1107/1107.3275.md

Hate networks revisited: time and user interface dependence study of user emotions in political forum

Pawel Sobkowicz
KEN 94/140, Warsaw, Poland*

Antoni Sobkowicz
Mechatronics Department, Warsaw University of Technology Sw. Andrzeja Boboli 8, 02-525 Warsaw, Poland

The paper presents analysis of time evolution within an Internet political forum, characterized by large political differences and high levels of emotions. The study compares samples of discussions gathered at three periods separated by important events. We focus on statistical aspects related to emotional content of communication and changes brought by technologies that increase or decrease the direct one-to-one discussions. We discuss implications of user interface aspects on promoting communication across a political divide.

I. INTRODUCTION

Internet discussion fora are very fertile grounds for research of human communication patterns and social structures in statu nascendi – that is when the links between communicating people can literally be observed as they form, together with the content, timing structure and emotional tone of the messages. Many Web sites provide tools that allow the users to express their views, comment important topics and reply to posts by other users. Resulting discussions are often only slightly moderated, allowing various kinds of expressions, ranging from elaborate texts to single words or emoticons; from polite discussion to exchanges of obscenities. The Internet allows the users to remain relatively anonymous, thus they are free from anxiety of expressing extreme views due to possible retributions typical face to face contacts or formal correspondence. At the same time, certain stability of nicknames allows recognizability within the discussion platform, so that social networks may form, grow and evolve. The resulting social networks stretch across geographic distances, social status, age and political divides. The discussion fora have attracted significant research attention: Mullen and Malouf [1], Kelly et al. [2, 3], Schuth et al. [4], Schuth [5], Wu et al. [6], Wu and Huberman [7], Gómez et al. [8], Grabowski et al. [9], Grabowski [10], Kulakowski et al. [11], Tsagkias et al. [12], Lee et al. [13], Schweitzer and Garcia [14], Si et al. [15, 16], Ding et al. [17], Ding and Liu [18], Chmiel et al. [19]. These works described multiple characteristics of user behavior, both from statistical point of view, describing social network properties, and from social dynamics perspective (opinion spreading, emotions expressed by the users). There are also works focused on discussions spurred by personal blogs (Jeong [20, 21], Mishne and Glance [22], Leskovec et al. [23]) and of networks formed by the blogs themselves (Adamic and

Glance [24], Trammell [25], Hargittai et al. [26]).

The importance of such studies results not only from the freedom of expression mentioned above, but also from the variety of motivations driving specific discussion fora. These may vary from helpful assistance (e.g. in computer technology), virtual gatherings of aficionados of particular activity (sports, music fans, entertainment...), reviews and opinions concerning specific products or services (hotels, gadgets, books...) to political discussions.

In our previous work (hereafter referred to as Paper I, Sobkowicz and Sobkowicz [27]), we have presented results of studies of Internet discussions powered by strong negative feelings, within highly polarized Polish political environment. Paper I compared statistical properties of such interactions with those of less contentious topics (for example sport or computer technology discussions) and presented a simple simulation model, in which large role was given to pairwise exchanges of comments between individual participants. Our motivation was to see, if there are particular properties of social networks that are formed by linking representatives of conflicted sides. Such hate based networking is quite unusual outside the Internet, because in real life voluntary social links are mostly based on common interests and views. We have found that the network based on negative emotions may be quite extended, and that statistical behavior shows many similarities to networks based on cooperation and shared interests, e.g. power law distribution of indegree and outdegree. These observations are in agreement with those of Chau et al. (Chau and Xu [28, 29]) who studied the network structure of ‘hate groups’ and Chmiel et al. (Chmiel et al. [19]), who analyzed a large dataset of discussions of the BBC political, religion and news fora.

The goal of this work is to extend the scope of Paper I in two directions. The first is to broaden the scope: we monitor the same news site for over two years, looking for elements that remain stable and those that change. The second direction is to study effects of change in visual presentation of the discussions and other elements of the user interface on resulting social networks, emotion levels and capacity to communicate.

During the two years that have passed since gather-

---

*Electronic address: pawelsobko@gmail.com; URL: http://countryofblindfolded.blogspot.coming of data for Paper I, the political split in Poland has significantly increased. Tragic crash of the plane carrying the President Lech Kaczynski resulted in snap presidential elections. As no candidate received a majority of votes in the first round, a second round was held on 4 July 2010 in which Bronislaw Komorowski, candidate of Platforma Obywatelska (Civic Platform, PO) defeated Jaroslaw Kaczynski, candidate of Prawo i Sprawiedliwosc (Law and Justice, PiS), twin brother of the president who died in the catastrophe. The elections had relatively high turnout (55.31%) and have split the voters almost half in half (53.01% to 46.99%).[49]

High emotional content of pre- and post-election discussions resulted in apparently unbridgeable split of the society at all levels. Voting data of the National Electoral Commission show that there is significant geographical correlation between the voter preferences and city size and industrialization level. In many cases direct contact between the supporters of the two camps is limited. On the other hand the Internet, allows such contact, with little or no limitations on content or attitude.

In the current paper we shall denote the four sets of comments by their time of origin: January 2009 (JAN09, covered in Paper I), second half of July 2010 (JUL10) and two datasets gathered in February 2011 (FEB11 and FEB11Q). The difference between the two latter datasets shall be explained in next section. Our goal is to examine statistical aspects of the discussions not covered in Paper I (such as emotion distribution) but also to look into general changes brought by the passage of time. In this aspect, the JUL10 dataset stands out, as it has been gathered just after the loss of the elections by the PiS candidate, when his supporters were in highly emotional state, some of them denying legitimacy of the voting result, some accusing the president-elect of treason.

II. NEW DATA DESCRIPTION

A. Data sources

As in Paper I, we have gathered our data from discussion fora related to news items published in category Politics by the Internet branch of the largest Polish newspaper, gazeta.pl. While the news source is not neutral (the newspaper has clearly an anti-PiS stance), the discussion participants from both camps use openness of the forum as a convenient ‘battleground’. Posting comments requires a registered user account, and while other users see only the other users’ nicknames, the privacy policy warns users that IP address and other data are logged and may be given to the relevant state agencies in the case of lawbreaking. There is possibility for a single person to register multiple times, but as we have no access to such information we treat each nicknames as a separate user.

In addition to posting comments, registered readers may score the other comments via simple thumbs

up/thumbs down mechanism (such evaluations are part of the JUL10 and FEB11 datasets). The current score is displayed along the post, in boldface, green/red font to increase visibility (see Fig. 1).

Non-neutrality of the forum, resulting from the political sympathies of the newspaper that provides the basic news items, is strengthened by automatic hiding of posts that have strong negative score. There is only ‘comment hidden’ and score visible during normal browsing and the post may be made visible by individual clicking on dedicated link. It should be noted that this visual hiding of the comment is separate action from administrative deleting comments reported as illegal, aggressive etc. The latter option is exercised relatively infrequently (for less than 1% of the posts). In our analysis we have found that the automatically hidden comments are not significantly more abusive than the rest, the distinction is that they represent minority view. As a result, the first impression that viewers of the web page have is of greater uniformity of political opinions, due to combination of administrative mechanisms and particular mix of participants.

B. Effects of user interface changes on communication characteristics

In addition to searching for effects of passage of time, we have identified another significant significant factor that might have influenced the user behavior. At the time of writing of Paper I, the forum interface allowed users to post replies to other users’ comments with a single click. This resulted in the observed high proportion of exchanges of posts between pairs of users. The graphical form of presentation of the discussion threads favored easy recognition of such exchanges by other users, which increased the ratio of comments directed to comments (rather than to the original news stories)

Since mid 2010, the forum has technically split into two branches. The first, www.gazeta.pl (Fig. 1), does not allow direct individual responses to specific posts. Thus, there is no longer a simple mechanism supporting quarrels that we have described. The change has been criticized by many users, some of them calling for boycott of the forum. The critics openly state that they want to interact with the other users. Interestingly, these criticisms come from users from both political camps, and are probably the sole topic on which the two groups agree. Throughout the paper we shall refer to this as the ‘new interface’.

The second, less popular site, forum.gazeta.pl (Fig. 2), has preserved the old capacity of one-click replies (and the quarrels that we have observed in Paper I). Tree-like structure of the posts is prominently visualized (rather than flat time-dependent sequence), so that the most active exchanges between pairs of commentators are immediately visible and draw a lot of attention from other users. We shall refer to it as the ‘old interface’.It is worth noting that the base news stories are essentially the same on both datasets. This situation gives us unique opportunity to study the influence of presentational aspects in Internet communications on the user behavior and expression of emotions. To focus on differences due to the user interface we gathered, in February 2011, two parallel datasets corresponding to new interface: www.gazeta.pl – dataset FEB11 and to the old interface: forum.gazeta.pl – dataset FEB11Q.

Despite the lack of tools facilitating exchanges of posts between users in the new forum, we observed a variety of impromptu editing tricks aimed to indicate that comments are related not to the main story but to other users. Typically, such posts are often started with reference to someone’s nickname (often preceded by @ sign for visibility). Sometimes the reference is hidden inside the comment. In the analyzed sample we had to identify such discussions through human reading. The process could not be automated, as users frequently misspelled other users nicknames. Each of the samples analyzed consisted of several thousand posts in a few tens of threads. To analyze the social networking statistics we have focused on discussions with more than 50 posts, typically in the range of 100–500. Selection of the threads used for analysis was based solely on their size, with no pre-screening of content of comments. We note here that the size of a discussion obviously depends on the ‘hotness’ of the commented topic. Some news stories were in themselves quite provocative (e.g. commenting negatively on prominent PiS politicians) so that one could expect that they would rise a lot of comments. But sometimes long threads resulted from relatively low profile news. The use of only selected, long discussions in our analysis makes direct comparisons of post statistics with data gathered from total records of other Internet discussion fora (Si et al. [15, 16], Ding et al. [17], Ding and Liu [18], Chmiel et al. [19]) impossible. For example, our datasets would have smaller number of posts with no links to other posts simply due to the fact that we avoided news items which generated very small discussions.

In the discussions using the new user interface we were able to identify only a few extended quarrels involving pairs of users, the longest comprising of 4 posts. This is drastically shorter than the exchanges in JAN09 and FEB11Q data, where we observed many exchanges longer than 10 consecutive posts. This is obviously due to technical properties of the new portal which make such exchanges more difficult to maintain – a user has to watch for replies to his/her own comment without graphical guideline to help, which makes responding more difficult. The lack of visibility of quarrels also diminishes interest in joining-in by onlookers. As a result, the new interface promotes many more self-contained comments, which do not relate to other users.

III. NETWORK ANALYSIS

A. User and comment network properties

Within each thread, each discussion forum, we may treat users as forming a directed social network, with links provided by comments directed at other users. Comments attached to source news (and not to other users’ comments) may be considered as marking the presence of active but isolated nodes.

Table I shows, for each dataset, the basic network parameters. The old interface with its one-click reply mechanism facilitated extended network formation. For both JAN09 and FEB11Q datasets the largest connected component comprised of more than 60% of the users. One can think in these cases about forming a percolation network for information travel among the users. And we recall here that the users are largely coming from opposing political camps. Thus the existence of such network shows that conflicted users at least see the arguments and narratives used by their opponents. On the contrary, much weaker network connection of the new forum with majority of users posting comments which do not relate to other users might indicate their focus on their own viewpoint only.

The new user interface has diminished the possibility of user-to-user communication. The majority of comments are now directed at news source. Many users are thus unconnected to others. With a similar number of posts, the largest connected component is almost three times smaller, there are more than 4 times more isolated users and 6 times less links. In Paper I we have identified pairwise exchanges of comments as the driving mechanism of network formation. This is well represented by a large number of pairs of users linked by multiple connections (multi-edge pairs). Between JAN09 and JUL10 the number of such pairs has fallen 8 times. The FEB11Q data preserve some of the highly networked characteristics of the JAN09 dataset, e.g. significant size of the largest connected component, small percentage of isolated users. We attribute smaller number of links to the lesser popularity of the forum, hidden, as we noted, deeply within the newspaper Web site.

Overall, our datasets contained 6404 users, defined as distinct nicknames. Out of these, 5132 were present in only one of the sets, 940 in two of them, 274 in three and 58 users participated in discussions contained in all four datasets. Presence in more than one set of data, which means long term presence in the forum, is correlated with general activity. The average number of posts per user for the small core of 58 users was 27.2, compared to the overall average of 4.03. For the users who were found in at least three datasets, the average number of posts was 16.6. Another way of looking at the extended activity measures is to check the most active users (as given by the overall number of posts). From the top ten, characterized by an average of 152 posts per user, two were present in only one set, two in two datasets, four in three and twoFIG. 1: ‘New’ user interface to news discussion forum, www.gazeta.pl. The interface shows each comment in full and allows one-click evaluation (thumbs-up, thumbs-down), together with the status of previous evaluations. To reply to a specific comment written by someone else, users resort to direct mentioning of the author in the text of the comment. No tree structure is visible. The discussion forum is placed directly beneath the full news item.

in all four.

Figures 4–6 present distributions characteristic network measures of user activity for the four datasets. We focused on user indegree, number of posts written by a user and number of discussions he/she participated in. Most of these distributions are relatively well described by power laws, the only significant deviation is in FEB11 data on number of threads, which shows unusual behavior at low range.

The number of posts written by users and network outdegree should not be confused: as noted in Table I many of the comments do not connect two users, being addressed to the original news article, so they are not counted in the usual network analysis. For the new in-

terface user-to-user comments are only a small part of the total number of posts.

B. User political affiliation statistics

Statistical properties of the comment fora may be analyzed from two points of view: looking at users and at discussion threads.

The first element is the relative number of participants from the opposing camps. As we have already noted, the news source (www.gazeta.pl) could hardly be called neutral. It shows strong pro-PO sympathies, actively participating in election campaign. Not surprisingly, supportersThe screenshot shows a forum interface with the following elements:

• Publicystyka: The site's main category header.
• Polec znajomemu: A button for sharing the article.
• Komentarze do artykułu: A link to view more comments.
• NEWS SUMMARY: A box containing the main article title, "PiS poprosi Europę o pomoc w ratowaniu polskiej demokracji", and a brief description of the article's content.
• COMMENT TEXT: A box containing a user's comment, "PiS poprosi Europę o pomoc w ratowaniu polskiej...", the author "6.smiech", the date "06.02.11, 14:03", and the comment body text.
• REPLY TO COMMENT BUTTON: A button labeled "+ Odpowiedz" for replying to the comment.
• DISCUSSION VIEW FOCUSING ON SOCIAL STRUCTURE: A box containing a threaded list of comments and replies, showing the hierarchical structure of the discussion.

FIG. 2: 'Old' user interface to news discussion forum, forum.gazeta.pl. The interface facilitates replies to individual comments and visualize the tree structure of the discussions. The screenshot presented here highlights an example of a quarrel between two users. On the other hand to read each comment in full the viewer must click on particular link. Discussion forum is placed beneath a summary of the news item, rather than the original story, where only number of comments is provided.

of PO form the majority of readers and commentators. Table II summarizes the ratios of posts and participating users identified as supporters of the two major combatting parties and those with unknown sympathies. In all cases the PO supporters formed roughly 50% or more of the participating users, PiS being a significant minority. This polarization of the forum participation has been observed by the users themselves, and often mentioned in the posts.

Figure 3 presents distribution of sympathies of users for each of the threads in JUL10, FEB11 and FEB11Q. In July 2010, the polarization of the forum was at the highest level in the studied periods, but in all cases there are quite large deviations from the averages in particular

threads. Despite overall PO dominance, in both datasets from February 2011 there are threads where PiS supporters posted a majority of comments.

The UNK category is actually comprised of two distinct groups. The first are the users who openly declare support for one of the remaining parties in Poland or openly against both PO and PiS. The second, roughly the same in size, are those users for whom assignment of political support was impossible to determine from the content of the posts.

User affiliation was determined first within each dataset, and then compared between them. Political affiliation has been remarkably stable for the studied forum. Only fifteen users changed their sympathies, all of

	JAN09	JUL10	FEB11	FEB11Q
Forum interface	OLD	NEW	NEW	OLD
Number of threads	47	27	27	50
Number of posts	7592	7179	6447	4591
Number of users	1613	2752	2187	1527
Number of links between users (percentage of posts)	4754 (62.6%)	770 (10.7%)	1172 (18.2%)	2286 (49.8%)
Largest connected component (percentage of the users)	1106 (68.5%)	440 (16.0%)	589 (26.7%)	947 (62.0%)
Isolated users (percentage of the users)	457 (28.3%)	2156 (78.3%)	1507 (68.9%)	522 (34.2%)
Multi-edge pairs (percentage of the users)	718 (44.5%)	92 (3.3%)	160 (7.3%)	310 (20.3%)

TABLE I: Comparison between JAN09, JUL10, FEB11 and FEB11Q network properties. JAN09 and FEB11Q facilitate user-to-user exchanges, and show much greater proportion of links and pairs of users connected by multiple links. The difference is especially visible in the number of isolated users (i.e. users who posted a comment directed at the main news story, without commenting other users) and in the relative size of the largest connected component of the user network, which reaches over 60% in the old, quarrel-promoting, interface.

	JAN09	JUL10	FEB11	FEB11Q
Forum interface	OLD	NEW	NEW	OLD
PO users	62%	78%	62%	46%
PiS users	22%	15%	20%	29%
UNK users	16%	7%	19%	25%
PO posts	63%	79%	65%	55%
PiS posts	25%	17%	24%	26%
UNK posts	12%	4%	11%	19%

TABLE II: Comparison between JAN09, JUL10, FEB11 and FEB11Q user political affiliations. Small differences between ratios calculated for users and for posts reflect generally lower activity of the users whose affiliation is undefined and higher activity of the committed users. This confirms the hypothesis of the forum being used as ‘battling ground’ between the two opposing parties. July 2010 data were gathered just after the presidential elections won by PO candidate, when his supporters were elated by the outcome.

these changes happening between July 2010 and February 2011. In one case, the change was from PiS to PO support, in another from UNK to PiS. Thirteen PO supporters, in all cases dissatisfied with perceived lack of activity by the PO government, changed their affiliation, twelve of them switching to support to newly formed political parties, and only one declaring support for PiS. Taking into account that the number of changes is only 0.23% of the total number of participants, we may con-

clude that participation in political discussions did not encourage change political opinions and support in general sense.

IV. DISCUSSIONS CONTENT

One of our goals was to observe if there are characteristic features of the content of posts that persist/changeFIG. 3: Distribution of user affiliations for threads in JUL10, FEB11 and FEB11Q datasets. In July 2010 dominance of PO supporters was very high in all studied discussion threads. In February 2011 we observe some threads where there is similar dominance, but also threads where supporters of the two parties are almost in equal numbers or even when PiS supporters dominate. The latter situation is especially interesting, bearing in mind strong pro-PO stance of the newspaper.

with the passage of time. As in Paper I, we have divided the posts into several categories:

Agr - comment agrees with the covered material (either the original news coverage or the preceding comment in a thread);

Dis - comment disagrees with the covered material;

Inv - comment is a direct invective and personal abuse of the previous commentator;

Prv - provocation - comment is aimed at causing dissent, often, but not always only weakly related to the topic of discussion;

Neu - comment is neutral in nature, neither in obvious agreement or disagreement;

Jst - ‘just stupid’ comment, which is unrelated to the topic of discussion, but without malicious intent;

Swi - comment signifying a switch in participant’s position leading to agreement between two previously opposing commentators.

Table III summarizes the distribution of various types of comments. Firstly we note relatively stable ratios of positive (Agr), negative (Dis, Inv, Prv) and neutral comments. The average values for all datasets are 17%/70%/12.5%, clearly indicating that majority of comments are confrontational in nature. There is significantly lower number of posts classified as neutral in JAN09 data. We observe an increase of neutral comments as we move to later datasets. Another observation is that the old interface discussions contain more agreeing

comments than the new interface sets, a finding which we shall discuss in more depth later on.

Looking at the distribution of comments linking within and between supporters of both parties, we observe that the largest number of links are between the groups supporting conflicted parties (inter-faction), rather than within each group (intra-faction). This stands in contrast with observations of Adamic and Glance [24], Hargittai et al. [26] who analyzed links between political blogs in the United States. The main difference is that blog entries are, at least in principle, deliberative in nature. The process of creation of the network is also radically different. The links between blogs are inserted by the authors as integral part of the blog text; to support the presented point of view. Moreover, many blog entries combine multiple topics, data sources and links. In contrast, quick comments in discussion fora are usually focused on one topic (either the original source or some previous comment) and are rather reactive. The links are usually to a single comment only, with very few exceptions. It is easier (and more rewarding emotionally), in a short time and space, to attack an opponent than to construct elaborate presentation of one’s own position by linking with other supporters. In fact, in research focused on blog content rather than on simple presence of direct links, even this more deliberative medium shows dominance of attack approach. As Trammell [25] noted in the analysis of blogs during the US 2004 elections, more than a half of the blogs discussed the opponent, and out of these, almost 80% contained an attack.

To analyze this issue further we have calculated the ratio of observed links in each category (inter-faction, intra-faction, to and from users of unknown sympathies) to the number of links expected if one assumes no preferences inFIG. 4: Comparison of indegree of users in various datasets (clockwise from bottom left: JAN09, JULY10, FEB11 and FEB11Q). Filled circles are normally binned observation data. Open circles are averages from 20 runs of simulations.

commenting. Such numbers would be given by the appropriate ratios of PO and PiS supporters in each dataset. The resulting ratio, named $\alpha$ factor, is presented in the last column in Table III. In all datasets, we observe that intra-faction $\alpha$ is less than 1, which means that there is much less motivation to post comments addressed to one's own group members. Conversely, the inter-faction $\alpha$ is always greater than 1, reflecting the willingness to start/continue discussions with the opponents. We note that in the most heated debate, JUL10, which took place just after the presidential elections, the inter-faction $\alpha$ is greater than 3 and for intra-faction it is equal to 0.36. This documents very high disassortativeness shown by the commentators who took the trouble to address their

posts to other users, despite the lack of the easy tools in the new interface.

Interestingly, the less popular, old interface discussions in FEB11Q show significant difference from the other sets. There is much smaller number of abusive comments (Inv). Also, $\alpha$ factors are much closer to 1, indicating less preferential linking to political opponents. As FEB11Q dataset shares a lot of network properties with JAN09, and, at the same time, it is based on the same newspaper content as the FEB11 dataset, this diminished assortativity is probably due to changes in the interface and change in general user attitudes between 2009 and 2011. We will return to these issues in the conclusions of the paper.FIG. 5: User activity measured by number of posts written by the user (clockwise from bottom left: JAN09, JULY10, FEB11 and FEB11Q). Filled circles are normally binned observation data. Open circles are averages from 20 runs of simulations.

V. ANALYSIS OF EMOTIONS

A. Human analysis of emotions and other users' evaluations

In Paper I our analysis focused on the content of the user comments, classifying them mainly according to the expressed views and opinions as presented in the previous section. Here we are expanding the analysis by looking at emotions expressed by user comments. Such emotions have been recently the subject of intensive studies (e.g. Schweitzer and Garcia [14], Chmiel et al. [19], Prabowo and Thelwall [30], Mitrović et al. [31]), aimed at quanti-

tative description of emotional motivations. While there is a rough mapping between the goals of the comments (informative, provocative, quarrelsome) and the associated emotions, we note that these categories are not completely the same. Posts categorized as agreements may be expressed in highly emotional fashion – or stated neutrally. Similarly, disagreement may be stated in with or without agitation. Even invectives may be expressed through vulgar, impolite language or through sarcasm, in a cold and calculated manner. The two directions of analysis: content and emotion are thus complementary.

The method used in this work is an extension of the approach used in Chmiel et al. [19]. Instead of simple +1, 0, -1 scale of emotions indicating positive, neutralFIG. 6: User activity measured by number threads that users participate in (clockwise from bottom left: JAN09, JULY10, FEB11 and FEB11Q). Filled circles are normally binned observation data. Open circles are averages from 20 runs of simulations. The origin of the deviation of FEB11 data from power law is unknown.

and negative emotional expressions, we use a graded scale from +1 to -3, described below:

• +1 Positive emotions, expressed as support for another post or statement by politician described in the news item.
• 0 Neutral emotions: statements of facts (either agreeing or disagreeing with the target of the post), explanations; worded in neutral language.
• -1 Light negative emotions: expressed via single instances of vulgar language, suggestions that politicians or other users are idiots, etc.

-2 Strong negative emotions: repeated use of vulgarities, comparing politicians or other users to Nazi/Stalinist personages, excessive use of capitalized letters etc.

-3 Excessive negative emotions: posts combining many elements described above.

The categorization of emotions in each post is done by human reading. Emotions are assigned to two categories: related to general political issues and politicians (e.g. ‘I hate politician X’) or directed at other users of the forum (e.g. ‘You are an idiot’). The total emotion expressed by a post is then calculated as sum of the two constituents.

Connection type	Agr	Dis	Inv	Prv	Neu	Jst	Swi	Subtotal	$\backslash alpha$ factor
JAN09
Intra-faction (PO-PO, PiS-PiS)	16.9%	2.1%	2.1%	1.1%	1.1%	0.4%	0.1%	21.9%	0.51
Inter-faction (PO-PiS, PiS-PO)	0.6%	32.8%	17.1%	2.7%	0.8%	0.0%	0.0%	54.0%	2.00
Factions-UNK (PO/PiS-UNK)	2.5%	11.1%	3.1%	1.9%	1.7%	0.4%	0.2%	20.9%	1.52
Comments by UNK	0.6%	1.2%	0.7%	0.2%	0.2%	0.1%	0.1%	3.1%	0.19
Subtotal	20.6%	47.2%	21.1%	5.9%	3.8%	0.9%	0.4%
JUL10
Intra-faction (PO-PO, PiS-PiS)	12.4%	1.4%	1.0%	1.8%	5.4%	0.3%	0.0%	22.5%	0.36
Inter-faction (PO-PiS, PiS-PO)	1.1%	28.1%	25.8%	8.2%	8.8%	0.0%	0.0%	72.0%	3.07
Factions-UNK (PO/PiS-UNK)	0.8%	0.9%	0.8%	0.4%	0.9%	0.0%	0.0%	3.8%	0.59
Comments by UNK	0.2%	0.4%	0.1%	0.0%	0.9%	0.0%	0.0%	1.7%	0.19
Subtotal	14.7%	30.9%	27.7%	10.5%	16.0%	0.3%	0.0%
FEB11
Intra-faction (PO-PO, PiS-PiS)	9.8%	1.7%	1.7%	2.1%	6.4%	0.0%	0.0%	21.8%	0.52
Inter-faction (PO-PiS, PiS-PO)	0.8%	26.1%	18.8%	7.3%	6.5%	0.0%	0.0%	59.6%	2.44
Factions-UNK (PO/PiS-UNK)	1.1%	3.5%	1.6%	1.1%	1.5%	0.0%	0.0%	8.9%	0.59
Comments by UNK	0.5%	3.8%	1.9%	0.7%	2.7%	0.2%	0.0%	9.8%	0.53
Subtotal	12.4%	35.2%	24.0%	11.2%	17.0%	0.2%	0.0%
FEB11Q
Intra-faction (PO-PO, PiS-PiS)	15.6%	2.2%	0.8%	3.8%	6.5%	0.4%	0.0%	29.2%	0.99
Inter-faction (PO-PiS, PiS-PO)	0.3%	18.9%	11.7%	3.9%	2.4%	0.0%	0.0%	37.2%	1.41
Factions-UNK (PO/PiS-UNK)	2.3%	5.7%	2.0%	1.2%	2.6%	0.3%	0.0%	14.0%	0.74
Comments by UNK	3.3%	8.4%	1.7%	1.9%	4.1%	0.2%	0.0%	19.6%	0.77
Subtotal	21.4%	35.2%	16.2%	10.8%	15.6%	0.9%	0.0%

TABLE III: Statistics of comment type between various groups of users for the studied datasets (two identified factions and neutral or unidentifiable class UNK). Only comments linked to other comments are classified. The $\alpha$ factor denotes the ratio of observed number of comments linking within faction group, between groups and to/from agents with unknown affiliation to values expected from the user affiliations if the posts were placed randomly. For example, for inter-faction comments we observe $\alpha$ value higher than 1, which means that users prefer to address the supporters of the opposite faction. Value smaller than 1 corresponds to lower preference for placing a comment.

Figure 7 presents distribution of emotions expressed in posts JUL10, FEB11 and FRB11Q datasets. It is worth to note much larger ratios of positive comments in threads 3, 22 and 25–27 in JUL10 set. In all these cases they were concerned with news items regarding comments strongly against PiS politicians made by perceived ‘outsiders’ of the political field. In one case (thread 22) this was a comment by ex-prime minister from Social-Democrat Left Alliance party. In four other cases, the positive response was directed at an ‘enfant terrible’ of PO, who some time later decided to form his own political party. Positive emotions were mostly of the form of personal support for courage and decisiveness of single persons, perceived as acting outside political establishment. In FEB11 dataset there is only one discussion with high ratio of positive emotions. In thread 26, positive emotion has been generated by relatively large number of PiS supporters, expressing the admiration for statements by PiS politician accusing the PO government of treason. A reverse situation is present in threads 27 and 32 of FEB11Q set: here the PO supporters express their

positive emotions at statements by the prime minister and by a popular sportsman.

B. Automatic emotion recognition

In addition to assignment of emotions to comments based on human reading, we have constricted simple analysis engine to detect emotions from the texts. Engine calculates emotion based of content of the post, length of the post and average of previous emotion of posts made by user.

Word list is built using two separate sources. The first contains popular polish swearwords. The second is built from words commonly used in community to make fun and/or irritate other users. Words in the second list are taken from subset of dataset (three randomly selected threads, about 600 posts). They include intentional misspellings of names and word-games on political party names, turning them into near-swearwords. Each word has its own emotion power property $P_n$ , where $P_n$ is in-FIG. 7: Distribution of emotions within threads chosen for analysis in JUL10 and FEB11 datasets. Figure shows percentages of posts within each thread with total emotions ranging from +1 to -4 (the latter is possible if one post expresses negative emotions against both politicians and forum users). Darkening shades of gray represent emotions from +1 to -4, indicated also for clarity in the JUL10 dataset.

teger ranging from 1+ (positive) to -3 (very negative), assigned before the evaluation process.

Emotion detection algorithm works as follows:

• • it reads a comment, divides it into words and counts them, and remembers count of the words as $L$ ;

• • it searches for words from the list, each hit adds $0.8P_n$ to post score $S$ . We also store total hits count $H$ ;

• • average score $U_i$ for the author of the post is calculated as arithmetic mean of her/his previous average $U_{i-1}$ and current score of the post $S$ ;

• • emotion value of the post is calculated from score through $E = S (\ln (25H/L + 2))^2 - U_i/5$ , this form of expression was chosen to arrive at best reproduction of the results for the three threads used in training of the system;

• • to allow for situation where a single highly negative word is used in a very long post (which ‘dilutes’ the emotional impact of the post) or when previous posts of the user were overwhelmingly negative, the negative emotion is reduced by looking at ratio of the words from the list found in comment to its length $H_p = H/L$ . For $H_p < 0.09$ if negative emotion was lower than -1 it is reset to $E = -1$ , for $H_p < 0.20$ negative emotions lower than -2 are reset to $E = -2$ .

• • all the other results of $E$ are rounded to the nearest integer value.

Human and automatic emotion recognition values were compared via average dataset emotion value and through distributions of posts with total emotion equal to +1, 0, -1, ..., presented in Table IV. The automatic algorithm, although ‘trained’ only on a very small subset of our data, gives reasonable results when compared to human analysis.

The quality of the automatic procedure may also be measured by correlations between automatic and human assignment of emotions to all posts in a forum and for thread averages. For the JUL10 dataset, correlation coefficients between human and automatic assignment of emotions were calculated for the whole set of posts ( $C_p^{JUL10} = 0.18$ ) and for threads ( $C_T^{JUL10} = 0.60$ ). For the FEB11 dataset the correlation coefficients were ( $C_p^{FEB11} = 0.26$ ) and for threads ( $C_T^{FEB11Q} = 0.54$ ), while for FEB11Q the respective values are $C_p^{FEB11} = 0.20$ and $C_T^{FEB11Q} = 0.57$ .

VI. READER EVALUATION OF POSTS: TRACING THE INVISIBLE USERS

The users of Internet fora are not limited to active authors of comments. There are those who read the messages but do not post comments. Internet slang calls

Dataset	Average emotion	Ratio of comments with emotion of
		+1	0	-1	-2	< -3
JUL10, human analysis	-0.539	0.059	0.497	0.321	0.099	0.023
JUL10, automatic analysis	-0.530	0.039	0.503	0.345	0.090	0.023
JUL10, simulations average	-0.485	0.009	0.423	0.410	0.076	0.003
FEB11, human analysis	-0.460	0.017	0.585	0.330	0.060	0.010
FEB11, automatic analysis	-0.490	0.026	0.557	0.312	0.088	0.017
FEB11, simulations average	-0.383	0.045	0.544	0.392	0.017	0.000
FEB11Q, human analysis	-0.338	0.046	0.630	0.272	0.045	0.006
FEB11Q, automatic analysis	-0.390	0.041	0.594	0.309	0.043	0.011
FEB11Q, simulations average	-0.351	0.048	0.564	0.375	0.012	0.000

TABLE IV: Comparison between automatic and human detection of emotions for various datasets. The parameters of the automated analysis algorithm as described in Section VB have been tuned to give best fit to just three threads in JUL10 dataset. Despite this, the results for both the average emotion and for the assignment of emotions of posts into different levels of emotion seem quite good.

these users ‘lurkers’, and usually there is little that we can do to study the statistical properties of this group.

Fortunately, in our case, in addition to post factum study of the entire dataset, we may use data on how forum readers evaluated the posts shortly after they were written. This feature is present in the ‘new interface’ datasets (JUL10, FEB11) thanks to thumbs-up/thumbs-down buttons. This provides a measure of reactions of readers to the views presented by other users. Similar measures are available at other sites, for example as tools in establishing usefulness of user advice in self-help fora or in review discussion boards. In the context of general discussions reader evaluations were analyzed by Gómez et al. [8] and Gonzalez-Bailon et al. [32] for slashdot; by Hsu et al. [33], Jamali and Rangwala [34], Khabiri et al. [35] for digg, and by Lange [36] for YouTube.

We used two numbers for each post: the total number of evaluations it has received, $N_P$ , and resulting reader opinion (difference between thumbs-up and thumbs-down votes), $O_P$ . The first value measures the interest that given post has received from reader community. The second is result of an interplay between post political positioning and distribution of political sympathies of reader community. Analysis of the correlations between the two variables may provide some information regarding the ‘invisible’ part of the forum users: readers who do not post. On average, in JUL10 dataset each post has received 10.8 evaluations; in FEB11 the number is smaller, 7.6. The fact that there are many more thumbs-up/thumbs-down clicks then posts agrees with an expectation that the number of lurkers is greater than the number of writers.

Plotting the number of evaluations $N_P$ vs. resulting opinion $O_P$ (Fig. 8) shows interesting behavior. The distribution splits into two separate ‘wings’ of positive and negative opinions scaling approximately as linear functions of the number of evaluations. The absolute value of coefficients are somewhat below 1. The number of

posts in the positive ‘wing’ is greater than in the negative one, and, at least for the posts with highest numbers of evaluations, negatively judged ones are written by PiS supporters, whereas the positively evaluated ones by pro-PO users. For posts which received many evaluations, one can approximate $O_P \approx \pm 0.8 N_P$ for PO and PiS supporting posts respectively (indicated as black lines in Fig. 8) Using a simple assumption that evaluations are done by the supporters of the two conflicted camps, and that supporter of the post political view would always give positive opinion and vice versa, the PO supporter reader ratio would be at 90%, higher than the observed writer ratio.

We have also checked whether the number of evaluations received depends on the post/user political views. Left panels of figures 9 and 10 present normalized histograms of the average number of evaluations per post for users supporting PO, PiS and of unknown sympathies. In each case these distributions are well described by $ax \exp(-\lambda x)$ functions, similar for each political group. On the other hand, the histogram of opinions (right panels of the figures 9 and 10) shows radical differences between supporters of different parties. PiS supporters, who are minority in the forum receive very few positive opinions and distribution increases towards totally negative -1 opinion/number of evaluations ratio. Opposite phenomenon is seen for PO supporters. Unknown users are as likely to get positive as negative opinion.

VII. AGENT BASED MODELING OF THE USER ACTIVITIES

In Paper I we have presented a computer model of participation in political discussion fora. The model was further developed and applied to BBC discussion fora in Chmiel et al. [19]. We have used the same model framework for the four datasets studied here. Thanks to fullFIG. 8: Evaluations of user posts in JUL10 (left) and FEB11 (right) datasets. All points correspond to single posts. Different markers indicate posts by supporters of political parties (light gray circles majority PO users, black squares minority PiS, triangles – unknown affiliation). Gray lines show boundaries of fully negative/positive evaluations. Black lines are guide for the eyes, with $O_p = \pm 0.8 N_p$ . Assuming that only politically committed readers would give the evaluation, and that PO supporter would always approve pro-PO post and always disapprove pro-PiS post, and vice versa, the lines correspond to 90% of the evaluating readers being pro-PO, much higher proportion than for the comment writing forum users.

scope human analysis of the posts some of the parameters for the model, which were guessed or fitted in previous studies could be input from observations or used to compare the simulations to real world, for example the ratio of proponents of the two competing parties and of the neutrals/unknown affiliation users. The model uses, as input parameters, probabilities of posting a comment, addressing it to source or to other post, probability of response within a pairwise exchange between users (depending on whether it is written by a user sharing political sympathies or not). Because simulations use probabilistic distributions and give slightly different results at each run, we present results as averages of 20 such runs for each dataset. Table V presents average number of users, threads above length threshold, posts and links, as

well as the distributions of connections between different user groups.

The network properties (in-degree, distribution of the number of written posts and number of threads a user participates in) of simulated discussions are compared with observations in figures 4–6. With the exception of the distribution of the number of threads a user participates in for the FEB11 set, simulations reasonably reproduce multiple system characteristics, such as observed distributions reasonably well.

We have extended and changed the model used in Chmiel et al. [19] with respect to the evaluation of emotions. Previous approach assumed a fixed reaction of an author in response to political stance of the author of commented post: the resulting emotion was +1 if theirFIG. 9: Statistics of user evaluations as function of their affiliation in JUL10 dataset. Left panel: normalized histograms of number of evaluations received by a user, for the three classes of users (PO supporters: light gray circles, PiS supporters: black squares, unknown: triangles). Lines are fits with $ax \exp(-\lambda x)$ , similar results obtained for each user group. Right panel: histogram of average evaluation of posts written by a user, the same color scheme applies.

FIG. 10: Statistics of user evaluations as function of their affiliation in FEB11 dataset. Left panel: normalized histograms of number of evaluations received by a user, for the three classes of users (PO supporters: light gray circles, PiS supporters: black squares, unknown affiliation – triangles). Lines are fits with $ax \exp(-\lambda x)$ , similar results obtained for each user group. Right panel: histogram of average evaluation of posts written by a user, the same color scheme applies.

affiliations were the same, $-1$ if they belonged to opposite camps and either $+1$ or $0$ if a committed agent commented a message written by a neutral (neutral agent's comments were assumed to be always neutral).

Thanks to a deeper human analysis of our data we were able to propose a modification of this approach.

The reaction of an agent to a post is in simulations is taken as the nearest integer from Gaussian distribution with center and width determined by comparing the political views of the two agents. This model is based on observed the statistics of posts, where data for various author-target affiliation combinations were fitted to nor-

	JAN09	JUL10	FEB11	FEB11Q
Number of threads	45	28	30	55
Number of posts	7315	7073	6485	4560
Number of users	1557	2691	2168	1627
Number of links between users (as percentage of posts)	4576 (62.5%)	748 (10.5%)	1148 (17.7%)	2299 (50.4%)
Intra faction links percentage	22%	23%	27%	25%
Inter faction links percentage	53%	68%	45%	37%
Factions to UNK links percentage	12%	5%	13%	17%
Comments by UNK percentage	13%	4%	15%	21%

TABLE V: Main results of simulations of discussions (averages over 20 simulations of each dataset). For numbers of threads, users, posts and links, compare with Table 1, for the distribution of connections between supporters of political parties, compare with column Subtotal in Table 3.

mal distributions. As it turned out, the distributions could be divided into three groups, differing by position of the center and width of the distribution. The first corresponds to similarity of views: PO-PO, PiS-PiS, PO-Source (we assume all Source messages are pro-PO). The second is when the author and target "disagree" (PO-PiS, PiS-PO, PO-UNK, PiS-UNK). The third, "neutral" distribution was observed for posts written by authors with no identified affiliation, regardless of the target.

For the purposes of the simulations, small differences within each group were looked over and post emotion was drawn from normal distribution generalized for the appropriate author-target class. The values were those taken from observations for each of the datasets separately. For example, emotion in similarity case for JUL10 data was centered at $-0.4$ , for disagreement at $-0.8$ and for comments by unknown agents at $-0.3$ ; in all cases the width of distributions was assumed, $\sigma = 0.7$ . Due to rapid decrease of the normal distribution, the resulting distributions do not agree very well for extremely negative comments (with total emotions of $-3$ and $-4$ ) but give rather good values for the less emotional posts statistics.

VIII. DISCUSSION AND CONCLUSIONS

Looking at temporal changes in the observed discussion properties we note that the JUL10 data has the highest negative emotion content. As noted before, this dataset contains discussions which took place when the entire public debate was at its hottest, shortly after PiS candidate lost his presidential bid. During the election campaign, the candidate assumed quiet style, suitable for

mourning and reconciliatory stance. Both himself and his campaign team refrained from outright attacks on the opponent. After the lost election this stance changed radically, and politicians from PiS started immediate, vehement attacks on the president-elect and on PO government, accusing them of causing the plane crash or suggesting assassination. The tone of political debate has been immediately picked up by PiS supporters, including the forum users. At the same time, PO supporters were clearly triumphant after the win. This resulted in PO dominance in the dataset. Moreover, during the same period the polarization of the whole society led to increased separation of 'virtual' agoras for the supporters of PO and PiS. To check this we have briefly analyzed July 2010 discussions related to a well known pro-PiS blog site (kataryna.blox.pl). In these discussions PiS supporters are majority, and we found them to be characterized by mirror image network and emotion statistics.

Without communication the two communities of supporters of PO and PiS diverged in worldview, basing their opinions of totally separate sets of 'factual' data and authorities. The process is observed especially in the context of the Smolensk plane crash, where the pro-PiS readers have gathered around several Web pages containing claims supporting various conspiracy theories. These Web sites bear no relation to the official investigation and are cross-linked extensively among themselves, much like 9/11 conspiracy sites in the US. As a result the reader focused on such subset of sources may be under impression that it presents a coherent and complete view, because there is no reference to contrary data and explanations. Armed with one-sided information, the readers are strengthened in their political opinions and less ready to participate in discussions with opponents, and even lessto work for a consensus.

Such behavior has been studied in laboratory environment by Cohen [37]. He has noted that groups not only affect attitudes on key subject but may ‘define the very meaning of the objects in real world’. Opinions of trusted in-group members are accepted without deliberation, regardless of the merit of the content.

Discussing polarization and fragmentation of society, Stroud [38] has noted that such selective exposure may have two effects on the population. First, using only the information provided by one side would limit a persons ability to deliberate and choose. Second, such separation might lead to lesser tolerance and extremism. The existence of Internet discussion fora where conflicted sides are, at least, trying to communicate may be a way to combat such extremism. Allowing the users to interact with each other, even if this interaction is, in many cases, in the form of abuse and invectives may decrease some of the negative emotions. This supports the hypothesis put forward by Mutz and Martin [39] that media would surpass face-to-face communications across political divides. The discussions provide much needed exposure to dissimilar views and they complement the traditional media in this eye-opening role. We must, however, remember that simple exposure to the opposite views is no recipe for open-mindedness. Experimental study by Lord et al. [40] has clearly shown that opinions held before being exposed to other viewpoints strongly bias the reception. This bias covers not only debatable opinions but also acceptance of hard, empirical facts. People would choose only the evidence that supports their views. Thus, to achieve effective communication between conflicted groups, more is needed than just combining opposing messages in one place.

One of the main observations in this study is that the change of the user interface has led to drastically diminished number of inter-user exchanges. As most of these (in the old forum) have been in the form of quarrels between supporters of opposing political views, one may be tempted to claim that the interface change has led to lessening of direct confrontations and, presumably, level of conflict. This hypothesis is not confirmed by comparison of February 2011 data on discussions in two parallel fora with different interfaces. In fact, the ‘quarrel promoting’ old interface FEB11Q dataset shows much less emotion and more even distribution of links, indicated lesser recognition of who is it worth communicating with.

There may be several reasons for the less aggressive characteristics of FEB11Q discussions. The first is smaller role of the trolls. These are usually defined as users who post many comments aimed at provoking fights. Typical advice for discussion board users is ‘do not feed the troll’, and if followed by most participants this leaves trolls with significantly smaller indegree than the number of posts. There is no formal numerical threshold for ‘trollness’, but we may define as a troll any user who posted more than 15 comments in the analyzed discussion set, with average emotion less than $-0.3$ and inde-

gree smaller than $1/2$ of the posted comments. Under such conditions FEB11 dataset contained 23 trolls, and FEB11Q only seven. If the threshold is more stringent, average emotion less than $-0.4$ and indegree smaller than $1/3$ of posts then FEB11 set includes 13 trolls, while FEB11Q none.

The second difference comes from the opposite end of the user spectrum: one time writers, who direct their posts at the source news message. The average emotion of such users in FEB11 set is $-0.45$ , while for the FEB11Q set is only $-0.35$ . One of the sources of the difference may be the anticipation of the scoring mechanism, present in FEB11 forum. Getting a large score is a goal for many users – this is clear from the texts of the comments. It is less important whether the score is positive or negative, what counts is being noticed. This quite naturally motivates the users to more extreme positions in the forum where the evaluation mechanisms are present. The reward of watching one’s own comment get noticed and evaluated is complemented by the possibility of making negatively evaluated post literally vanish from the forum, if the number of negative evaluations surpasses certain threshold.

The two fora differ also in the ways they promote reactions to agreeable/disagreeable post. The new interface makes it easy to use the evaluation buttons. Pushing them does not really require thorough reading of the post, much less formulating any response. All it requires is a recognition of ‘a friend’ or ‘an enemy’. In contrast, the old interface promotes exchanges of posts. Even though many of these consist of invectives, many users are forced to actually state what is it that they disagree with. Sometimes even explaining in detail why they disagree. Quite often flame and abuse on one side are met with calm and deliberation on the other side. While we have not observed bona fide opinion changes (comments of the Swi category), there were numerous occasions where discussion evolved from accusations and invectives to explanations and arguments. This would not be possible if the only tool at the disposal of the reader is a simple approve/disapprove button.

Slater [41] has postulated existence of a spiral of selective media use, leading to polarization of beliefs, actions and attitudes, which again influences media use to become more selective. This process is clearly visible in Polish society, moreover, it has reached another level of positive feedback. The whole landscape of the media has changed, adjusting to polarization of the population. There is a clear division of pro-PO and pro-PiS newspapers, weeklies and TV broadcasters. The polarization of media is so extreme that one of nation-wide TV stations has never invited any PO politician (even during four years PO has been in government) while prominently featuring PiS activists. The same split applies to journalists: lacking any pretense of neutrality and impartiality they have grouped themselves in journals that are clearly fighting on both sides of the political barricade. The positive feedback driving this change is not only via personalbeliefs of the journalists. Polarized media sells very well: a new weekly publishing only anti-PO texts has catapulted to second place in circulation numbers just after 3 weeks after introduction.

The descriptions of reality that are offered by so polarized media and selected by the users are impossible to combine into one coherent worldview. One side treats the plane crash as tragic accident, the other as a terrorist attack in which the government played an active role. For one side Poland is economically healthy and fast growing country, for the other it is on the verge of economical catastrophe or even already collapsed. Such divisions are so deep that they make communication between supporters of the two camps – especially face-to-face – extremely difficult, separating the society. It is interesting to note that in-group influence goes beyond simple conformism, but leads also to deep changes in memory, as shown by Meade and Roediger [42], Loftus [43], Wright et al. [44], Edelson et al. [45]. Such effects would influence the future opinions of the people who see only uniform or mostly uniform opinions within a political group. Thus, the effect of group pressure would have long-term effects beyond those characterized by Cohen [37]. Promoting exposure to contrary opinions, even if it does not lead to true value judgments and opinion/attitude change, decreases chances of such socially induced distortion of perception and memory, simply by observing that there are people with different views.

Participants of the discussion fora are, in this environment, rather special: they want to express their views in a way that reaches the other side. More: they want to communicate to the other side. We have already noted that the change of the user interface, making direct responses difficult, has been heavily criticized by many users from all political camps. In many cases, the users themselves emphatically noted that the negative reaction to the change in the user interface is ‘the only topic on which they agree’, which is confirmed largely by our analysis. More than half of the inter-group Agr type comments in JUL10 dataset (shortly after the new forum interface has been introduced) are agreements on this particular topic, brought spontaneously into otherwise political discussions. Forum administrators seem to become a common enemy for both political camps, because they deny the right to open (though bloodless) fight. This may be compared to behavior of football hooligans, who fight each other, but unite when confronted by police trying to separate them. But perhaps there is more positive view of this unity – the need to talk.

The new forum interface with its diminished capacity for pairwise exchanges is dominated by ‘one shot’ comments, which are not addressed to a concrete user but rather express writer’s view to world at large. To gain attention these comments are usually more provocative and do not have to refer to opposite views. In contrast, when a user posts non-injective reply to another post, he or she has, at least, to read it. Experimental study of Palmer [46] has shown that while simple exposure to

negative argumentation leads to more extreme positions, this effect may be moderated by actual evaluation of messages.

In a large scale meta analysis Pettigrew and Tropp [47] have identified three ways of reducing inter-group prejudice: increasing the knowledge about the other group, reducing anxiety related to contact and increasing empathy. The study shows that increased knowledge has the weakest effect in diminishing prejudice. On the other hand, Internet fora are not good environments to promote empathy: anonymity of the contacts prevents it. This leaves anxiety reduction resulting from intergroup contacts as a possible way of improving communication across political divisions. In discussion fora such contacts correspond to pairwise exchanges of posts, when users ‘talk’ to each other, even hidden behind the anonymity of the forum. A reply to a post is treated as personal communication, not as general statement.

MacKuen et al. [48] introduce an interesting concept of two idealized types of participants in political debates: deliberative citizen, who considers all arguments, including these opposite to his views and passionate supporter of one view, the partisan combatant. In real situations people behave somewhere between these two extremes. MacKuen et al. [48] argue that it is emotions that define the deliberative vs. combative stance. Moreover, effects of emotions would differ if the source is aversion due to negative feelings such as fear of the other group or anxiety, due to lack of knowledge or new situation. Aversion strengthens reliance on the views already held, while anxiety encourages seeking out more information. Experimental study has shown that when the situation is seen as familiar threats (leading to aversion) possibility of a compromise decreases dramatically. When the situation is treated as novel and challenging people are more open-minded and allow compromise to form.

With this in mind we might look at the discussion fora with more hope than fear. Despite the generally negative emotions, the discussions provide a way of exposing oneself to the views of others, and also require participants to formulate their positions in communicable way. Especially when two users engage in a discussion using explanations rather than invectives. Based on the results we suggest that to achieve better results in promoting understanding and bridging the polarization gap the technology used should facilitate user-to-user exchanges as much as possible. Instead of anonymous like/dislike polls, the interface should emphasize the dialog and also visibility of such dialogs to other users. It is important that onlookers would see that it is possible to communicate across political division. Of course, large part of these messages would be, as we have shown provocative or abusive. But the comparison of FEB11 and FEB11Q data shows that bigger fraction of person-to-person exchanges diminishes the negative emotional and confrontational status. Properly designed Internet discussion fora might become one of the forces that pull people together, despite their differences, towards democratically desir-able goal of participation and communication (Stroud [38]), even if they remain fixed in their opinions. Taking into account observations of Pettigrew and Tropp [47] and MacKuen et al. [48] one might postulate a change in the role of moderators. Instead of simply hiding or deleting posts that are considered to break the rules of communication, the moderators might visually promote discussions that are based on novel argumentation and informative exchanges rather than those using invectives and provocations. This would diminish the visibility of trolls and provocateurs, and ‘set the tone’ of the forum, without actually banning the extremist views. As we argue here, visual presentation plays important role in shaping the reactions of the readers.

Another argument for facilitation of user-to-user comments has been proposed by Kelly et al. [3]. He argues that extremist users who have nothing interesting to state are actually avoided by other users, in a form of self-moderating mechanism. Such users, called by Kelly ‘fringe’, showing some of the troll-like characteristics, were identified by specific network linking patterns by analyzing Usenet discussions. The analysis of their contributions has shown that they were, indeed, social undesirables, whose contributions consisted of hatred and racism.

Comparing Kelly’s observations with our data, we did not observe ‘fringe’ users, their role taken by trolls (who, however, has much greater acceptance level within their own camp). Importantly, only fora with easy user-to-user communication allow the user community to isolate undesirables. Without such communications all users are left with the option of writing to general audience, without the feedback of responses. And to get the responses, posts have to be interesting (even if abusive or provocative) which promotes deliberation. This is exactly the optimistic conclusion of Kelly et al. [3].

Lastly, we must comment on a particular weakness of our work that has surfaced recently. Throughout the study we have assumed that all users participated in the discussions because they wanted to freely express their views and reactions to other users. In May 2011, one of newspapers has announced[50] that Polish state prosecutor is conducting an investigation covering the use of paid services, consisting of posting inflammatory comments on discussion boards, by political parties and associated institutions. The phenomenon, was reported to cover many Web sites, and has supposedly become ‘a regular industry’, with rates of 0.1-0.5 USD per posted comment. Indeed, posts containing accusations of being ‘on the payroll’ of PO or PiS are quite frequent in the studied datasets, either aimed at particular users on in general, but there were no proofs. Should the investigation prove that the phenomenon is truly widespread, we would face the situation where some of the conclusions (although not the raw observations) presented in this paper would have to be reconsidered. At least part of the emotions expressed by the comments would be faked, scripted according to some Public Relations instructions prepared by political parties. Such comments would rightly deserve being in the ‘provocative’ category. Some of the users we have treated as separate (and included in statistical measures) might turn out to be multiple avatars of a smaller number of ‘professional’ participants.

In such a case our observations would have to be reinterpreted. The paradigm of free, even anarchic, exchange of views between individual members of society, would have to be replaced by a game between the ‘professionals’ (representing PR branches of the political parties) and the rest of society interested in politics. While interesting from research point of view, this picture is a very disquieting one.

---

• [1] T. Mullen and R. Malouf, in AAAI symposium on computational approaches to analysing weblogs (AAAI-CAAW) (2006), pp. 159–162.

• [2] J. Kelly, D. Fisher, and M. Smith, in Online Deliberation Conference (Stanford University, 2005).

• [3] J. Kelly, D. Fisher, and M. Smith, in Online Deliberation: Design, Research, and Practice, edited by T. Davies and S. P. Gangadharan (CSLI Publications, 2009), pp. 83–93.

• [4] A. Schuth, M. Marx, and M. de Rijke, in Proceedings of the 9th annual ACM international workshop on Web information and data management (ACM, 2007), pp. 97–104.

• [5] A. Schuth, Master’s thesis, Universiteit van Amsterdam (2007).

• [6] F. Wu, B. A. Huberman, F. Wu, and B. Huberman, Internet and Network Economics pp. 334–341 (2008).

• [7] F. Wu and B. A. Huberman, Public discourse in the web does not exhibit group polarization, arXiv:0805.3537 (2008).

• [8] V. Gómez, A. Kaltenbrunner, and V. López, in Proceeding of the 17th international conference on World Wide Web (ACM, 2008), pp. 645–654.

• [9] A. Grabowski, N. Kruszewska, and R. A. Kosiński, The European Physical Journal B 66, 107 (2008).

• [10] A. Grabowski, The European Physical Journal B 69, 605 (2009).

• [11] K. Kulakowski, M. J. Krawczyk, and P. Gawronski, Hate: no choice. agent simulations, arXiv:0908.2692 (2009).

• [12] M. Tsagkias, W. Weerkamp, and M. De Rijke, Advances in Information Retrieval pp. 191–203 (2010).

• [13] Y. Lee, J. Shim, H. Cho, and G. Woo, in 2010 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery (2010), pp. 456–463.

• [14] F. Schweitzer and D. Garcia, The European Physical Journal B 77, 533 (2010).

• [15] X. Si, Y. Liu, F. Xiong, Y. Zhang, F. Ding, and H. Cheng, Physica A: Statistical Mechanics and its Applications 389, 3711 (2010).

• [16] X. Si, Y. Liu, H. Cheng, and Y. Zhang, in Advanced Computer Theory and Engineering (ICACTE), 2010 3rd**International Conference on (2010), vol. 5, pp. V5–96–V5–99, ISSN 2154-7491.

• [17] F. Ding, H. Cheng, X. Si, Y. Liu, F. Xiong, and B. Shen, 4, 571 (2010).

• [18] F. Ding and Y. Liu, The European Physical Journal B 78, 245 (2010), ISSN 1434-6028.

• [19] A. Chmiel, P. Sobkowicz, J. Sienkiewicz, G. Paltoglou, K. Buckley, M. Thelwall, and J. Holyst, Physica A 390, 2936 (2011).

• [20] A. C. Jeong, The American Journal of Distance Education 17, 25 (2003).

• [21] A. C. Jeong, Distance Education 26, 367 (2005).

• [22] G. Mishne and N. Glance, in Third annual workshop on the Weblogging ecosystem (2006).

• [23] J. Leskovec, M. McGlohon, C. Faloutsos, N. Glance, and M. Hurst, in SIAM International Conference on Data Mining (SDM 2007) (2007), pp. 29406–29413.

• [24] L. Adamic and N. Glance, in Proceedings of the 3rd international workshop on Link discovery (2005), pp. 36–43.

• [25] K. Trammell, Public Relations Review 32, 402 (2006).

• [26] E. Hargittai, J. Gallo, and M. Kane, Public Choice 134, 67 (2008).

• [27] P. Sobkowicz and A. Sobkowicz, The European Physical Journal B 73, 633 (2010).

• [28] M. Chau and J. Xu, International Journal of Human-Computer Studies 65, 57 (2007).

• [29] M. Chau and J. Xu, in Pacific-Asia Conference on Information Systems, Kuala Lumpur, Malaysia (2006), pp. 1527–1549.

• [30] R. Prabowo and M. Thelwall, Journal of Informetrics 3, 143 (2009), ISSN 1751-1577.

• [31] M. Mitrović, G. Paltoglou, and B. Tadić, The European Physical Journal B - Condensed Matter and Complex Systems 77, 597 (2010), ISSN 1434-6028.

• [32] S. Gonzalez-Bailon, A. Kaltenbrunner, and R. Banchs, Journal of Information Technology 25, 230 (2010), ISSN 0268-3962.

• [33] C. Hsu, E. Khabiri, and J. Caverlee, in 2009 International Conference on Computational Science and Engineering (IEEE, 2009), pp. 90–97.

• [34] S. Jamali and H. Rangwala, in International Conference on Web Information Systems and Mining, 2009. WISM 2009. (2009), pp. 32–38.

• [35] E. Khabiri, C. Hsu, and J. Caverlee, in 3rd International AAAI Conference on Weblogs and Social Media ICWSM (2009), vol. 9.

• [36] P. Lange, Journal of Computer-Mediated Communication 13, 361 (2007).

• [37] G. Cohen, Journal of Personality and Social Psychology 85, 808 (2003).

• [38] N. Stroud, Journal of Communication 60, 556 (2010).

• [39] D. C. Mutz and P. S. Martin, American Political Science Review 95, 97 (2002).

• [40] C. Lord, L. Ross, and M. Lepper, Journal of Personality and Social Psychology 37, 2098 (1979).

• [41] M. Slater, Communication Theory 17, 281 (2007).

• [42] M. Meade and H. Roediger, 3rd, Memory & Cognition 30, 995 (2002).

• [43] E. F. Loftus, Learning & Memory 12, 361 (2005).

• [44] D. B. Wright, A. Memon, E. M. Skagerberg, and F. Gabbert, Current Directions in Psychological Science 18, 174 (2009).

• [45] M. Edelson, T. Sharot, R. J. Dolan, and Y. Dudai, Science 333, 108 (2011).

• [46] C. Palmer, The effects of conflict on attitude stability and extremity, Available at SSRN: http://ssrn.com/abstract=1733983 (2010).

• [47] T. Pettigrew and L. Tropp, European Journal of Social Psychology 38, 922 (2008).

• [48] M. MacKuen, J. Wolak, L. Keele, and G. Marcus, American Journal of Political Science 54, 440 (2010).

• [49] For detailed data see Web page of the National Electoral Commission (http://prezydent2010.pkw.gov.pl/PZT/EN/WYN/w/index.htm).

• [50] Polska The Times, May 30th, 2011.