An Eyetracking Study on Human Categorization

Table of Content

List of tables

List of figures

1. Introduction

2. Theoretical Setting
2.1 Linguistic Background
2.1.1 Human Categorization
2.1.2 Previous Research
2.2 Online Advertising

3. Eyetracking Study
3.1 Hypotheses
3.2 Method
3.3 Results and Discussion
3.3.1 Results and Discussion: Word Group
3.3.2 Results and Discussion: Picture Group

4. General Discussion

5. Future Research

6. Conclusion

References

Appendix

List of tables

Table 1: Summary of hypotheses

Table 2: Results of the maximum fixation analysis for the object category (word group)

Table 3: Results of the maximum fixation analysis for the brand category (word group)

Table 4: Results of the maximum fixation analysis for the abstract category (word group)

Table 5: Results of the maximum fixation- & last fixation analysis for the object category (word group)

Table 6: Results of the maximum fixation- & last fixation analysis for the brand category (word group)

Table 7: Results of the maximum fixation- & last fixation analysis for the abstract category (word group)

Table 8: Results of the maximum fixation-, last fixation- & dwell time-analysis for the brand category (word group)

Table 9: Results of the maximum fixation analysis for the object category (picture group)

Table 10: Results of the maximum fixation analysis for the brand category (picture group)

Table 11: Results of the maximum fixation analysis for the abstract category (picture group)

Table 12: Results of the maximum fixation- & last fixation analysis of the object category (picture group)

Table 13: Results of the maximum fixation- & last fixation analysis of the brand category (picture group)

Table 14: Results of the maximum fixation- & last fixation analysis of the abstract category (picture group)

List of figures

Figure 1: Final results of the object category (word group)

Figure 2: Final results of the brand category (word group)

Figure 3: Final results of the abstract category (word group)

Figure 4: Final results of the object category (picture group)

Figure 5: Final results of the brand category (picture group)

Figure 6: Final results of the abstract category (picture group)

1. Introduction

As a global network, the World Wide Web (henceforth WWW) enables peo­ple from all over the world to interact with each other and to share and exchange information. In the context of Web 2.0, social networking sites, personalized homepages, e-mails, and other communication services have become more and more important since 2002. According to Internet World Stats (2009a), the number of Internet users rose from 558 million in 2002 to 1,596 million in March 2009. North America shows the highest Internet penetration rate with 74.4%. In comparison, 48.9% of Europeans are connected to the WWW. On average, 23.8% of the world's population uses the Internet. The Web's global scope and its easy accessibility have turned the Internet into "the most recently developed mass medium and communication tool of the 21st century" (Janoschka 2004: 9). As such, the WWW has become increasingly important in Online Marketing. Due to its global reach, Internet can be used perfectly to inform a large number of people about a product or service. Regarding this aspect, the WWW plays a crucial role in a company's communi­cation policy, particularly in advertising.

Analyses about users' online activities revealed that e-mail services and the use of search engines to find information take up the most of users' time online (cf. Internet World Stats 2009b). For 60% of Internet users, checking e-mails represents the most important online activity. At the same time, search engines have become more and more important. The number of people accessing search engine websites rose from 33% in 2002 to 49% in 2008 (cf. PEW Internet 2008). Hence, advertising on e-mail- and search engine websites reaches the most Internet users and is most effective. Statistics of eMarketer (2008) about the online advertising spending of US companies shows that online advertising on search sites is most important: 40% of the budget spent on online advertising was used for paid ads on search sites in 2008.

Due to the importance of Search Engine Marketing, the following research paper is concerned with online advertising on search sites. Advertisements displayed on search engine websites consist of text only and are triggered via keywords. This means that online advertisers choose keywords which determine when to show their advertisement. Of course, it is not solely the selected keywords which determine if the ad is displayed or not. There are other important factors, such as the quality of the website and the budget the advertiser is willing to spend. However, on the whole, the keywords and the ad text are highly important in keyword-based Online Advertising. This points to the fact that the success of an online advertisement is mainly determined by language and the Internet user's interaction with language. If there are words which are more likely to be used as search terms by Internet users, then online advertisers should include these words in their ad text and their keyword list rather than other words. The better the keywords and ad text match the user's search term, the higher the probability that the ad will be displayed and the better the ad performance.

This paper aims to give an answer to the question of which words are preferably used by Internet users when they place a search query. Concluding from this, the thesis intends to provide advertisers with recommendations for keywords that ensure a successful ad performance. In order to deal with this issue, the linguistic theory of the basic level concept will be applied to keyword- based Online Advertising. According to the basic level advantage, basic level words carry the most information, are most distinctive and are preferably used in speaking (cf. Rosch et al. 1976).

The research paper is structured as follows. Firstly, the theoretical background of the study is explained. With regard to the linguistic basics, important facets of human categorization and an overview of previous research are presented. Furthermore, this section describes the fundamentals of online advertising. The second part of the thesis focuses on the eyetracking study, which was conducted, and its results. Moreover, I provide a detailed description of the study's design and formulate hypotheses concerning the results of the study. Most importantly, this section deals with the presentation, evaluation and interpretation of the results. Based on these, this part involves a general discussion of the results including implications for online advertisers. Thereupon, I mention some ideas for future research. In conclusion, the thesis' main points are summarized.

2. Theoretical Setting

The following section deals with the theoretical setting of the eyetracking study. First of all, some linguistic basics on human categorization as well as an overview of previous research concerning the basic level advantage are given. Secondly, I point out important aspects of online advertising in order to provide a better understanding with respect to the marketing background.

2.1 Linguistic Background

2.1.1 Human Categorization

Real world objects do not crop up randomly with certain attributes. As Rosch et al. (1976) found, features such as have wings mostly appear with have feathers (383). Further examples can easily be detected in our everyday life: Barking is mostly associated with dogs, wet is more likely to be found with rain than with sun, the attribute have leaves often goes together with have blossoms. Hence, our world does not consist of "intrinsically separate things” (Rosch et al. 1976: 383), but the occurrence of attributes with certain objects in our world is structured. This, in turn, has an impact on our perception of the world. There must be an equivalent system in our mind organizing real world objects and attributes. From an economic point of view, this approach makes sense when considering the infinite number of objects in the world and the finite capacity of our mind. A structure within our mental storage of words, namely the mental lexicon, is required in order to enable us to easily access new unknown objects and their attributes. Otherwise, our mind would not be able to deal with the extensive information load of the external world (Radford et al. 1999: 193).

In order to approach the idea of the mental structure, one should think of language as a medium to describe the world and to express thoughts. In general, a language consists of lexemes, which are defined as the "abstract notion of words" (Radford et al. 1999: 66). When describing the world or expressing thoughts, one makes use of lexemes. According to Frege (1982: 210), a word or, to use the linguistic term, lexeme is three-fold. It consists of its linguistic sign and its referent, which constitutes the meaning of the word by relating it to a definite object in the external real world. The third dimension of a lexeme is its sense, which further determines the meaning of the word by forming a mental representation of this word in the human mind. The sense forms the lexical entry of a word in our mental lexicon. Since the sense of words should be understood as an abstract mental image, the notion "concept" (Saeed 2003: 33) is preferred in the following. Concepts do not occur independently from another. Rather, they are related to each other and organized in a "gigan­tic network of interconnected concepts” (Hudson 1995: 33). This complies with the idea of Radford et al. (1999) that there is a "structure of the lexicon” (193) rather than seeing "the lexicon as a list of lexical entries” (193). The knowledge that arises from the connection and relation of concepts is called conceptual knowledge.

The most frequently used type of conceptual lexical relations is taxonomy. According to Kay (1971), a taxonomy consists of the finite set T of Taxa (for example, the set of all animals) and is based on the relation "strict- inclusion-of-set” (868): A set x includes another set y if every member of y is a member of x. However, at least one member of set x must not be included in y. Hence, set x precedes set y, which Kay refers to as immediate precedence. This implies a one-way relation within the taxonomy (869). Further axioms defining taxonomic structure are the idea of the unique beginner and the notion of partition. Concerning the unique beginner, there must be one member of the taxa which includes every other member, such as plant. As to the axiom of partition, each set must be divided into subsets. Moreover, the levels of a taxon vary with the degree of set inclusion: the greater the inclusiveness of a set, the higher its position in the taxonomy. This multi-layered approach implies that the different levels within a taxonomy carry different amounts of information. Kay (1971) uses level one to refer to the highest level of a taxa, the unique beginner (for example, plant); level two addresses the taxon which immediately precedes level one (for example, tree). The level that succeeds level two is level three (for example, oak). In accordance to Rosch et al. (1976: 383), these different levels are referred to as levels of abstraction in the following study: Level one is named the superordinate level, level two is called the basic level and level three is referred to as the subordinate level.

However, one has to note that there are different types of lexical relations among taxonomies. The lexical relation of hyponymy is most important for the topic at stake. Hyponymy is based on meaning inclusion and defined by "An X is a kind/type of Y" (Radford et al. 1999: 197). For example, tree and flower are referred to as co-hyponyms of the superordinate plant. Another type of lexical relation is meronymy, which establishes a part-whole relation among its members. For instance, finger is a meronym of hand. Both, hyponymy and meronymy are types of taxonomies in the sense that "we meet more general terms as we ascend to higher levels" (Radford et al. 1999: 196). Additionally, there are kinds of lexical relation which are not taxonomic, for example, synonymy (identity of meaning) and antonymy (opposites of meaning).

Overall, taxonomies allow us to structure concepts inherent to our mind and to access this conceptual knowledge via its relational nature. Note that it is not the "amount of knowledge" that is important but more "the integration of the meaning of a new word into our existing knowledge" (Saeed 2003: 38). Taxonomies enable us to categorize objects, infer information from one object to the other and, thereby, gain new knowledge about the object.

Besides the easy acquisition of knowledge and the economic advantage of information storage through categorization processes, human categorization built upon taxonomies is extremely important in cognitive processes. For example, if one searches for something, for instance something to cut with, one usually comes up with knife, scissors or saw. Depending on the purpose, one will choose the appropriate tool. This short example shows a taxonomy with cutting device as the superordinate. Knife, scissors, and saw form the basic level while specific tools, such as blade or buzz saw, are subordinates. Via taxonomies one is able to conclude from attributes to objects or the other way around. Furthermore, one can connect certain objects to a more general idea of this object or to more specific objects. This reveals the importance of human categorization on cognitive processes. What is crucial for the present study is the fact that people normally come up with knife and scissors rather than blade or cutting device. This hints to one preferably used level within taxonomic structures. This is not only the focus of this research paper. Previous research has also been conducted on this topic.

2.1.2 Previous Research

Early research on human categorization suggested that there is a middle level of categorization, which is preferably used (cf. Brown 1975). The most influential research was conducted by Rosch et al. (1976), who referred to this middle level as the basic level. The study provided not only a definition of the basic level, which has frequently been recalled in further research papers, but it also mentioned some implications. The basic level was defined as "the level at which categories carry the most information, possess the highest cue validity¹ and are, thus, the most differentiated from one another” (Rosch et al. 1976: 383). In terms of taxonomies' strict-inclusion, the basic level was considered as ”the most inclusive level” of all levels of abstraction (Rosch et al. 1976: 385).

Rosch et al. (1976) proved this definition as well as the special status of basic objects in a series of experiments. Stimuli were objects of nine taxonomies consisting of six non biological taxonomies (musical instrument, fruit, tool, clothing, furniture, vehicle) and three biological taxonomies (tree, fish, bird). The structure of these served as a guideline for the following study and helped to form the taxonomic stimuli of my study. The research of Rosch et al. (1976) tested the basic level feature "most inclusive” extensively by analyzing attributes common to the objects, motor movements, similarity in shape and the average shape of these objects. For the first two, subjects were asked to list attributes or motor movements. The relevance of the listed attributes was tested by a judgment task. The similarity in shape focused on the ratio of overlapping and non-overlapping parts and it was calculated by a computer program. The average shape of objects was tested by an object naming task and a category naming task. By using different types of tasks, Rosch et al. (1976) provided substantive evidence in favor of the basic level. The results showed that the most attributes and motor movements were listed for the basic objects while less attributes were named for the subordinate level. The least number of attrib­utes and motor movements was found for the superordinate level. This revealed that basic objects are most inclusive, and that they carry the most information. In the analysis of similarity in shape, basic level objects were proven to be the most differentiated from one another. Examining the average shape of objects showed that while there was no average shape for superordinate objects, stimuli at the basic level were easiest to identify. This re-verified the "most inclusive” aspect.

There was one unexpected finding concerning the levels of abstraction for biological objects: The suggested superordinate level (tree, bird, fish) appeared to be the basic level. The hypothesized basic level revealed characteristics of the subordinate level. This was considered when preparing taxonomies for the following experiment.

In the second part of their study, Rosch et al. (1976) provided implications for the basic level concerning mental representations of class, object recognition, development of categories, evolution of categories, and linguistic usage. Again, Rosch et al. (1976) used different types of tasks to provide solid evidence for the implications: a signal detecting paradigm, a priming paradigm, matching task, a category verification task, an object naming task, and an analysis of the American Deaf Language. Within the category verification task, Rosch et al. (1976) showed that superordinates were verified the slowest and basic objects were identified most rapidly, followed by subordinates (413). From this finding, the experimenters concluded the basic first hypothesis proposing that basic objects are processed first. This motivated Rosch et al. to speculate about information processing of superordinates and subordinates. The most important finding in terms of this study was the following: During an object naming task, subjects used mostly basic level names to refer to the objects. Furthermore, Rosch et al. (1976) showed that basic level names are first acquired by children (426). Due to an analysis of American Deaf Language, names for basic level objects evolve first (427).

The experimenters also addressed other important issues. Firstly, they mentioned that a person's basic level evolves from "an interaction between the potential structure provided by the world and the particular emphases and state of knowledge of the people who are categorizing" (1976: 430). Hence, the environment of the perceiver and his cultural background play a crucial role in human categorization. In this context, Lakoff (1987) even described the basic level as "human-sized" (51). Rosch et al. (1976) mentioned a second issue: prototypicality. This means that one member of a category has more attrib­utes/fewer attributes in common with the other members of that category/the contrasting category than any another member of that category/the contrasting category (433). Thirdly, Rosch et al. (1976) addressed the issue of expertise knowledge and its influence on the basic level (432). They suggested the emergence of two or more basic levels as an increase of expertise knowledge but did not provide evidence.

This hints to the shortcomings of the study. The idea of the highest cue validity within the basic level was contradicted by Medin (1983) and Murphy (1982) who proved that cue validity as an indicator of basic level membership is highest at the superordinate level. Furthermore, the basic first hypothesis was investigated by other research papers with varying opinions about it. Lastly, the experimenters' suggestion concerning the emergence of two basic levels due to expertise knowledge was proven to be false in a study by Tanaka & Taylor (1991).

However, in terms of the present study, Rosch et al. (1976) provided extensive research on the basic level and offered a useful definition. The researchers highlighted the universality of the principle of categorization: It is not the context of categorization, but "the principle of category formation that is claimed to be universal” (Rosch et al. 1976: 435). The different types of tasks provided substantive evidence in favor of the basic level. Moreover, further research, such as the subsequent presented study of Tanaka & Tversky (1991) on expertise knowledge, was motivated by Rosch et al. (1976).

In the study by Tanaka & Taylor (1991), the experimenters tested the effects of expertise knowledge on the structure of classification. The study compared dog and bird experts with novices using a feature listing task, an object naming task and a category verification task. In doing so, they were able to prove the influence of expertise knowledge on categorization. For example, experts listed more attributes for subordinate objects than novices. In contrast to novices who mainly used basic level names, bird experts named objects with subordinate names. With regard to the dog experts, these preferred to use basic level names. The reason for this discrepancy between bird and dog experts is the following: The way dog experts usually interact with the object does not require the knowledge of specific names. This further confirmed the finding of Rosch et al. (1976) on the importance of the interaction between the perceiver and the external world as well as its impact on human categorization. Moreover, while novices verified category names fastest at the basic level, experts verified categories equally fast at both the basic level and subordinate level. This proved that the status of the basic level cannot be "eliminated" (Tanaka &Taylor 1991: 476). At the same time, this contradicted Rosch et al.' implication that experts verify category names faster at the subordinate level than at the basic level. This was based on their basic first hypothesis. Contrary to Rosch et al., who suggested a downward shift of the basic level, Tanaka & Tversky (1991) proposed an "increase of accessibility of the subordinate level as a function of expertise" (478). Hence, the basic level is not affected by expertise knowledge. Furthermore, Tanaka & Taylor (1991) pointed out that expertise does not necessarily mean that one has to be an expert regarding a certain object. Expertise knowledge can also refer to the daily interaction of the perceiver with an object or to cultural expertise, meaning that a product might be especially important in a culture.

To fully understand the basic level concept, one has to tackle the following question: Regarding the different levels of abstraction, why is the basic level special? Rosch et al. (1976), as well as other studies, pointed out that the basic level carries not only the most information, but that basic level objects are also the most differentiated from objects of contrasting categories. Hence, informativeness and distinctiveness are two crucial components of the basic level. According to Murphy & Lassaline (1997), the basic level's informativeness highlights its usefulness; its distinctiveness limits the number of categories (107). Otherwise, an infinite number of categories would be required. While the subordinate level is extremely informative, it is not very distinctive (for example, peach and nectarine). In contrast, the superordinate level is not very informative, but it is highly differentiated compared to other superodinates (for example, fruit and furniture). This so-called differentiation theory forms the underlying principle for the basic level and was proven by the following two studies (cf. Murphy & Lassaline 1997).

Murphy & Smith (1982) tested the theory for artificial categories while Murphy et al. (1985) analyzed it for natural categories. Both studies used a category naming task and tested how rapidly subjects named the categories of the objects. Both experimenters focused their studies on the question of why do subjects tend to name objects fastest at the basic level, slower at the subordinate level and slowest at the superordinate level. Both studies confirmed the hypothesis: It is not name length, frequency or early learning of basic level objects, but the basic level’s informativeness and distinctiveness that causes the difference in category naming tasks. Murphy et al. (1985) examined in particular the effects of typicality and atypicality on category naming. As Rosch et al. (1976) already speculated and as Murphy & Brownell (1985) proved, atypical members were named faster with the subordinate word than with the basic level name. For example, penguin was recognized faster than bird. How­ever, Murphy et al. (1985) questioned to which extent typicality influences object naming. In their study they provided some suggestions, but they also pointed this out as a shortcoming of their study. With regard to the present study, informativeness, distinctivencess, and typicality effects will be considered as important factors influencing the participants' choice of objects.

Even though research primarily focused on biological and artificial ob­jects, other studies applied the basic level concept to environmental scenes, actions, events, and social traits (Tversky & Hemenway 1983; Rifkin 1985; Morris & Murphy 1990; Cantor & Mischel 1979; Cantor et al. 1980). In this regard, two studies applying the basic level to emotions and products will be presented. Shaver et al. (1987) dealt with emotions and how these are organized mentally. Based on the data from a sorting task, they performed a cluster analysis in order to find out which attributes are grouped together. The results showed that negative or positive emotion formed the superordinate level. Shaver et al. identified five types of emotion (joy, anger, love, sadness, fear) at the basic level. Attributes such as pleasure, shock, and regret were sorted into the subordinate level (1069). This reveals that the basic level can also be applied to a more abstract notion. This study motivated the present research paper applying the basic level concept to abstract notions. The taxonomy of emotion found by Shaver et al. (1987) will be used in the study presented.

Furthermore, Sujan & Dekleva (1987) conducted research about the basic level in the context of products. They hypothesized that product class (for example, car) forms the superordinate level, product type (for instance, sports car) refers to the basic level and a specific brand name (for example, Nissan 300ZX) depicts the subordinate level. To test this, Sujan & Dekleva (1987) selected seven product classes (cars, cameras, tape recorders, video cassette recorders, electronic games, tennis racquets, and foods) and identified product types as well as specific brand names. Similarly to Rosch et al. (1976), they asked the subjects to list attributes of each object and analyzed the inferences. The results showed that the most attributes were found for the product types, fewer characteristics were named for the specific brand name and the least at­tributes were given for the product class. This proved that product types carry the most information, they are most inclusive and, compared to the product class and the brand name, the product type is most differentiated. The findings of this study will be applied in the present study with regard to stimuli on products. However, for the subordinate level, general brand names were chosen, such as Audi or Mercedes, instead of specific brand names. This was due to the difficulty of presenting a specific brand name or model as a picture and distinguishing it visually from the general brand.

In summary, the presented research papers offered a detailed definition of the basic level and helped to form taxonomic stimuli for the present study. Additionally, the papers delivered an insight into important issues, which have to be considered when studying the basic level concept: interaction of the perceiver with the external world, expertise knowledge, prototypicality, informativeness, and distinctiveness. In the present study, the findings on these topics will be of importance when analyzing and explaining the test results. The research paper follows up by applying the basic level concept to online advertising. Some basics on this emerging promotion medium are presented in the following.

2.2 Online Advertising

As an "applied discipline of marketing” (Janoschka 2004: 10), advertising plays a crucial role in a company's communication policy. So far, informing customers about a certain product or service has been done via television, radio or newspapers. However, as mentioned in the introduction, since 2002 the WWW has emerged as a new advertising medium reaching masses of people. Companies in the European Union spent approximately 9.1 billion € on online advertising in 2008, which is 31% more than in 2007. In the United States, spending on online advertising rose by 46% to 13.6 billion € in 2008. The increasing importance of online advertising becomes apparent when comparing global turnover rates of traditional and online advertising. While the global revenue of TV, radio, newspaper, etc. grew by only 8% in 2008, sales for online ads increased by 23% worldwide (cf. Absatzwirtschaft Online 2008).

Displaying ads on search sites is the most dominant format of online advertising. Although other formats, such as ads embedded in emails or video ads, are utilized, online advertising on search sites amounted to $11 billion in 2008 and it will increase to $12.9 billion in 2009. This represents 40% of the overall US online advertising spending (cf. eMarketer 2008).

These numbers are reason enough to look at how advertising on search sites works and why it is so successful. Basically, search sites serve two purposes. On one hand, they fulfill a seeking function: Users are driven by an information need and, hence, they place a search query to find the necessary information. This information retrieval is not always informational, but it can also be navigational or transactional. When the information need is transactional, the user is interested in carrying out an action, such as downloading a file or completing an online purchase. If the search query belongs to the navigational category, the user is looking for a specific web site. If the user seeks information on a topic, s/he will conduct an informational search query (cf. Border 2002). The search results which are returned according to the user's search query are called organic search results. They are selected algorithmically and are non-paid. On the other hand, search sites fulfill an advertising function. Depending on the search terms, paid sponsored search results show ads to the users. These ads are specifically targeted to a user's search query and highly relevant to the user. By linking the advertised product or service to a user’s search query via the search term, a search engine site manages to provide contextually interesting ads to the user and to reach only relevant audiences. This complementary combination of seeking and advertising function makes search sites a customer-oriented advertising medium. Advertising on search sites does not always aim at transactions or sales, it can also contribute to branding, which means to grab a user's attention for a product (cf. Google 2008).

There are several search sites which offer keyword advertising programs, such as Yahoo!, Microsoft and Altavista (Lammenett 2007: 88). However, the most successful keyword advertising program is offered by Google. With a market share of 76%, Google AdWords is far ahead of its competitors. Yahoo!, in contrast, reaches only 3% of the market share (cf. ebrandz 2009).

So far, the most important question has not been tackled: How are online ads triggered? As mentioned in the introduction, advertising on search result pages is keyword-based. An ad is only displayed if the search query is related to the selected keywords of the advertiser. For example, if somebody searches for guitar, ads on music lessons for playing the guitar or ads of shops selling guitars are displayed to the user. Note that 68% of users only look at the first page results and 92% at the first three pages (cf. iProspect 2008). Hence, in order to achieve a good ad performance, it is highly important that the ad is displayed on the first pages. During an auction, it is determined if an ad is shown and at which site the ad appears. One advertiser competes with other advertisers having the same or similar keywords included in their keyword list. Key factors of this process are the keyword's cost-per-click (henceforth CPC) bid and the Quality Score (henceforth QS). CPC implies that the advertiser only pays when a user clicks on the ad. By constituting a maximum CPC bid, the advertiser measures advertising costs and ensures that s/he never pays more than the given bid. An ad's QS ensures that only high quality advertisements are displayed on the search engine's website. The QS is calculated by its clickthrough rate², the relevance of the ad text to the search terms, and the relevance of the keywords to the search query as well as other factors. Having a good QS does not only improve ad rank but also diminishes the CPC.

After being displayed on the search site, the advertiser has to draw the user's attention to his ad to achieve clicks. Since the advertiser pays per click, only interested users and potential customers should click on the ad. Keywords and the ad text are extremely important in attracting a user's interest, capturing their attention, awakening desire and convincing them to click on the ad. Requiring the user's interaction by convincing him to click on the ad sheds immediate light on the ad's performance and on customer's behavior.

3. Eyetracking Study

After giving a short introduction on the study's theoretical setting, the following section represents the main part of the research paper. Firstly, hypotheses are formulated and the design of the study is described. Secondly, I present the study's results following a discussion.

3.1 Hypotheses

The study aims to examine which words Internet users utilize in their search queries. The results are supposed to provide online advertisers with suggestions concerning their selection of keywords.

Considering the features of basic level objects, such as "most inclusive”, "carry the most information" and "most differentiated" (Rosch et al. 1976: 383), Internet users will tend to use basic level words when conducting a search query. Hence, the basic level advantage examined in previous research is expected to become apparent in the context of online advertising. This was tested for search queries concerning objects, products/brands and abstract notions. Participants were split up in two separate groups. Each of them was confronted with a different test condition. While one group of participants made their decision by looking at pictures, the other group was presented with words. The hypotheses below should be confirmed for both the word and the picture group. Primarily, presenting participants with pictures was supposed to ensure that subjects' decisions were independent from the written text of words. Furthermore, it gave some information on the effects of using image ads.

Since decision making displays a cognitive process, eyetracking was a suitable method to use here. It gives an insight into participants' cognitive practices. Note that the human eye can only receive detailed information from 2° of its visual field covering 200° (Richardson & Spivey 2004: 2). Within this so-called foveal region, objects are fixated and perceived clearly. The eyetracker monitors saccades, fast eye movements, and fixations. The interplay between these reflect participants' interaction with the stimuli. By means of the basic level concept, the following hypotheses will be tested:

H1a: For stimuli of the object category, participants will strongly prefer words/pictures at the basic level to super- and subordinate objects.

H1b: In the object category, pictures/words at the superordinate level will be chosen the least.

H2a: For stimuli of the brand category, basic level objects will be selected most frequently by participants.

H2b: Compared to the object category, the number of chosen subordinates in the brand category will greatly increase. Words/pictures at the superordinate level are chosen the least of all levels of abstraction.

H3a: In the abstract category, subjects will mostly search for words/pictures at the superordinate level.

H3b: Objects at the subordinate will be selected the least for the stimuli of the abstract category.

The following table summarizes the hypotheses for each category. Levels of abstraction are ranked top-down. Levels, whose objects were returned at most, take first place and levels, whose objects were rarely selected as search terms, take last place. The symbol "t" indicates that, in comparison, more words/pictures at the subordinate level were chosen in the brand category than at the same level in the object group.

illustration not visible in this excerpt

Table 1: Summary of hypotheses

3.2 Method

Material The test was composed of 22 stimuli. Every stimulus consisted of a short story and one taxonomy (see section 2.1 for a definition of taxonomy). Each story ended with a question asking the participant to search for something. Seven possible answers were given per story. These were arranged in the form of a taxonomy. While the stories were the same for participants of both groups, taxonomies differed regarding the presentation of the answers. One group of participants saw the answers in written text. These subjects formed the word group. For another group of participants, answers were presented in form of pictures. This group was called picture group. The taxonomic structure of the answers was as follows: one superordinate, three basic level names and three objects at the subordinate level. Content and arrangement of the taxonomies were the same for the word group and the picture group. Most of the pictures were taken from the Internet and were then edited by the vector graphics editor inkscape in order to adapt them visually to the other pictures of the taxonomy. Other pictures had to be newly created. To minimize the influence of colors on the participant's answer, the pictures within one taxonomy were either all black and white or all colored. The arrangement of taxonomies between stimuli varied. For example, some taxonomies were aligned top-down while others were ordered bottom-up. This should avoid subjects developing systems or rules for which word or picture to choose.

Of the 22 stimuli, 12 were experimental stimuli and 10 were filler stimuli. For the word and the picture group, test stimuli were divided up equally in three categories: object category, product/brand category, and abstract category. Taxonomies of the object category referred to the following topics: Getrank, Mobel, Tier, Essen. The arrangements dealt with (non-)biological objects. The­se were set according to the findings of Rosch et al. (1975). For example, the taxonomy Tier consisted of the superordinate Tier, three basic level objects Hund, Katze, Fisch and three names at the subordinate level Dackel, Pudel, Mops. Taxonomies of the product/brand category (Auto, Schuh, Fast Food, Notebook) were concerned with the product class at the superordinate level, for instance Auto. The product type, for example Van, Kleinwagen, Cabrio formed the basic level. Brand names, such as Opel, Audi, Mercedes, built the subordinate level. The structure of the arrangements was drawn from the previously presented study of Sujan & Dekleva (1985). However, for the following study, general brand names, such as Audi, were preferred to specific model names, such as Audi A3. This was due to the difficulty of identifying specific brand names when presented as pictures. Taxonomies of the abstract category focused on conceptual notions: Emotion, Liebe, Religion, Staatsform. According to the findings of Shaver et al. (1985), they were arranged such that the abstract notion, such as Emotion, formed the superordinate level. Forms of the abstract notion, such as Wut, Freude, Angst, built the basic level. Characteristics or detailed shapes of the abstract notion formed the subordinate level, for instance, Angst vor Feuer, Angst vor Spinnen, Angst vor Spritzen.

There were four test stimuli in every category. This amounted to 120 test stimuli in the word group and 120 test stimuli in the picture group. Filler stimuli were not categorized. Three of the filler stimuli showed only the same answers. For the other seven filler stimuli, possible answers were selected randomly, however, the answers were still related to the story. Note that the explications, the stories and the taxonomies, which consisted of words, were in German. This ensured that participants, who were all native speakers of German, fully understood the study.

Participants 20 University of Mannheim students, who all use search engines regularly, participated in the present study. (In fact, more than 20 students were tested because some were discarded due to measuring problems of the eyetracker. However, in the end, 20 participants provided usable results.) Ten students acted in the word group, and the other ten students formed the picture group. They were 21 to 28 years old and came from diverse fields of studies. The students participated voluntarily.

Method As the title states, eyetracking was used for testing. For the following study, the SMI iView X RED was used: a contact-free, remote-controlled eyetracker which measures eye and head movements by an infrared eye camera attached below the computer screen (cf. SMI 2009). The method is sensitive to the shortest and fastest eye movements. Due to technological progress, participants were allowed to move their head freely. Eyetracking enables an authentic and non-disruptive test setup. Additionally, it offers a real-time analysis of eye movements and cognitive processes.

Procedure Since only one eyetracker was available, participants were tested one after another. Each participant was seated in front of the computer with the eyetracker attached below the computer screen. The procedure differed slightly for the word group and the picture group. Contrary to the word group, the picture group had to be prepared for the visual stimuli. Hence, a priming task had to take place before the actual test. Some pictures were not easy to identify, such as Partnerschaft or Demokratie. Participants had to learn which picture depicted which object beforehand in order to avoid misunderstandings during the test.

Before starting the test, each participant in the picture group watched a 7:06 min. long slideshow. Every picture, which appeared in the test, was shown for three seconds. Each picture was presented in the form of a card such that the picture itself showed up on the left half and the adequate word showed up on the right side of the card. After the slideshow, the eyetracking study started.

Participants were calibrated in order to calculate deviations of the participants' eyes. This is a crucial step when conducting an eyetracking study. Deviations between 0° and 0.5° on the x- and y-axis were considered as excellent; deviations between 0.5° and 1° were still acceptable. Participants with a deviation larger than 1° were not considered for the following study. After the calibration, participants were presented a short introduction explaining the test. The opening text already told the subjects that they were going to see pictures. Furthermore, they were informed that there were neither wrong nor right answers, but that the answers solely reflect personal decisions. The introduction was followed by two practice items in order to familiarize the participant with the eyetracker and to illustrate the test procedure. Furthermore, the examples provided the possibility to clarify misunderstandings and to answer questions, if there were any. Before the test started, subjects were asked another time if they comprehended everything to make sure that the test could be carried out without interruptions. During the test, subjects were rotationally presented with a story and the appropriate taxonomy. Trials of the object, the brand and the abstract category as well as fillers alternated so that two of the same group never followed one another. Subjects were asked to look at the pictures and decide which object to choose by fixating on this object long, for a few seconds. After this long fixation, they should immediately turn to the next stimuli. Between and within the stimuli, participants proceeded manually by pressing a button. This avoided time pressure on the subjects and enabled them to take as much time as they needed to read the story and to decide which object to choose. The overall test length varied with each subject; however, on average, it lasted 20 to 25 minutes.

[...]

---

¹ "Cue validity is a probabilistic concept; the validity of a given cue x as a predictor of a given category y (the conditional probability of y/x) increases as the frequency with which cue x is associated with category y increases and decreases as the frequency with which cue x is asso­ciated with categories other than y increases” (Rosch et al. 1976: 384).

² The clickthrough rate (CTR) is the number of clicks an ads receives divided by the number of times the ad is shown (cf. Google 2008).