Perspective-based search: a new paradigm for bursting the information bubble

In a classic IR system, users access the search results either from a general perspective (nonpersonalized methods) or from their own perspective (personalized methods). Therefore, they are unable to view the searched topic from the viewpoint of different beliefs. In many circumstances, they may even be unable to identify different viewpoints existing on the search topic. This can cause the formation of information bubbles and lead to false perceptions of reality.

For example, suppose that a user believes in a particular religion and that religion is dominant in the user’s country. If the person wants to research other religions, a personalized search system may bias the results towards the user’s beliefs (e.g., by ranking the webpages viewed more by the user’s compatriots higher), thus leaning towards a point of view criticizing other religions. This can be unfair and result in severe misjudgments and more polarized viewpoints, which can trap the individuals into their limited understanding of the truth, i.e., into their own information bubbles. Also, another problem may happen when the user is in a minority by biasing the user’s search results (according to information such as the user’s social connections and geographical location) towards what the user is reluctant to (unless an exhaustive profile on the search topic is available indicating the user’s beliefs). The bias may even be inherent to the source data (White and Hassan 2014; Kulshrestha et al. 2017), and the search results may be biased even without any personalization and thus violate the openness norm (Bozdag and van den Hoven 2015) (e.g., when the number of followers of a religion is significantly lower compared to other religions and they have produced less online content). These problems can happen in almost any topic in which there are conflicting opinions, like many political, social, and even scientific topics.

Moreover, in these situations, formulating the queries that convey the real user needs with the traditional personalized paradigm may be challenging. For example, if users search for critiques on their religious beliefs, personalization techniques may bring up the results confirming their own beliefs like the documents written by the members of the same religion, quoting the critiques in a distorted manner and answering them unfairly, i.e., documents written in a similar manner to what Garrett (2013) mentioned about partisan information seekers: “There is little evidence that partisans are driven to form balanced opinions; instead, they seek information in order to better defend their position or ridicule others”. However, if the user were able to view the different existing perspectives on the topic in a categorized format, the user would be able to read the ones of interest and conduct research in a more fair fashion and with less partiality about the topic. As another example, one might want to know the attitudes of a country’s political orientation about a political issue. By including the country name in the query, the search engine normally brings up the pages mentioning the country name, whereas a page that describes a topic from a country’s political viewpoint does not necessarily mention the country name.

As a solution, in PBS different possible perspectives on the search topic are presented to the user, and the user chooses the desired perspective if wanted. This puts control in the users’ hands and allows the user to view the topic from different perspectives. For example, a Russian may search about “2014 Crimean crisis” and choose USA as the desired perspective or vice versa. Or a Christian may search a religious topic and mention Islam as the perspective.

Another possible type of perspective in PBS is to search from the viewpoint of a particular user (or group of users) in a social network, e.g., when a user is interested in another user’s preferences and wants to search a topic from the user’s point of view or wants to know what the user and those who are similar to the user think about a particular topic. Viewing the world from the perspectives of different people may also help the user in thinking outside the box and come up with more creative ideas (Teevan and Yu 2017). For instance, assume that in a social network for multimedia content a user’s taste of movies is similar to another user. By using PBS, a user can easily search the query “movie” with determining the similar user as the perspective to obtain the movies of interest. We call these perspectives social networks members’ perspectives. This type of perspectives could be considered similar to the traditional personalized search but with selecting a particular user (or group of users) as the perspective by the searcher, instead of using the searcher as the perspective.

It’s important to note that this type of perspectives does not interfere with users’ privacy since it considers only the public profile of the user. Moreover, some users may even try to enrich their profiles to help others better understand their beliefs. For example, a religious figure may add information and (or) links to websites that represent preferred beliefs and lifestyle to help the believers of that religion view things from the same perspective. Thinking this way, PBS may become an interesting media for publishing thoughts and lifestyle information. We discuss this in more detail in Section 4.2.

Data sources could also be used as perspectives. For example, if an individual wants to know the attitudes of the Republican Party towards a particular political or social topic, the individual could choose a news agency with Republican leanings, e.g., Fox News (Morris and Francia 2010), as the perspective and obtain the results from that point of view. Similarly, one could search by choosing Xinhua as the perspective to retrieve results from the viewpoint of the Chinese government. The important point is that unlike traditional IR operators like “site:” or “insite:”, the search results should not be limited to the documents of the chosen data source and the data source is only used as a guidance to what really the user intends to see. This allows the user to simply enter what the user is searching for without sacrificing the retrieval recall. An interesting application of this type of perspective is to provide a search facility in a website (e.g., a news agency) in which the users can search the Web from the perspective of the website.

Demographic factors could also be used as perspectives. Users with different personalities, expertise, age ranges, language proficiencies, etc. could have different information needs. Allowing users to choose the demographic perspective themselves allows them to indicate their search goals more conveniently. For example, mentioning “kid” as perspective along with query “poem” could allow the mother of a child to find childish poems easily. As another example, a user searching about food may choose a country as the perspective to become familiar with the food culture of the country.

Some, but not all, of these factors are taken into consideration in some search systems (e.g., search engines for children (Gossen 2015)), but presenting them in a uniform format, as in PBS, can be beneficial. More details on the benefits of this unification are explained in Section 4.1.5.

Although there are some tools for retrieval based on the sentiments (see Section 5.4.1) of the results (e.g., Zheng and Fang 2010; Demartini 2011; see Section 5.3.1), PBS offers a straightforward way of integrating sentiment-based search into general search engines. This can be done by considering sentiment-related perspectives (e.g., negative and positive) in PBS.

The search component is the main retrieval component of a PBS system. This component retrieves the results for the query and with respect to the selected perspective(s). At first, when the user issues a query and still no perspective is selected, the search component considers a default perspective in the retrieval. The default perspective can be either “My Perspective” (which results in personalized results) or “General Perspective” (which results in nonpersonalized results). Then, when the user selects one or more perspectives, the component retrieves the results based on the perspective(s).

The perspective extraction component is responsible for identifying and categorizing relevant perspectives in PBS. For example, it identifies what types of viewpoints are available for the searched topic, or it explores a social network’s users and suggests specific pundits as perspectives, or it detects that there are conflicting opinions on the topic and considers positive and negative perspectives as the relevant perspectives.

Also, since many relevant perspectives may exist for a query, and considering the importance of the UI, this component ranks the extracted components to allow the UI component to better present them to the user. Some of the current research problems on retrieval (ranking documents) could also be raised about this ranking (ranking perspectives). For example, a system can personalize the ranking based on the user’s preferences.

One of the most important points that demands special attention in PBS is the user interaction. Some UI approaches are used in similar IR tasks. For example, in search result clustering (SRC), the clusters are mainly represented as lists or hierarchies (see Fig. 3). However, we think that the traditional representation of concepts in the form of lists or hierarchies has not been very successful, and we present our suggestion.

One of the reasons why most of the Web clustering engines (Carpineto et al. 2009) were not welcomed with open arms by the users may be that their over-informativeness and complexity make the users confused and disoriented (e.g., see the sidebar in Fig. 3). Therefore, UI is important in PBS. One of the solutions to this issue is to rank all the possible perspectives and display only the most interesting ones to the user instead of displaying a mess of perspectives and creating confusion. Figure 2 shows an example UI, in which a perspectives area is added to a search results page. In this UI design, the default list of results is shown as the search result and the perspectives of interest are also shown to the user; the search perspective can then be changed if desired. In the example, “Republican Party” and “Democratic Party” are suggested as perspectives to an American user searching about the abortion laws, and the user can change the search perspective with just a single click. Moreover, an option to view more perspectives on the topic is provided to the user beneath the suggestions to present additional perspectives to the user.

In addition to the core components, some supplementary components may be integrated into a PBS system. For example, a search topic may relate to many perspectives, and viewing all of them may be time-consuming for the user. A PBS system may employ a perspective search component to allow the user to search among the related perspectives. The component can help the user to find the desired perspectives by showing perspectives that are relevant to the perspective search query. For instance, the user may enter US as the perspective search query, whereas the system uses United States of America; the perspective search component retrieves the perspective as the result of the user’s perspective search. Figure 4 shows an example of perspective search in which auto-completion is also offered.

One approach is to obtain the information by user collaboration and from user-generated content (based on Web 2.0 principles). This approach has been used in many websites and is shown to be practical, effective, and successful. We can name Wikipedia,¹ Pinboard,² Stack Exchange,³ and reCAPTCHA⁴ as some successful examples of this approach, in which their content is generated by the users and the focus is not on artificial intelligence for generating the content. A similar approach can be employed for PBS. In other words, users help to identify perspectives and relate them to documents. For example, users define perspectives by providing the perspective labels (e.g., the Republican Party) and the corresponding Wikipedia pages as the descriptions. Meanwhile, users tag the webpages with the defined perspectives, similar to folksonomy websites like Pinboard. It should be noted that this approach is prone to mistagging and employing methods for detecting mistags may become necessary over time. Other crowd-sourcing methods are also possible. For example, one might employ games to collect the required information (gamification) (von Ahn 2006). Another approach to obtain the information from users is discussed in Section 4.2.

Some websites such as Debatepedia,⁵ Debate.org,⁶ Debatebase,⁷ and ProCon.org⁸ provide information on controversial issues, conflicting opinions, different sides of debates, etc. Another approach is to use data on these websites to identify perspectives and the corresponding webpages. Also, the Argument Web (Bex et al. 2013) and Argument Interchange Format (Chesñevar et al. 2006) can facilitate obtaining the required data and semantics in a machine-readable format. Although this approach may be a low-cost solution to obtain the required information, the available data may be limited and available only for some topics. This can make its applicability limited.

An extension to the approach suggested in Bennett et al. (2015, p. 1) also seems promising. They propose using “the aggregate search behaviors of others in the beneficiary’s demographic groups” in searching on behalf of others (see Section 5.3.6). Based on the existence of selective exposure in humans, search engines can estimate the membership of a user to a group, say Muslims, by observing their interactions with the system. After identifying many members of the group, the search engine can then observe their behaviors in other topics, e.g., searching for a restaurant, and use their aggregate search behaviors in retrieving the results from the perspective of that group, possibly resulting in Halal restaurants. Moreover, by observing different behaviors in different groups on a search topic, the search engine can infer that the groups have possibly different opinions on the topic, and thus suggest the groups as perspectives in the searches on that topic. For example, by observing that in the United States those who are inclined toward the Republican party and those who are inclined toward the Democratic party behave differently on searches about “progressive tax”, it can deduce that the Republican party and the Democratic party are two candidate perspectives for the search topic in the United States. This approach can be considered as an implicit form of using user-generated content, in which the required information is implicitly extracted from the collaborations of users with the system.

Another approach is to use machine-learning tools to automatically extract the required information. This approach demands special attention from academia and raises new research questions on how to extract the required information automatically. Some results in previous works suggest that the automatic extraction is indeed possible. For example, Lin et al. (2006, p. 6) stated that, “The results suggest that the choice of words made by the authors, either consciously or subconsciously, reflects much of their political perspectives. Statistical models can capture word usage well and can identify the perspective of documents with high accuracy.” Thus, unsupervised approaches similar to the ones used for topic modeling (Blei and Lafferty 2009) may be useful for identifying perspectives. However, at least at the early stages, some kind of supervision may be required for learning such methods. The abovementioned approaches may be used for the supervision. For example, data on websites such as Debatepedia may be used to learn the methods. Moreover, top-down methods for creating a Semantic Web (Stevenson and Ciravegna 2003; Iskold 2007c) (e.g., see Li and Huan 2008) may be helpful. Some of the methods used in the papers mentioned in Section 5.4 may also be applicable. In particular, extracting the required information for some perspective types may be easier than the others. For example, the information for the sentiment perspectives may be relatively easily extracted using methods similar to the one used in Zheng and Fang (2010). Moreover, a bootstrapping approach, similar to the one used in Park et al. (2011a), may be employed—that is to say, the system tries to automatically extract the required information with an acceptable precision, and then the users can amend the information.

Because of the existence of social phenomena like homophily, analyzing the link structure of the Web and social networks can help in revealing the existence of polarizations and finding stances on topics. For example, Dori-Hacohen (2017) showed that Wikipedia articles exhibit homophily, i.e., linked pages are more likely to have the same controversy label than nonlinked pages. It leverages the phenomenon to detect controversy in Wikipedia. Thus, one way to obtain the required perspective information is by analyzing the link structures of the Web and social networks, which requires further research.

As time passes and PBS becomes more widespread, development of ontologies/semantic networks/knowledge graphs (OSKs) for relations of concepts/topics and perspectives could facilitate the search as well as production of more useful and exhaustive OSKs (before that, available OSKs like DBpedia⁹ can be used for this purpose). Using Semantic Web inference techniques is also promising. For example, the system may use the information (e.g., click-through data) about the perspectives of the topics similar to (according to the OSK) the searched topic to enrich its information about the perspectives of the searched topic. This can help reduce the sparseness problem. As an example of the use of OSKs in PBS, suppose that the user selects “Russia” as the perspective when searching about a political issue; the system can detect that “Vladimir Putin” is the president of Russia according to an OSK, and thus also retrieve pages tagged with the “Vladimir Putin” perspective but are not explicitly tagged with the “Russia” perspective. Also, website owners currently use mark-ups and tagging (e.g., with Schema.org tags) to better introduce their contents to search engines and to raise their ranks in searches. The very same idea can be employed in PBS. In other words, content producers may tag different parts of their websites with perspective tags, representing the relevancy of those perspectives, which can be a rich and reliable source of information about perspectives. This approach is similar to the bottom-up approach of creating a Semantic Web by embedding semantic annotations into the data (Iskold 2007b). The success of the approach in the Semantic Web (e.g., Guha (2013) stated that two years after the launch of Schema.org over four million Internet domains use its markup) suggests that it may be a promising solution for implementing PBS.

One of the main advantages of PBS is that it promotes users’ awareness of the information bubble problem. This is done by presenting the users with different viewpoints on their search topics and informing them that search results can be different from each viewpoint. This is in contrast to some other approaches like enabling serendipity (Section 5.3.2), diversification (Section 5.3.3), and using intermediary topics (Section 5.3.5). This characteristic is important for the search engine to have the value of transparency discussed in Bozdag and Timmermans (2011). It allows the users to understand the limitations of their bubbles and try to get out of their bubbles if the topic really matters. In addition, this can promote critical thinking and avoid believing falsehoods.

PBS also informs the users about different viewpoints on the search topic and allows them to view the results from those viewpoints. By informing individuals about different stances on the topic, PBS can prevent individuals from moving to extremes and forming polarized views of the topic.

One of the challenges in search personalization is to correctly handle the user’s context. For example, suppose that a user has always been a fan of a brand like Apple, and the user’s profile is tied to it, but for some reason the user is now searching for positive or neutral information on a rival like Microsoft. In this case, a personalized search engine may bias the results toward the unintended ones. As another example, the search results for a believer of a religion, say Christianity, may be biased when searching about another religion, say Islam, in a personalized search system. These biases may trap the users into their own information bubbles.

Retrieving the appropriate results for the user in different contexts and avoiding unwanted biases was a challenge in the traditional personalized search engines. PBS solves the problem in a straightforward way by allowing the user to control the search. This control lets the user exploit the power of personalization without being stuck in the information bubble. PBS thus conforms to the value of identity discussed in Bozdag and Timmermans (2011).

A characteristic of PBS that is of paramount importance is that it respects users’ autonomy. Unlike several other approaches (e.g., see Sections 5.3.2, 5.3.3, and 5.3.5), the approach taken by PBS puts the users in the driver’s seat and allows them to choose when and how they want to view the topic from other perspectives. Even, PBS allows the users to knowingly insist on their own viewpoints and choose to view the topic from their own perspectives (e.g., selecting their political party as the perspective), although after they are apprised that there are other perspectives. Thus, PBS conforms to the value of autonomy discussed in Bozdag and Timmermans (2011). Respecting users’ autonomy also avoids several possible negative civic and ethical implications (see Bozdag and van den Hoven 2015; Dori-Hacohen et al. 2015) and increases the trustworthiness of search engines.

PBS can be provided to the user without much modification to the well-established UI of search engines. The user can perform normal search operations and PBS does not intervene in the day-to-day search tasks. This is in contrast to the some approaches such as clustered presentation (see Section 5.3.4) in which a totally different UI is presented to the user. This characteristic is crucial for popularizing a system aiming at mitigating the bubble, because a different UI will probably require providing PBS in a different webpage, which can dramatically narrow possible users.

Moreover, in many circumstances being informed about different perspectives on the search topic is not important to the user and manipulating the original, possibly personalized, search results to cover different perspectives (e.g., in the diversification approach; see Section 5.3.3) may even reduce user satisfaction. The weaker performance is also shown to exist in other coercive mitigation approaches (e.g., see Graells-Garrido et al. 2016a). Because of reduced satisfaction, such an approach may not be an appropriate option for search engine provider companies and may cause conflicts of interest. Our suggested UI for PBS shows the user the default search results without any manipulation, but it also informs the user about the existence of different perspectives (Fig. 2) and allows the user to intentionally select a perspective. This can avoid reducing the user satisfaction, while helping to mitigate the information bubble. Not having this property can be a big barrier to the adoption of a solution for bubble mitigation by major search engines.

Although each of the perspective types proposed in Section 2.1 could benefit users in several aspects, the power of PBS is not limited to the advantages of its individual perspectives. Viewing all those points of view in a unified form, named perspectives, puts a lot of flexibility into the paradigm. First, the user could become familiar with all of the features in a uniform format, which can reduce the overhead of feature adoption considerably. This decreases the cost and time of learning how to interact with the system and a better experience is delivered to the users. This can be contrasted with the previous approach in which the user had to visit a website for kids’ search, another one for sentiment-based search, etc.

Second, by uniforming many capabilities, a common ground for research is established in which synergy can take place. Moreover, new research problems arise and new facilities can be delivered to users. An example (mentioned in Section 3.1.2) is personalization in ranking perspectives: ranking the perspectives with respect to both the query and the user’s preferences. Another example is perspective search, which was discussed in Section 3.1.4.

Furthermore, by representing many points of view in a unified form, combining different points of view (perspectives) would also be possible, which lets the user refine the point of view and gradually add more perspectives.

The filter bubble was coined by Eli Pariser in his book The Filter Bubble: What the Internet Is Hiding from You (Pariser 2011). Pariser discussed three challenges with respect to the personalized filters: (i) “… you’re alone in it. A cable channel that caters to a narrow interest (say, golf) has other viewers with whom you share a frame of reference. But you’re the only person in your bubble. In an age when shared information is the bedrock of shared experience, the filter bubble is a centrifugal force, pulling us apart.” (p. 10) (ii) The filter bubble is invisible. When you visit a conservative or liberal news source, you may know that you are visiting content curated to present a particular viewpoint. But when you use a personalized information system, you probably do not know what the system thinks about you and how it models you and whether it is right or wrong. The system decides what you should see and you may not even be aware of the existence of such personalized filters. (iii) You don’t choose to enter the bubble. When you read a particular newspaper, you have actively decided which newspaper to read and probably have some ideas about its leanings. You are the one who decides what kind of filter to use in obtaining information about the world. But, with personalized filters, you do not make such a decision. Filters come to you and they are hard to avoid, because they drive up profits for the companies who use them.

The filter bubble problem can arise in different systems. For example, in a search engine it may arise through personalization algorithms. In a recommender system, it may arise through recommender algorithms. And in a social network news feed, it may arise through algorithmic curation (Berman and Katona 2018) of content.

The concept of filter bubble has attracted much attention, resulting in many studies. Davies (2018, p. 14) redefined filter bubbles as socio-technical recursion: “…users coming back for more because they like the suggestions offered to them which have been inferred from their previous searches.” Some research validated the existence of filter bubbles (e.g., see Nguyen et al. 2014), whereas the other findings (see Bechmann and Nielbo 2018; Dutton and Fernandez 2019) suggested that the filter bubble problem may not be that severe. Tran and Herder (2015) introduced a method for detecting the filter bubble in ongoing news stories. Bozdag and van den Hoven (2015) and Bozdag (2015) provided interesting explanations on the relations between the filter bubble and different democracy models and also the relations between tools introduced to mitigate the filter bubble and the democracy models. Moeller and Helberger (2018) reported on the effects of algorithmic news recommendation systems or filter bubbles. Resnick et al. (2013) reviewed some approaches for bursting the information bubble and promoting diverse exposure.

The echo chamber theory (Sunstein 2001a, 2001b, 2007) has its roots in the social science concept homophily (McPherson et al. 2001) and in the psychology-related concept selective exposure (Sears and Freedman 1967; Frey 1986). Cavoukian and Chibba (2018, p. 37) explained: “In news media an echo chamber is a metaphorical description of a situation in which information, ideas, or beliefs are amplified or reinforced by transmission and repetition inside an ‘enclosed’ system, where different or competing views are censored, disallowed, or otherwise underrepresented.” Echo chambers are shown to exist in social media (An et al. 2013a). One may argue that the echo chamber phenomenon is more severe in cyberspace, because in the real world people are compelled to have social interactions with many people (not necessarily with those they want to) and this makes them exposed to different topics and viewpoints. Garrett (2013) authored an interesting paper on selective exposure that reflects the state of the research area that offered some criticism on the echo chamber concept.

As opposed to the traditional system-centered IR methods, personalized methods rely on a user-centered approach and try to tailor the results to the user’s preferences. Indeed, personalized search usually tends to bias the results towards the interests of the user according to the available information. Personalized search can be considered as a tool for implementing PBS. There are many studies on personalization in IR that are beyond the scope of this paper; readers are referred to Ghorab et al. (2013), Bouadjenek et al. (2016), Pasi (2010), and Steichen et al. (2012) for more information.

Personalization could severely lead to the problem of information bubble. Unlike traditional personalized search systems in which the only perspective is “My Perspective”, in PBS different existing perspectives about the search topic are presented to the user and the user can select the desired perspective. Thus, PBS can deliver the benefits of personalization to the user without sacrificing breadth of vision or compromising autonomy.

An IR approach related to PBS is SRC (Carpineto et al. 2009). Because a query may be related to multiple topics or subtopics, clustering search results and presenting the clusters to the user can help the user search more conveniently. For example, for the query “Java”, the results are categorized into different categories such as “Programming Language”, “Drink”, “Geographic Location”, etc. Figure 3 shows an example of the output of such system, Yippy¹⁰ (formerly Clusty (Iskold 2007a)). Some other related tools are KeySRC (Bernardini et al. 2009), Lingo (Osinski and Weiss 2005), Lingo3g¹¹ (Osinski and Weiss 2007), and SnakeT (Ferragina and Gullì 2004).

SRC is different from PBS in four aspects. The first and the most important difference is that in SRC the results are categorized based on their facets (e.g., programming language, drink, island, etc. for query Java), whereas in PBS different types of perspectives could be considered but only one is a perspective based on the result facets. Thus, SRC could be considered as a special case of PBS.

Moreover, in SRC clusters are expected to mostly represent concrete concepts that are explicitly mentioned in the documents (e.g., using Java as a programming language in a document), whereas perspectives in PBS may refer to abstract (possibly subjective) concepts not necessarily mentioned directly in the documents (e.g., being from a liberal viewpoint).

In addition, SRC is mainly focused on helping the user find and select the desired search results, whereas the focus in PBS is on mitigating the information bubble.

On the other hand, in the popular approach to SRC, the categories are represented as lists or hierarchies, which are confusing to the end user and waste time finding the appropriate results. This kind of UI is very likely to make users reluctant to use the provided categories. The common approach to UI in such systems may be one important reason why these systems are not so popular. Carpineto et al. (2012, p. 15) stated that “… while it [clustering] is less effective for partial subtopic coverage because of the additional cognitive effort required on the part of the user to browse through the cluster labels before examining any document”. The problem is resolved by the UI suggested for PBS, in which the results are presented in a simple format with minimum intervention the user’s common searches (see Section 3.1.3).

Another IR approach related to PBS is faceted search (Tunkelang 2009; Wei et al. 2013). Zelevinsky (2010) defined facets as: “navigable and summarizable properties, tagged onto the records in the system.” Faceted search is commonly used in e-commerce websites (e.g., Amazon¹²) and is also used in other search systems on metadata-rich collections (e.g., library.oum.edu.my/oumlib/iportal_search/faceted); items are gvien tags such as type, category, author, subject, format, and price that can be used to navigate the collection for the desired items. The main difference between faceted search and SRC is that in faceted search items are, presumably manually, tagged with facets, whereas in SRC the clusters are generated at the query by clustering the search results. Hearst (2006) provides a more detailed comparison between faceted search and SRC.

There are three major differences between faceted search and PBS. First, faceted search is mainly defined in the domain of documents with manually tagged facets, whereas PBS is broader and can be used in a Web search. Second, similar to SRC, faceted search is mainly related to facet perspectives, whereas PBS covers different perspective types, e.g., belief perspectives. Lastly, faceted search is focused on facilitating exploration of the collection by the users, whereas the most important goal in PBS is mitigating the information bubble.

One approach to mitigate the information bubble is to promote the user’s awareness about different dimensions of the task, e.g., search or recommendation by UI facilities and visualization techniques. For example, Carson (2015) proposed recommending that Google allows users to manually toggle between personalized and nonpersonalized results (e.g., through a switch icon). This causes users to pay more attention to the information bubble phenomenon and understand its existence as well as empowering them to control it. Much earlier in 1990s, Negroponte (1995, p. 154) coined “The Daily Me” to refer to personalized news summaries and “The Daily Us” to refer to nonpersonalized news summaries, and then suggested: “These are not two distinct states of being, black and white. We tend to move between them, and, depending on time available, time of day, and our mood, we will want lesser or greater degrees of personalization. Imagine a computer display of news stories with a knob that, like a volume control, allows you to crank personalization up or down. You could have many of these controls, including a slider that moves both literally and politically from left to right to modify stories about public affairs.” Stracke (2014) suggested allowing the users to modify the amount of personalization in search by enabling or disabling categories of search signals (e.g., search history, browsing history, and geography). Nagulendra and Vassileva (2014) suggested visualizing the filter bubble and allowing the user to control what resides inside it and what does not. The user controls which categories to be inside the bubble and which users to be inside it for each category. Similarly, El-Arini et al. (2012) proposed transparent user models, i.e., user-interpretable badges such as “liberal” or “Apple fanboy”, for personalization, which can make the filters more transparent to users and allow them to correct them. Demartini and Siersdorfer (2010) and Demartini (2011) proposed showing the sentiment polarity (see Section 5.4.1) of each search result next to it so the user can view results with different sentiments. Demartini (2011) also allowed the user to filter the results based on their sentiments. Aktolga (2014) proposed a similar approach based on sentiments but also suggested showing the times, i.e., time references mentioned within documents. Gao et al. (2018) designed a system that provides visual hints and recommendations of opinions on different sides with different indicators. They concluded that the proposed system “is promising in raising users’ awareness of their own stances and preferences, as well as mitigating selective exposure to information” (Gao et al. 2018, p. 10). Zheng and Fang (2010) presented a retrieval system that delivers search results in “for” and “against” categories and tries to find the topics inside each category and show the corresponding topics (two topics discussing the same query perspective but expressing different opinions) next to each other. Liao and Fu (2014a) considered displaying a source position indicator, showing both valence (pro/con) and magnitude (moderate/extreme) of positions on controversial topics and examined its effects on users’ selection and reception of diverse opinions. They also considered and examined the effects of a source expertise indicator, i.e., a design feature that explicitly distinguishes information sources with higher competence (Liao and Fu (2014b). Liao et al. (2015) also considered providing aspect indicators (e.g., displaying whether a comment is about side effects or effectiveness of a medicine) to mitigate selective exposure. They argued that providing the aspect indicators to users can reduce their resistance to diverse opinions and encourage them to explore information from different perspectives, since it can potentially narrow down their biased beliefs to a subset of aspects, and thus reduce cognitive dissonance. Schwartz (2012) explained Bing’s elections portal that can filter news sources with right or left political leanings. Similarly, The Wall Street Journal provides “Blue Feed, Red Feed”,¹³ in which it categorizes the Facebook posts on some controversial topics into two liberal and conservative categories and displays them side by side. Oh et al. (2009) evaluated the effect of displaying the political orientation (see Section 5.4.5) of blog posts in the weblog search task. Munson et al. (2013) and Munson (2012) developed a browser widget that shows the relative amount of the user’s liberal and conservative visited pages to nudge the user to have a more balanced reading behavior. Munson and Resnick (2010) also evaluated the effect of simple presentation techniques such as highlighting agreeable items or showing them first on readers’ satisfaction.

Some general differences between works in this approach and PBS are:

Maccatrozzo (2012, p. 2) proposed that another approach for mitigating the information bubble is to enable serendipity; they defined serendipity as “the ability to make fortunate discoveries by accident”. Maccatrozzo et al. (2017, p. 1) also used a similar approach and defined serendipity as “making a pleasant and relevant discovery that was unexpected”. Recommending serendipitous items to the user can mitigate the information bubble in two ways: users become aware of items outside of their own bubble and these items may catch users’ attention and encourage them to explore new paths and visit unprecedented items. Exploring these new items may also affect future recommendations.

Serendipity can be contrasted with PBS in several ways.

Another popular approach to mitigate the information bubble is to diversify the results in search engines or recommender systems. Diversification has been employed for other goals (e.g., tackling query ambiguity; see Santos et al. 2015) and several methods are proposed for it (e.g., Dang 2014).

Demartini (2011), Giachanou et al. (2014), Aktolga and Allan (2013), Aktolga (2014), Kacimi and Gamper (2011, 2012), and Meguebli et al. (2014a) considered diversifying the search results according to their sentiments. Aktolga (2014) considered diversification by time and opinionatedness. Kacimi and Gamper (2011, 2012) and Meguebli et al. (2014a) used the semantic similarity of documents (the Jaccard similarity of topic elements and (or) key concepts of documents) to obtain semantic diversity, which can be considered a form of aspect-based diversity (see Section 5.4.3). Kiritoshi and Ma (2014, 2015, 2016) addressed sentiment-based diversification of news, and they also considered a form of aspect-based diversification by using two measures “difference in factor coverage” and “difference in details”. Park et al. (2013) introduced agenda diversity and questioned its methods. Graells-Garrido (2014), Graells-Garrido and Lalmas (2014), and Graells-Garrido et al. (2016b) considered diversifying user timelines (e.g., in Twitter) according to the geographical dimension (see Section 5.4.6) of micro-posts. Wang et al. (2015) attempted to employ online analytical processing (Han et al. 2011) techniques to generate brief personalized summaries delivering news updates in multiple dimensions (such as entity, topic, and time). Although this paper does not focus on diversification or bursting the information bubble, its approach of presenting only updates of the news without repeating what the user has already read may indeed cause it to present diverse information to users. Munson et al. (2009a, 2009b) addressed diversifying news aggregators according to political orientations (see Section 5.4.5). Yom-Tov et al. (2014) focused on increasing exposure to different political opinions and discussed the role of language models (Zhai 2008) on people’s exposure to opposing viewpoints. When the language model of a document is similar to that of an individual, the document is more likely to be read despite describing an opposite viewpoint. On the other hand, An et al. (2013b) placed doubt on the wisdom of exposing people to diverse opinions.

Diversification can be contrasted with PBS in several ways.

Some studies (Park et al. 2008, 2009, 2010, 2011a, 2012; Ma and Yoshikawa 2009; Chhabra and Resnick 2012, 2013) considered presenting different dimensions of news events as separate clusters to users to mitigate media bias. In clustered presentation, news is clustered according to its aspects (see Section 5.4.3). The clusters are presented to users to make them aware of the presence of different dimensions and help them explore different dimensions, thereby mitigating the biases incurred by reading only from a single news source.

Clustered presentation can be contrasted with PBS in several ways.

Two reasons for the creation of the information bubble are homophily (McPherson et al. 2001) and selective exposure (Sears and Freedman 1967; Frey 1986). In a social network context, users tend to connect with like-minded people and only read agreeable information. One approach to mitigate the information bubble is to recommend users with dissimilar viewpoints. However, this approach may provoke cognitive dissonance (Festinger 1957). Therefore, some studies (Graells-Garrido 2014, 2015; Graells-Garrido et al. 2014, 2015, 2016a; Bleackley 2017 and in Liao et al. (2015) in the context of a political forum) proposed using intermediary topics, i.e., topics that users with distant views on some sensitive issues have in common, to overcome the cognitive dissonance and provide opportunity for a connection.

This approach is different from PBS in two aspects.

Sometimes, a user wants to go out of the bubble and explore new ideas. Some methods facilitate this exploration. Taramigkou et al. (2013) proposed a methodology to help users explore new items that may be serendipitous and interesting to them. They implemented their methodology in a music recommender system. The general procedure of the method is as follows. The user selects the destination genre (identified by Latent Dirichlet Allocation (Blei et al. 2003)) and the system identifies the destination user who is a best match for the destination genre. Then, the shortest path between the user and the destination user is calculated and a predefined number of artists are selected per each node in the path. The list of artists is then presented to the user.

Bennett et al. (2015) considered the problem of searching on behalf of others, e.g., trying to find cool clothes for a teenage child. Although this approach does not pay attention to the information bubble, it can be considered a way of searching according to the interests of people other than the searcher, and thus a way of viewing the world from outside of the searcher’s information bubble. This approach is similar to using our demographic perspective; however, Bennett et al. (2015) did not provide a technical method for realizing on-behalf-of searches and focused on studying these searches according to Web search logs and characterizing this type of search.

Doris-Down et al. (2013) introduced a mobile application to bring people with different political viewpoints together for a cup of coffee and a face-to-face conversation to combat echo chambers.

These methods aim to help the users willing to go out of their bubbles, whereas PBS aims to raise user awareness of the bubbles and provide the required tools to mitigate them in searches.

One reason people may get trapped in bubbles is that they are not aware of the existence of other opinions and competing viewpoints. Thus, some systems aim to make the users more conscious of disputed claims and encourage them to judge the claims more critically. For example, Ennals et al. (2010) implemented a browser extension, called Dispute Finder, to alert users when the information they read online is disputed. They also proposed a popup interface with articles for and against the claim when a highlighted claim is clicked on. Rbutr¹⁴ is another extension that aims to tell users when the webpage they are viewing has been rebutted, disputed, or contradicted elsewhere on the Internet. Dori-Hacohen and Allan (2013) considered automatic identification of controversial webpages and suggested alerting users about controversial issues, for example, through a warning at the top of the webpage stating that “This webpage represents one of several perspectives on a controversial topic”. Also, Hypothes.is¹⁵ is an annotation tool that allows the users to add comments or critiques on webpages at the page-level or sentence-level and collaborate with others, which can help in improving the epistemic quality of information.

Moreover, sometimes a user is aware of controversy in an issue, but does not precisely know each group’s views on the issue and is unable to reach a conclusion. Kriplean et al. (2012) developed ConsiderIt, which invites the participants to create pro and con lists about issues. It also enables users to adopt the pros and cons contributed by others into their own lists. These lists can help everyone better explore and understand facets of issues. This approach can foster deliberation and promote more balanced civil discourse.

Kriplean et al. (2014) employed reference librarians to integrate an on-demand fact-checking service into a crowdsourced voters’ guide. Each user could request that a librarian fact-check a pro or con point contributed by another user and the librarians “… reviewed these requests, identified verifiable claims, researched the accuracy of those claims, and posted a public evaluation and report about each claim.”

This approach can be contrasted with PBS in several ways.

Sentiment¹⁶ analysis is a salient domain of research in IR, which “deals with the computational treatment of opinion, sentiment, and subjectivity in text” (Pang and Lee 2008, p. 1). Several studies discussed in Section 5.3 (Demartini and Siersdorfer 2010; Zheng and Fang 2010; Demartini 2011; Kacimi and Gamper 2011, 2012; Aktolga and Allan 2013; Aktolga 2014; Giachanou et al. 2014; Graells-Garrido 2014; Liao and Fu 2014a) considered sentiment in their research. Because sentiment analysis is well-known in IR, we do not discuss it further here. For more information, readers are referred to Pang and Lee (2008).

Sentiment can be considered as only one of the many possible perspectives in PBS.

A lot of topics are controversial, i.e., there is not an obvious consensus among nearly all people. For example, many of religious and political issues are controversial. Some studies address the problem of detecting controversy or identifying stances (“an overall position held by a person toward an object, idea or proposition” (Somasundaran and Wiebe 2010, p. 1)) in a controversial topic. Dori-Hacohen et al. (2015) elaborated on this and discussed several related issues, including the definition of controversy and the related ethical concerns. Awadallah et al. (2012) presented OpinioNetIt, an opinion base of facets, opinion holders, and their opinions, which are automatically extracted from news and other web documents. OpinioNetIt includes information such as a politician’s position—support or opposition—on a controversial topic. Agrawal et al. (2003) addressed partitioning authors into opposite camps within a given topic. Park et al. (2011c) considered unsupervised classification of news articles based on disputant relations. They identified opposing disputant groups dynamically and classified the articles based on their positions. Paul et al. (2010) proposed a method for summarizing contrastive viewpoints in opinionated texts. Ren and de Rijke (2015) addressed the task of summarizing contrastive themes, in which they consider a viewpoint to be “a topic with a specific sentiment label” (p. 1), similar to Paul et al. (2010), and a theme as “a set of viewpoints around a specific set of topics and an explicit sentiment opinion” (p. 1). Somasundaran and Wiebe (2009, 2010) addressed classifying stances in product debates and ideological debates, respectively. Thomas et al. (2006) used transcripts of U.S. Congressional floor debates to determine whether the speeches represenedt opposition to or support of the proposed legislation.

Although sentiment analysis is related to controversy detection and stance analysis, they are not identical. Awadallah et al. (2012, p. 1) stated that: “… political controversies are much more complex and opinions are often expressed in subtle forms, which makes determining pro/con polarities much more difficult than with product reviews for cameras, movies, etc.—the typical objects in prior work on opinion mining.”

Our concept of perspective can be contrasted with the concepts of controversy and stance in several ways.

Content producers view the same events from different viewpoints and select different aspects of reality. Several methods focus on identifying these aspects. Park et al. (2008, 2009, 2010, 2012) introduced NewsCube, an online news service aiming at mitigating the effect of media bias. NewsCube automatically clusters news according to their aspects and provides multiple viewpoints on a news event of interest to the readers (see Section 5.3.4). Park et al. (2011a) proposed NewsCube2.0 and exploited user collaboration to determine the viewpoints of and the framing¹⁷ methods used in the news articles. Using user collaboration, NewsCube2.0 can deal with various types of articles and different forms of framing methods, beyond straight news articles and fact selection, which were the focus of NewsCube. Chhabra and Resnick (2012) proposed CubeThat, which takes a similar approach to NewsCube but allows the user to change the cluster assignment of a news story. Similarly, Ma and Yoshikawa (2009) employed a clustered presentation of news to deliver diverse news reports on a certain news event. Meguebli et al. (2014a) and Kiritoshi and Ma (2014, 2015, 2016) implicitly considered aspects for diversifying news recommendations by exploiting semantic dissimilarity and sentiment dissimilarity measures. Kiritoshi and Ma (2014, 2015, 2016) employed a “difference in details” measure to consider the differences in the amount of details of news articles. Kacimi and Gamper (2011, 2012) employed a similar approach using semantic dissimilarity and sentiment dissimilarity for diversification in the search task. Kacimi and Gamper (2012) also provided a summary of opinions, including different sentiments and arguments, covered by the results of the search. On the other hand, Meguebli et al. (2014b) focused on determining users’ political orientations by comparing their aspects to the aspects extracted from Wikipedia pages.

This concept can be contrasted with our concept of perspective in several ways.

Ideologies are reflected in people’s statements. For example, two editorials written by a Palestinian author and an Israeli author about the Israeli–Palestinian conflict are expected to be divergent and induce contradictory conclusions. Lin et al. (2006) addressed the problem of identifying ideological perspectives on the bitterlemons corpus,¹⁸ which includes editorials written by Israeli and Palestinian authors on the Israeli–Palestinian conflict. Lin and Hauptmann (2006) also considered the problem of determining whether two collections are written from different perspectives. Lin et al. (2008) further proposed a statistical model for determining ideological discourse in an unsupervised manner.

Identifying ideological perspectives is fundamentally different from sentiment analysis about a particular item. One’s sentiment is different for each item, while one’s perspective is probably more static, presumably underpinned by their ideologies or beliefs about the world (Lin et al. 2006).

This concept can be contrasted with our concept of perspective in several ways.

Some works focused on political orientations (Table 1). For example, Kim and Lee (2014) considered the problem of determining the political orientations (liberal, conservative, and moderate) of news articles based on orientations of Twitter users. Oh et al. (2009) used SVM (Cortes and Vapnik 1995) and Naïve Bayes to classify political viewpoints of blog posts. Zhou et al. (2011) addressed the problem of classifying political leanings of news articles based on the assumption that conservative users vote for conservative articles more often, and similarly for liberal users and articles. Also, Kulshrestha et al. (2017) performed their analysis for quantifying search bias with respect to the Democratic and Republican parties. Park et al. (2011b) used commenters’ sentiment patterns for identifying political orientations of news stories. Meguebli et al. (2014b) proposed an unsupervised approach for identifying users’ political orientations based on a knowledge base extracted from Wikipedia.

The border between the political orientation category and the ideological perspective category may not be obvious at first glance. We put methods having a generative nature in the ideological perspectives category, as ideology is somehow generative and can generate new thoughts. On the other hand, methods we put in the political orientations category have a discriminative nature, for example Oh et al. (2009) used SVM to classify political orientations.

This concept can be contrasted with our concept of perspective in several ways.

A number of papers considered some dimensions of documents. For example, Aktolga (2014), Aktolga and Allan (2013), and Singh et al. (2013) considered the time dimension, i.e., time references mentioned within documents; Aktolga (2014) and Aktolga and Allan (2013) also considered the sentiment dimension and Singh et al. (2013) also considered the geographical dimension. Odijk (2016) considered the framing and time dimensions and focused on four generic news frames: conflict, human interest, economic consequences, and morality. Liao et al. (2015) used aspect indicators, i.e., interface features indicating the focused aspect of information (e.g., indicating whether the comment was about effectiveness or side effects of a medication) to mitigate selective exposure. Liao and Fu (2014a) also studied how source position indicators, which indicate positions of information sources (e.g., pro/con), impact user’s information-seeking behavior. They extended their approach by also considering source expertise indicators, i.e., “design features that explicitly distinguish information sources with higher competence” Liao and Fu (2014b, p. 1).

Several of the works in this category focused on tasks other than search, e.g., Liao et al. (2015) and Liao and Fu (2014a, 2014b) reported on systems for browsing comments. In addition, the dimensions are not necessarily used as viewpoints. For example, in Singh et al. (2013), time and geographical dimensions were used as filters that the user can use to dig into specific subsets of the search results. In contrast, in PBS, the focus is on the viewpoint. For instance, for the query “tourist attractions”, a document may not mention a geographical location, but is relevant to its corresponding perspective since, e.g., it mentions a tourist attraction near it.

Another concept employed in previous works is demographics. For example, Graells-Garrido (2014), Graells-Garrido and Lalmas (2014), and Graells-Garrido et al. (2016b) considered geographic factors. Gossen (2015) considered age and focused on search engines for children. Also, Bennett et al. (2015) considered demographic groups (e.g., men over 50 years) and focused on searching on behalf of others.

Demographics can be considered as only one of the many possible perspective types in PBS. In addition, the works in this category do not necessarily have mitigating the information bubble in mind. For example, Gossen (2015) focused on search engines for children, whereas PBS allows a user (possibly an adult) to search on behalf of a child.