, but the distribution was heavy tailed (see Fig. 2). 5.5 5 4 3.5 30 1.35 1.3
We now consider whether the number of contributions from each party are similar, or ‘balanced’ within pairwise interactions and conversations. For a given sequence of tweets, there are two ways to compute balance, we can either consider the ratio of means b = hmax(nA, nB)i/hmin(nA, nB)i or the mean of ratios = hmax(nA, nB)/ min(nA, nB)i. We will use the subscripts ‘I’ and ‘C’ to denote whether these have been calculated for interactions or conversations respectively. Since we only consider reciprocated interactions, both quantities are well-defined and we would generally expect b < . For the total number of interactions between pairs, we find that bI = 2.424 and I = 3.457. Thus on average, one party contributes around 3 times as much as the other. For the sub-set of conversations, we find that bC = 1.148 and C = 1.425. These are much closer to 1, and hence more what we would expect from typical, balanced conversations. The distribution of conversation contribution ratios is plotted in Fig. 3(a), which illustrates that conversations are most likely to be balanced, but some extremely unbalanced conversations do occur. In Fig. 3(b), for each 100 x a m n 50 00 101 Balance 102 50 nmin 100 150 minimum conversation contribution nmin = 1, 2, 3, . . . , we compute the mean of the maximum contribution nmax. There is a roughly linear trend (the grey line is nmax = 1.148nmin + 1), which further illustrates conversational balance. By allowing multiple @ signs in one message, a Twitter user could send a tweet to several recipients simultaneously, facilitating multi-user conversations or multicasting. Note that because of the 140 character limit there is a physical limit on how many users each message can be multicast to.
In this part of analysis, our aim is to – Identify multi-users exchanges; – Determine how many users typically engage in them; – Identify their time-frame, pace and how balanced they are.
In addition, are all users equally involved, or do some dominate the discussion? Are the same people at the heart of di↵erent multi-user conversations? What are the enablers and inhibitors of conversation flowing in the sense of pauses between consecutive contributions? 3.4
Identification of multi-users conversations
The reciprocated mentions data represents a directed multi-graph G (where
an edge from A to B implies at least one edge from B to A), thus
multiuser exchanges correspond to strongly-connected6 subgraphs of G with k > 2
participants. We ran a non-recursive version of Tarjan’s algorithm [
In order to take the time elapsed between consecutive messages into account, we use the same threshold method explained in subsection 3.2, this time demanding for an exchange to be a“conversation” that there is a contribution from all parties and got relatively similar results (see Fig 4). The number of exchanges which had contribution from all parties was at peak around 9 and 11 hours. We took a threshold of 9 hours which gave us 334 multiuser conversations of sizes 3, 4, and 5 (see Fig 5a).
Most users (out of 646) in our dataset were involved in just one multi-user conversation, but a small number were involved in multiple conversations. The users’ involvement in multi-user conversation is illustrated in Fig 5b.
When examining the time-frame of multi-user exchanges, we found that the correlation coecient between the total number of exchanges between clique members and the average di↵erence between consecutive exchanges was 0.244 (see Fig 6a). This was not surprising, since we would expect lively conversations (with lots of exchanged messages) to have a relatively fast pace, in contrast to a casual exchange of messages with longer di↵erences inside our chosen 9 hour time-window. The same picture is obtained from looking at the median time di↵erences between consecutive messages across di↵erent clique sizes (see Fig 7 Maximal cliques are the largest complete subgraphs containing a given node. 6b). We also investigated how balanced multi-user exchanges were, although this situation is more complicated than in the pairwise case.
Firstly, we looked at the di↵erence between the number of tweets received and sent by individual clique members. For each node, we computed the di↵erence of their in-degree and out-degree. We summed up the positive values8 and to normalise, we divided by the total number of exchanged messages. In this way, we obtained a percentage of ‘unreciprocated’ messages, where reciprocity is not 8 Clearly the number of sent and received messages within a group are equal, thus summing the di↵erences between in- and out-degree over individual members in the group is by definition equal to zero. toward a sender but toward a whole group. We show the histograms for the di↵erent sizes of cliques in Fig. 7a. Across all clique sizes and in most of the multi-user conversations around 30% messages were unreciprocated. In a small number of conversations of 3 or 4 users a larger percentage were unreciprocated, i.e. they were dominated by certain members, but also a large number of cliques were very balanced (with unreciprocated messages at 0 10%), meaning every individual received and sent a similar number of tweets.
Finally, we looked at so-called ‘floor-gaining’ [
We looked at conversations in Twitter, based on the underlying structure and timings in approximately 4 million UK tweets with mentions over a period of 28 days. We structured the data as a multigraph to make use of graph algorithms. We proposed a simple method of identifying conversations between pairs of users, based on a time-threshold on the time-to-next tweet, and found evidence that a threshold of 9hrs gives a good indication of distinct conversations. We observed that the conversations detected using this method appeared to be balanced, meaning that each party involved contributed approximately equally to the conversation. This was not the case within more general interactions, in which one agent typically contributed around three times as much as the other.
Although finding cliques in graphs is computationally demanding, because of the sparsity of interactions patterns within the data-set, extracting multiuser exchanges was feasible and relatively fast. We were able to find all cliques within the graph and, using the threshold method, identify conversations for up to a maximum of 5 users. Most of those exchanges were fast-paced. We also found that the number of messages in multi-user exchanges was reciprocal to the average time di↵erence between them. When looking at the balance of multi-user conversations, we found that most exchanges are dominated by just one or two users, with some evidence of well-balanced group exchanges in between 3 users. Regarding the number of received and sent messages by each individual in a group, we found that some were dominated by one or two users, but also some were well balanced.
Further work needs to be done using content information to explore how topics flow through multi-user exchange and if there is any relationship between time-di↵erences between messages and topic. We hope that the insights gained from our analysis could help to develop an understanding of the mechanisms and dynamics of Twitter conversations, with potential scope for generating models of micro-blogging behaviour.
This work is partially funded by the RCUK Digital Economy programme via EPSRC grant EP/G065802/1 ‘The Horizon Hub’ and EPSRC MOLTEN EP/I016031/1. We would like to thank Datasift for the provision of the data analysed, and to Colin Singleton and Bruno Gon¸calves for very useful feedback and comments.