Search Engines in Learning Contexts: A Literature Review

The web is one of the primary sources of information for finding learning oriented documents. In addition, the main suitable way to find information and documents on the Internet is by using search engines. Search engines are constantly improving in terms of selection algorithms and in terms of the Human Machine interface (HMI). Also, these search engines are the basis of a new field of research called Search-As-Learning. The Search-As-Learning explores information search environments to enhance learning during user search tasks. This work focuses on our view of the state of the art in the field of Search Engines in learning context and Search-As-Learning, stressing on the most recent research. We conclude by highlighting the current shortcomings on improvement of the learning aspect within search engines, and present next work which will be the association of a layer above the traditional search engines to promote the appropriation of content during search task for a learning context


Introduction
Search engines are the most popular way to find information today. Even more, the use of a search engine to perform a web search is the most popular activity on the Internet [1]. In the past, documentary search was carried out mainly in libraries. Nowadays, searching for documents via the Internet is an essential practice for Internet users, particularly for teachers and learners. In this momentum, search engines have evolved into a refined search and sophisticated display of search results [2].
In terms of interface, it can be said that any search engine displays the return of queries using a page called Search Engine Result Page (SERP) [3]. Researches on studying the layout of search results pages are ongoing. Early results on this have shown that the layout of the results influences user selection [4]. In the literature, internet queries are mainly categorized into three classes: navigational queries, informational queries and transactional queries [5]. From the year 2007, new results have shown that the representation in enriched snippet [6], [7] on the result page was essential for informational queries [6], [8]. Today, in addition to traditional snippets, search engines are full of rich snippets [9], for example: home pages, events, products, job offers, Images, Tweets, Advertising, Videos, Addresses, Maps, News, Knowledge Graph.
eLearning [10] is extremely important in education nowadays. We underline for example that this method has been serving as the main relay during the period of expansion of covid19 [11]. Just recently, we have the advent of Search-As-Learning (SAL) which is a subarea of eLearning, surrounding learning by doing information search tasks on a system like Google, Bing, Yahoo Search, etc. Thus, researchers recognize search systems as rich online spaces in which users can learn and discover new knowledge while interacting with online content [12]. Around SAL, new research has focused on the acquisition of knowledge during information retrieval tasks on search engines [13], [14], [15], [16], [17]. In addition, it has been shown that SAL mainly relates to informational search task [13], [15], [16]. However, for a search task which is oriented learning as example, SERP produces relevant snippets, but also, irrelevant excerpt related to advertising, products, events and the like. Thus, one consequence is that these non-educational snippets grab the attention [18] and clicks of learners to produce cognitive overload. Therefore, with SAL, a new challenge in the field of information retrieval is to optimize results with regard to educational objectives [19], [20].
This work is mainly aimed to present a literature review around search engines in learning contexts. We first present some bases on information retrieval in traditional search engines, then we follow with the state of the art about Search-As-Learning. Finally, we conclude by highlighting the limitations in terms of the search result page in a knowledge acquisition context.

State of the art on search engines in learning contexts
This section presents related works that examines the different web strategies (algorithmic and human-machine interface) used in setting up SERP, and then present the state of the art in the subdomain of Search-As-Learning. Particular attention will be paid to the limitations of existing approaches to SERP oriented for learning.

Information retrieval and search engines
After significant advancements in search engine selection and ranking algorithms, researchers looked at various aspects of SERP and explained how their design affects user behavior. Here we provide a summary of the various aspects that have been investigated. We first focus on the selection and ranking of results in search engines in general, then we study: the layout of the SERP; the size of the SERP; how the text of the snippet is generated; and how much text should be presented in each summary of results. Some work on these has been done by David Maxwell et al [8].

a. Information Retrieval Models
In order to find the information that corresponds to a user need, classical Information Retrieval Support System (IRSS) are used. They are based on an architecture where the user is just a consumer of the content. In general, an Information Retrieval (IR) model is characterized by the document and query representation model "F" as well as the document-query matching process "RSV (q, d)". It was formally defined by a quadruple (D, Q, F, RSV (q, d)) [2] where:  D is a set of document representations,  Q is a set of queries,  F is a framework for modeling document representations, queries and their relationships,  R associates a real number to document-query-pairs (ranking),  RSV (q, d) is a ranking function which associates a real number with a query q ∈ Q and a document representation d ∈ D. Such ranking defines an ordering among the documents with regard to the query q.
There is a wide variety of models in the literature [2]. Figure 1 presents a classification of the most important, mainly distributed around three families: Boolean, vector and probabilistic models.  The main quality of a search engine is the efficiency of its ranking system [23]. The main filing systems are as follows: ─ Sorting by relevance: the relevance sorting method is used by most search engines [24]. Sorting by relevance allows you to evaluate the ranking of a page according to five factors: the weight of a word in a document; the density of a word; the weight of a word in the database, the correspondence; the close relationship. ─ Sorting by popularity: sorting by popularity favors web pages with a large number of backlinks. This sorting method was used by Google in 1998 to create the Pag-eRank algorithm [25]. ─ Sorting by audience: sorting by audience measurement favors web pages which have the highest "quantity" and "quality" of visits.

a. Search engine result pages and their presentation
Amongst the first works on SERP is that of Dumais and Chen on the presentation of snippets of results [26], [27]. This work investigated approaches to automatically classify result summaries for users, similar to the categorization approach used by early search engines. Subjects in a user study found the interface to be easier to use than the traditional "ten blue links" approach. Indeed, they were 50% faster to find the information displayed in the categories. This work was then extended by Dumais et al [27], where they explored the use of hover text to present additional details on search results based on user interaction. The search was found to be slower with the hover, possibly due to the fact that explicit decisions about when to (respectively not to) search for additional information where needed (respectively not needed).
Arrangements based on a grid have also been tested [28], [29], [30]. For example, Krammerer and Beinhaur [29] examined the differences in user behavior when interacting with a standard list interface compared to a tabular interface, as well as a gridbased layout (summary results placed in three columns). Grid users spent more time reviewing the summary results. The approach has shown promising results in overcoming problems such as positional bias [31], as observed by Joachims et al [32].
Marcos et al [9] performed an eye-tracking study, examining the effect of user behavior during their interactions with the SERP, and determining whether the richness of the results snippets provided on a SERP had an impact on the users' search experience. The result demonstrates that the representation of an enriched text document during the presentation of the results could have a positive effect on the evaluation of the relevance and the reformulation of the queries [33]. However, classification remained the most critical factor in the perception of relevance. The enriched summaries were also reviewed by Ali et al [34] in the context of navigation tasks. After this review, a positive relationship is shown between the length and informative nature of the abstracts and their attractiveness. However, we do note a loss of performance.
b. Generation of the snippet text Results summaries let users know whether a document is likely to be relevant or not [35]. As a result, research was undertaken to examine different types of snippets and the length of an snippet. The work initially focused on how these extracts were to be generated [36], [37], [38], [39], [40]. This early work was proposed to summarize documents by query (query-oriented summaries) or keywords in context, as opposed to simply extracting representative sentences or main sentences from the document [41]. Tombros and Sanderson [39] showed that subjects in their study were likely to identify relevant documents more precisely when using summaries of extract text associated with a query, compared to summaries of simply extract text generated from the first sentences of a given document. Recently, text-based snippets have been shown to be preferred on mobile devices [42].
When constructing snippets using query-oriented summaries, Rose et al [43] found that snippets offered a perception of the quality of the result. If the snippet texts contained truncated sentences or many fragmented sentences, users perceived the quality of the results more negatively, regardless of their length. Kanungo and Orr [44] maintained that readability is a crucial presentation that must be taken into account when generating a query-oriented text summary. Clarke et al [7] found that results with very short (or missing) snippets containing fewer query terms are not quite readable, and have a reduced number of clicks.
Recent work has examined the generation of snippets from more complex angles ranging from manipulation of underlying indexes [45], [46] to language modeling [35], [47], as well as the use of search data from users to improve the generation process [48], [49]. Approaches from previous generations may also fail to consider which parts of a document are actually useful for users. Ageev et al [48] incorporated postclick searcher behavioral data, such as mouse cursor movements and scrolling through documents, into a new model, producing behavior-oriented snippets. The results showed a marked improvement over a strong base for generating snippet fragments. Temporal aspects were also taken into account. Svore et al [50] conducted a user survey, showing that they preferred snippet text with trending content in snippets when searching for trending queries, but not for general queries.

c. Number of results per page
Examining behavioral effects on mobile devices when interacting with SERP has recently been the subject of much research (e.g. [51], [52]), with each device showing a different number of results. Recent research has shown that the value of RPP (Results per Page) can influence the behavior of document searchers [32], [53]. Understanding this behavior can help guide and educate those responsible for designing user interfaces.
In a Google report, Linden [54] indicated that users wanted more than 10 RPPs. Despite the fact that increasing RPP resulted in a 20% drop in traffic, it was speculated that this was due to the extra time required to send the SERP in addition. However, this drop in traffic can be attributed to other reasons. Oulasvirta et al [55] deals with the paradox of choice [56] in the context of search. This is because when there are more options for results, especially if those results are all very relevant, it can lead to poorer choice and worsen user satisfaction. Oulasvirta et al [55] found that presenting users with a list of search results with six items was associated with higher degrees of satisfaction, confidence in choices and a perception of caution than a list of 24 elements. Kelly and Azzopardi [57] broadly agreed with the findings of Oulasvirta et al [55]. The authors conducted an inter-subject study with three conditions, with subjects assigned to one of three interfaces: the base interface, indicating 10RPP (the "ten blue links"), and two interfaces displaying 3RPP and 6RPP respectively. The results showed that people using the 3RPP and 6RPP interfaces spent significantly more time reviewing the best results and were more likely to click on higher ranked documents than those on the 10RPP interface. The results also suggested that subjects using interfaces displaying less RPP found it comparatively easier to find relevant content than those using the 10RPP interface. However, no significant difference was found between the number of relevant elements found across interfaces. Currently, 10RPP is still considered the norm [58].

d. Snippet lengths
The lengths of the snippets were examined in different ways. A user study by Paek et al [59] compared a user's preferences and usability with respect to three different interfaces for displaying snippets of results. In the first case in their work, for information retrieval tasks, the interfaces displayed a normal SERP (i.e. a two-line fragment snippet for each summary, with a clickable link); an instant interface, where an expanded extract was displayed by clicking on it; and a dynamic interface, where hovering the cursor would trigger the expanded code snippet. Instant View has been shown to allow users to complete given search tasks in less time, with half of participants preferring this approach. The pioneering work of Cutrell and Guan [6] explored the effect of different lengths of snippets (short: 1 line, medium: 2-3 lines, and long: 6-7 lines). They found that longer snippets significantly improved the performance of information tasks (e.g., "Finding an airport address"). Users got better performance for information queries as the snippet length increased. This work was followed by Kaisser et al [60] who conducted two experiments that estimated the length of the result snippet based on the type of response (for example, finding a person, time, or location), and comparing the results of the selected snippet lengths to user snippet preferences to see if it could be predicted. The preferred extract length was shown to depend on the type of response expected.
Research work on examining which snippet sizes are appropriate for mobile devices is ongoing. This is important since the screen sizes are smaller. Also, snippet text considered acceptable on a computer screen may involve considerable scrolling or sliding on a smaller screen. Kim et al [61] found that subjects using longer snippets on mobile devices exhibited longer search times and similar search accuracy in informational tasks. Longer reading times and frequent scrolling or swiping (with more view movement) were exhibited. So the longer clips didn't look very useful on a small screen. Thus, an instantaneous or dynamic approach (according to Paek et al [59]) could also be useful for mobile search.
The presentation of summary results has an important effect on a user's ability to judge relevance [35]. Relevant documents may be overlooked due to non-informative summaries. Conversely, in informative excerpts, irrelevant documents may be reviewed due to a misleading summary. Recent work by David Maxwell et al [8] showed that longer summaries also lead to an increase in the cost of examinations. However, in the context of a search in a learning context (searches in a learning context are mainly related to information queries [13], [15], [16]), some works are favorable to the display of a lot of information around the subject, to allow the user to explore various aspects of the subject. Thus, it would therefore not be in vain in a learning process, to read an element that would seem irrelevant because it fits into the contours or the contexts of the subject.
Remarks on search engines and learning. Works on information retrieval via search engines presented previously does not particularly integrate the specific case of learning. However, as searching information in learning context is informational [13], [15], [16], we understand from previous research that a result page in a learning oriented search framework must have 10 results per page [58], summaries must be query based [42], [43] with an informative text (6-7 lines) as underlined in the last paragraph of Section 2.1.2.a and in Section 2.1.2.d. Note also that the new field called Search-As-Learning, is the specific field investigating the improvement of search engines in a context of education and learning.

Search-As-Learning
In the literature, studies in the context of "Search-As-Learning" can be divided into five main themes: 1-search tasks related to learning, 2-exploratory search oriented learning, 3-expertise and learning, 4-evaluation of learning during search task, and 5optimization of search results for education.
The recent work of Ran Yu et al [15] presented a state of the art by grouping the work around Search-As-Learning in the following way: Detection of the intention of web search sessions from user interactions, understand the learning process and outcomes, and support learning while searching the web. Table 1 provides a correspondence between the literature around Search-As-Learning and the work of Ran Yu et al [15].

Table 1. Grouping research works around Search-As-Learning
Consolidation of the work of Ran Yu et al. [15] In a more specific context of the Search-As-Learning literature  [62] on how users engage in such tasks using search engines describes a taxonomy of web search tasks, based on 4 months of log data from Google, Yahoo, and Bing. This taxonomy included certain tasks related to learning, covering the exploration of subjects ("discover more information about a specific topic"), the search of fact diverse ("find facts about a person" and "find a specific miscellaneous fact") and procedural learning ("learn how to perform a task").
Subsequently, Eickhoff et al [63] analyzed the fraction of search sessions involving a procedural or declarative knowledge intention. They found after selection criteria in the log files that 3% were clearly learning work from procedural or declarative knowledge. Raman et al [64] also noted the large number of intrinsic diversity search tasks. In fact, the various intrinsic search tasks are exploratory web searches intended to explore and learn more about multiple aspects of a specific topic. Jansen et al [65] found that information retrieval is a learning process with unique search characteristics specific to particular learning levels of Bloom's taxonomy [66]. Other recent works have attempted to assess the motivation of users to perform information retrieval tasks. For example, Kim et al [67] characterized this motivation in terms of a user's willingness to search and browse documents which are well above his reading level.
Exploratory search oriented learning. Learning oriented search activities often involve multiple interactions in the search process and the processing of multiple sets of findings that need to be interpreted in depth by researchers. Marchionini [68] asserts that search activities that promote learning focus more on "the acquisition of knowledge, the understanding of concepts or skills, the interpretation of concepts, and the comparisons or aggregation of data / concepts". Therefore, search activities that support learning require human participation on a more continuous and exploratory way during the search process. Exploratory search, which emphasizes broader information-seeking strategies focusing on deeper understanding rather than rapid factual responses, has become an alternative paradigm for fostering learning and inquiry in research [68], [69]. White et al [69] propose learning outcomes as an important future assessment method for exploratory search [69]. The work of Ran Yu et al [15] provides an overview of the challenges and research approaches to detect, understand and support lifelong learning in web search.
Expertise and learning during search tasks. Expertise is a dynamic characteristic of users that reflects learning over time. Wildemuth [70] examined how domain expertise was reflected in users' choice of search strategies, concluding that domain novices tended to exhibit search strategies increasingly similar to those of more experienced users as they learned more about the subject. Previous work [71], [72] has characterized domain expertise and search behavior in terms of metrics that can be derived from search logs, generally focusing on long-term patterns of behavior from session to session.
In one of the first large-scale journal-based research studies to examine the functionality of session-level tasks where people explicitly seek new knowledge, Eickhoff et al [63] examined assignment changes for these expert measurements. They focused on two types of specific knowledge acquisition: procedural knowledge (how to do something) and declarative knowledge (knowing facts about something). The authors found evidence of both learning progress within a single session and the persistence of learning between sessions. A significant portion of the new query terms came from results page snippets and recently visited pages, showing that the search process itself has helped increase the user's domain knowledge. Other studies like Zhang et al [72] have attempted to predict domain knowledge from users' search behavior. Finally, N Bhattacharya et al [17] recently showed that the differences in participants' knowledge changing are reflected in their search behavior and visual measures on results pages and towards pages containing content. In his survey paper, Vakkari [73] described more in detail how learning occurs in the search process.
Evaluation of learning during search tasks. A few studies have attempted to identify indicators of learning during the search process. Vakkari et al [74] found that students' level of knowledge about their subject could predict the characteristics of search queries, in the sense that students who know less about the subject are likely to use fewer search terms more vague in their queries. In another study with medical students, Vakkari and Huuskonen [75] found that efforts in the search task did not lead to better search results, but improved research works on a subject. Several information seekers have designed search methods to study learning as a measure of search outcomes. In one of the previous studies measuring learning, Hersh et al [76] showed how search allowed students to answer more questions in a post-search questionnaire. Instead of a quiz, Kammerer et al [77] asked study participants to write a summary of the topic after using the MrTaggy exploratory search interface, to assess learning. The quality of the abstract was rated based on topic-specific criteria, including the number of reasonable topics, the overall quality of the topic description, and the number of arguments. Wilson et al [78] developed systematic techniques for measuring the depth of learning at three levels: quality of facts, interpretation of data in statements, and use of criticism. Still on the assessment of learning during search, recent work by Ujwal et al [13] has shown that users acquire important knowledge during search sessions on topics with which they are less familiar.

Optimization of search results for education
─ User interface and interaction: The web today has several online learning platforms. Many Learning platforms are environments offering massive data and con-tent for learning [79]. Indeed, learning by doing information search tasks is also held within these environments and they have interfaces of the following components for learning: forum, content, board, collaboration, messages which are important for learning and tailored to the profile of learners. The works around the MOOCs also allows us to have enough content and a search environment with an adequate interface taking into account the profile of the learner [80]. However, in general-purpose search engines, learning support is not taken into account also because of the versatile nature of these environments and the variety of tasks performed in them. A central question for research in this area is how interfaces can be adopted to improve learning performance, even in versatile search engine environments. An attempt aimed at improving user engagement in learning-oriented search tasks by providing a richer representation of retrieved web documents was made by Arora et al [81]. Specifically, they explored methods to find useful semantic concepts in retrieved documents, with the goal of creating improved document surrogates for presentation in the SERP. Kodama et al [82] studied the relationship between the Google mental models of college students and their information search skills. During the Kodama et al [82] experiment, the students showed limited understanding of how search engines work. Additionally, this research suggests that developers and interface designers make the inner workings of the engine more transparent to learners. Recently, Qiu et al. [83] produced a conversational interface for search in learning context which can improve user engagement, augment user long-term memorability, and alleviate user cognitive load. ─ Retrieval and ranking: As current search engines are optimized by taking into account a need for information without considering the learning intention of a query, research on the optimization of extraction and ranking algorithms based on learning needs are relatively rare. In 2007, I. Bosnić et al [84] propose a system indexing learning resources with SMIL which helps finding the right data in multimedia lectures or HTML web pages and improves the quality of eLearning content retrieval. Then, Sandler et al [85] examined the potential of two ranking models (paragraph recovery model, dependency-based re-ranking) with varying purposes to improve the performance of learning-focused search engines. In their research, Syed and Collins Thompson [86], [87] proposed to optimize learning outcomes by selecting a set of documents while taking into account the keyword density and domain knowledge of the learner. Their theoretical framework provides a solid basis for the further study of learning-oriented recovery techniques. Lu and Hsiao [88] studied the information-seeking behavior of users in programming language forums. They also designed a personalized information retrieval assistant that promotes learning through the modeling of user behavior and query refinement, thus showing significantly improved learning efficiency. Karanam [20] present an improved research environment, YouUnderstood.Me (YUM), which aims at supporting the learning of children from Kinder-garten to Grade 9 by retrieving documents that do not only meet the information needs of children, but also their reading skills.
Remarks on the improved works on Search-As-Learning. We realize that current work for the improvement of search engines in a Search-As-Learning context proposes a reclassification (Sandler et al [85]), a refinement of the query (Collins Thompson [86], [87], Lu and Hsiao [88]) or an environment (Pera et al [20], Qiu et al [83], Bosnić et al [84]). Nevertheless, shortcomings are observed because the proposals do not highlight a filter to eliminate content for non-educational purposes to avoid cognitive overload. Also, we do not notice a reorganization of the results which takes into account the level of Bloom's taxonomy perceptible from the user's request. Also, with regard to the result extract, it was mentioned previously in Section 2.1.3 that it would be desirable to have on the search result page, extracts of text focused on the queries with a satisfactory length. At present, no work exploits these aforementioned mentioned components.

Discussion
Throughout this paper, we propose a state of the art around search engines in a learning context. Thus, we start with a presentation of some basic concepts on Information Retrieval (IR) because, it is largely the advances in IR that lead us to search engines today. Then, we highlight some aspects of the principle of search engines selection and ranking. Next, we present some relevant works on the display side of search engine results by examining: SERP and their presentation, generation of the snippet text, Number of results per page and snippet lengths.
All of these aspects around SERP lead us to understand that today's search engines are designed for all types of searches tasks. Indeed, the emphasis is not particularly on optimization in learning contexts. However, search for learning being informational search task, we note according to these previous researches that:  snippets must be query oriented: so, remove irrelevant snippets (Ads, Events, etc.);  snippets must have fairly informative content (6-7 lines);  snippets should not be more than 10 results per page.
After doing this study around SERP, we turn this review around Search-As-Learning whose questions are more concise in terms of learning oriented search task. This presentation of works around Search-As-Learning categorizes researches according to the following themes: search tasks related to learning, exploratory search oriented learning, expertise and learning during search tasks, evaluation of learning during search tasks, optimization of search results for education.
Following the presentation of works around Search-As-Learning we note in our aforementioned remarks certain aspects which we consider important to take into account in the optimization of search engines for learning cases. These aspects are:  consideration of the level of the learner in accordance with the Bloom's taxonomy when classifying the results;  elimination of irrelevant contents for learning.
In fact, in our future works, we intend to optimize the search results with regard to educational objectives by relying on the remarks that we made in this literature review.

Conclusion
In this work, we provide a state of the art on search engine in a learning context. We started by recalling some information retrieval (IR) bases, then, we illustrated the search engine selection and display result policy. In terms of user interface, we brought out some research around SERP. After this literature review on contexts surrounding SERP, we brought specifically a state of the art on learning during search tasks. We note that Search-As-Learning addresses questions around human learning during the search tasks. Concerning the selection and ranking of algorithms in search engines, particularly the Search-As-Learning optimization algorithms, we realize that no algorithm is interested in determining the learner knowledge level from the request and the elimination of non-educational content from search engines. In addition, works on SERP agree that exploratory (informational) search should display fairly informative content in result snippets or summaries. Based on the state of the art, we can say that the advantages of informative snippets include the rapid assessment of the relevance of an extract on the results page, the time to time saving during the appropriation of the content, the ability for the user to better reformulate his requests during the search process, and the improvement of the readability of the extracts text. In our next research, our goal will be to provide a layer above the search engines that responds to two levels: algorithmic and interface levels. At the algorithmic level, we identify the level of the user from the request, and then eliminate the extracts of noneducational content (advertising, Tweets, Addresses, Maps, etc.) during the search to keep only snippets for educational purposes which will be classified according to the level identified in user request. At the interface level, we propose a result snippet model for educational content.