In this paper, we analyze di erences in perspective between timelines created by news agencies from various countries. By employing methods for date and headline selection, which have been extensively evaluated in previous work, we show several types of bias that exist in the media landscape. As users typically select only a small number of news sources to follow, they necessarily experience at least some 'tunnel vision', which is commonly associated with the so-called 'filter bubble'. By recognizing and emphasizing the peculiarities of the users' self-selected news sources, we can help them to break out of the bubble.
One of the tasks of a journalist is to monitor, gather, curate and contextualize the relevant information for the target audience. He needs to go through an enormous amount of records with information of very diverse degrees of granularity, in order to put information into context and tell his story from all significant angles, and, at the same time, he needs to reduce the noise of irrelevant content.
Many newspapers regularly publish manually created timelines, which allow the reader to gain a quick overview over events that span a longer time period - such as the Egypt revolution - and to answer questions such as: how and when did the event start? What were the main consequences of the initial events? What happened to the main protagonists in the event?
Though convenient for the reader, creating a timeline is expensive, as it requires
substantial expertise. Moreover, creating a timeline is a very subjective task and
therefore huge di erences can be found in expert-generated timelines, even if they are on
the same topic [
On the Internet, the filter bubble, as coined and popularized by Eli Pariser [
These are some of the reasons why summaries of events, such as timelines, di er wildly, depending on the nationality, audience, political viewpoints and other sources of subjectivity associated with a newspaper. At the same time, readers are known to select a small number of newspapers - or even just one - that best match their personality, viewpoints and interests. As argued by Paul Resnick in his keynote at UMAP 20111, emphasizing di erences in perspective is an important instrument for allowing users to break out of their ‘filter bubbles’.
In this paper, we investigate methods for automatically recognizing and emphasizing di erences between various timelines, and evaluate them by comparing timelines from well-known news agencies and newspapers. These methods are based on our previous work on date and headline selection for news event summarization; by employing these methods and the Shannon Diversity Index, we are able to recognize dates that were considered important in one timeline but not in others, and di erences in news coverage on dates that were considered important in all timelines.
The work is carried out in the context of the EUMSSI2 project, in which crossmodal analysis techniques are developed for analysing news articles, videos and audio reports. This will allow journalists - and media consumers - to relate these messages with one another and to understand the underlying events. 2
Many studies specific to timeline summarization, such as [
Prior approaches dedicated specifically to date selection include work by Tran et al.
[
Additionally, by using the Shannon diversity index [
In order to recognize and emphasize di erences between timelines, it is important to first find events and corresponding dates that are ‘subjective’, in other words dates that are not considered important by all news agencies.
Our approach for subjective date detection makes use of a corpus of news articles
(more details in Section 4) and consists of two steps. First, we use a random walk model,
which we proposed in [
Formally, in the first step, we build a date reference graph, which is a fully directed graph G = (V; E), where V is the set of dates mentioned in any text in corpus C, including publication dates. The edges E = fe(di; d j)g indicate that at least one text published on di refers to the date d j.
We represent each link between events as a multi-value tuple
e(di; d j) = (Mi j; f req(di; d j); Itemporal(di; d j); Itopical(di; d j)) to integrate di erent measures of date importance. The first value, Mi j = N1 expresses the prior transition probability between 2 dates where N = jVj. The other values express the strength of the connection between di and d j, modeled by the following aspects: frequency ( f req), temporal influence (Itemporal) and topical influence (Itopical). Our random walk model uses these perspectives to rank the collection of dates. The intuition is that when a date d j is referred to from either a past or future news article (published on di), it is likely involved in the events that are reported in that article.
In the second step, we compute the SDI for the top-K ranked dates, based on the distribution of news articles published on those dates. For example, a date that contains only news articles from the BBC is considered less diverse than a date that contains news articles from several large news agencies over the world. The computation of SDI is sketched as follows:
S DI(di) =
R X pi ln pi i=1 where R is the number of news agencies from which we collected data and pi is the proportion of news articles from news agencies pi. That measure quantifies uncertainty in predicting the agency identity of a news article that is taken randomly from the dataset, hence, suggesting how subjective (or non-diversified) a date is. 4
We present our preliminary results of two experiments on detecting subjective dates in the Crisis data4 dataset, which contains ground truth (GT) timelines (written by professional journalists) as well as a corpus of around 12K news articles that cover events that happened in the context of four news stories: Egypt Revolution, Syria War, Libya War and Yemen Crisis. The dataset is suitable for our purpose for the following reasons: (1) it is a heterogeneous dataset that contains news articles and expert timeline summaries from 25 well-known news agencies and; (2) it covers long-term stories that have been happening since 2011, making the date selection problem non-trivial for any system. 4 The Crisis data is currently available at http://l3s.de/~gtran/timeline/ 4.1
In this experiment, we aim to detect dates that have been included in the timeline of only one news agency (and not in other timelines). We use a cut-o K = 50 to select the 50 most important dates, using our random walk model, and rank them by their SDI score. We then compare these selected dates with those in our ground-truth timelines. The performance of the subjective date detection process is measured as the proportion of the top-10 dates (this can be extended to a larger number) that are included in exactly one timeline.
The result is described in Table 1, which shows that a significant number of
important dates and their events are reported by only one news agency. We took a closer look
into those events (see examples in Table 2) and found evidence that the events that are
included in only one timeline are typically not about the main theme of the ongoing
news stories, but about related aspects, such as business and human rights. As we
discussed in [
In this experiment, we aim to find out how many important dates are not included in any of the news agencies’ timelines. Similar to the previous experiment, we picked the top-50 important dates from the results of our random walk algorithm and calculated the proportion of dates that are not mentioned in any timeline. The results are shown in Table 3. The high number of ‘missed dates’ does not mean, however, that the ground truth timelines are of poor quality: as explained earlier, journalists need to be selective when creating timelines. The SDI values also confirm that journalists usually do a good job in selecting the dates to be included: for all stories, except for Libya, the average SDI of dates included in at least one timeline is higher than the average SDI of dates that did not make it into any of the timelines. In other words, dates that are included in at least one timeline are typically covered by more news agencies than dates that do not appear in any timeline.
To better understand which events were left out of the timelines, we analyzed several of those dates with a high SDI score but that were not included. Some example events are shown in Table 4. Except perhaps for the Libya article, these are events that readers may be interested in, but that they will not be able to find in the timelines of the news agencies that they are subscribed to. In this paper, we have shown that it is possible to automatically find dates and corresponding events in news stories that are likely to be covered in only one or a subset of manually generated timelines. Further, we have shown that there are important dates that are not incorporated in any timeline.
As we argued earlier, it is unavoidable that timelines are selective and, to a certain extent, subjective. Still, it would be very desirable to make readers aware of the peculiarities of the timeline that they currently inspect. This can be achieved by highlighting dates with large di erences in coverage between news agencies or timelines, by adding links to other timelines for dates that are not incorporated in the current timeline, or, alternatively, by constructing an annotated ‘timeline of timelines’ for those who wish to contrast the bias in their personal selection of news sources with news sources that they usually do not visit.
Conversely, journalists will be able to create better balanced - or, alternatively, even more argumentative - timelines with feedback on their current selection of dates and events. Even though - as we have shown in previous work - it is possible to automatically construct timelines by selecting the most relevant dates and headlines, still manual processing and editing would be needed to enhance the communicative qualities of the timeline, and to adapt it to the needs of the readers. 6
The work was partially funded by the European Commission for the FP7 project EUMSSI (611057)