Common search engines deliver quite good results when the user has a precise notion of what he is looking for. However, the user might have in mind additional prior information regarding the importance of speci c terms. Consequently, it seems desirable to avoid the latter and incorporate the knowledge into the query instead. Therefore, we propose a search user interface concept that supports users in modelling their uncertainty in a comfortable way, foster exploratory search and provide a compact yet informative representation of results. An implemented prototype demonstrates the feasibility of the concept. We also present results of a rst twostep usability study. The results indicate a good usability of the concept and show that even this novel concept meets user's expectations.
Modern search engines have become very powerful
tools, providing excellent results - even in areas beyond
basic document queries like nding a nearby dentist or
checking for the weather of next weekend. However,
they require textual input of keywords by the user,
who is not necessarily capable of formulating suitable
terms at the begin of the search, for instance because
he is new to the domain. Search engines also lack in
the ability to support the formulation of importance
of selected search terms as every term has potentially
the same impact on the result (apart from
contextsensitivity). Also functions for explicitly excluding
terms are either hidden - in most cases unknown to
users - or do not exist at all. A study on search query
logs conducted by Jansen et al. [
Another aspect of search engines is the presentation of results. It comprises in most cases just text containing the title, a text snippet and the URL. This gives no clue about the visual appearance of the actual document, which nevertheless could be helpful for a user's relevance estimation and recognition of previously visited websites.
In this paper, we present a concept designed to
overcome these disadvantages of current web search
user interfaces by introducing a novel query
formulation mechanism and a compact representation of a web
page's content as well as visual appearance. Both
being integrated into a search user interface prototype
that addresses some aspects of exploratory search [
The VIBE-system by [
Nitsche & Nurnberger [
The problem of reducing a web page's content to
a compact representation has been addressed in
various publications [
Our system consists of three major components (Fig. 1): A crawler, a backend and a frontend. This section deals with the rst ones, while the frontend will be described separately in section 4.
Crawler. The rich result representation prohibits
the utilization of APIs of common search engines as
they deliver too little information about the web page
and crawling these in real time is infeasible given a
reasonable number of results. Therefore, we
developed our own crawler computing a colour histogram, a
salient extract of the web page and a wordcloud as well
as the text document for indexing. First, a screenshot
of the web page is taken. For the colour histogram only
the top 600 pixels of the web page are considered as
most websites have a characteristic header. The pixel
values are clustered by KMeans [
tf (w)(logjBj + 1); (tf (w)log(jBj=tfB(w)); if w 2 B else (1) With B being the set of words in the corpus and tf respectively tfB the term-frequencies in the web page and in the Brown corpus.
Backend. Our database relies on Apache Solr without any profound modi cations. Besides a typical text eld, we added elds for additional features the crawler captures. Term weighting is implemented by the boosting mechanism of Solr. Communication with the javascript-based frontend is realized via HTTP and JSON encoding, which is natively supported by Solr. 4
The user interface consists of ve main elements, with
query formulation and result representation as the
most innovative ones. Query formulation is placed
at the top of the screen and the result representation
below. Both cover the entire width. Below them a
navigable web page preview is set, surrounded by
navigation buttons to the left and right. The small result
representation and the big preview follow the design
pattern of \overview and detail" [
Query Formulation. In order to constitute a query, terms are rst typed in the simple text-box as usual. After submitting the rst query, all terms move to the left side without overlapping each other. Note that this implies a positive initial weighting. For a re nement and the expression of uncertainty, terms can be moved horizontally. Fig. 3 depicts how the arrangement of terms a ects the query semantics. With x being the position of a term in the interval [ 1:0; 1:0], sgn(x) indicates whether the term is explicitly wanted or unwanted in the result documents. jxj denotes the condence of the former statement. Single terms can be removed by triggering a small remove button that pops up on mouse over. New terms can be added by clicking the query bar at the position associated with the wanted weighting of the new term and simply starting confidence
0 positve terms negative terms -1.0 1.0 to type. It is also possible to restart the entire search, i.e. to remove all terms, by clicking a single button.
Result Representation. Due to the elongated and { compared to common search user interfaces without preview { small result bar, the crucial goal in designing the result representation was to keep it as compact as possible and to allow a horizontal arrangement. Previously seen web pages should be recognizable and the content of unknown web pages should be as obvious as possible when looking at the result representation. Our approach consists of three di erent constituents (Fig. 4):
A colour bar (1) on the left as well as the background of the whole element indicate frequent colours of the respective web page. When colours are known in advance, it allows to quickly redis1 2 parts containing Word3Cloud important most text
cover a previously visited web page because colour
is a pre-attentive attribute [
An extract of the rendered web page's screenshot (2) provides a small preview of the most salient region of the web page. This might also support recognition and could additionally serve as a hint for the web page's topic.
The wordcloud (3) - it's computation is described above - gives a general overview of the website's content by putting an emphasis on words that occur rarely in general but frequently in this document and are hence more likely relevant for the current topic.
By dragging a result representation into the query bar (Fig. 5), the query can be manipulated depending on the results content. When there is an intersection between wordcloud of the element and query terms, they are shifted to the left giving them a more positive e ect. If there is no intersection, the most popular word in the wordcloud is added to the query. This way exploratory search is further supported in the proposed search user interface concept.
Implementation. The implementation of the
described concept is based on current web standards:
Javascript for the logic and SVG using raphael.js1 for
rendering graphics. Since the elements of the result
representation are not retrievable from common APIs,
we had to make a decision between re-crawling the
elements in real-time as soon as results from an API
are delivered or to build our own index with colours,
salient region and wordcloud directly stored. We
decided in favour of the latter as the re-crawling takes
too much time and results could not be presented
instantaneously. The index is based on an unchanged
(except for con guration) Apache Solr2 server. It is
lled with content by our own crawler that captures a
screenshot for colour and a salient region extraction as
well as it processes the html content, ending up with
term scores (see Formula 1) of the wordcloud. The
crawler is implemented in python using nltk [
An open question is how the system reacts in a
larger scale, but as we use Solr for storage and query
handling, we are con dent that the system scales well,
possibly by utilizing Apache Hadoop [
We conducted two evaluations with 17 participants in total, i.e. nine respectively eight participants each: A formative evaluation guided us for some design decisions. A summative one tested the nal prototype implementation. Note that the evaluation was originally carried out in German and translated to English for this paper. 5.1
The entire formative evaluation was implemented as an interactive form, where the study participants has been asked to interact with mock-ups of parts of the later implemented user interface. We o ered a discrete and a continuous version of the query formulation mechanism (Fig. 5) and tried to assess which one is easier to handle. Therefore, we created two challenges:
Query formulation. The rst task involved creating a query given the following brief note about the goals as well as the terms we wanted to be used: You are looking for a destination for your hiking vacations in the mountains, not necessarily in the alps as you have been there before. As you su er from vertigo you want to avoid climbing. The results in Fig. 6 show that the test users were able to formulate a proper query, i.e. putting the relevant terms to the left side of the query and the negative ones to the right side.
Query understanding. To solve the second task, users were asked to do the inverse. Given a nal query formulation, six di erent images needed to be ordered or removed. Five of them were images of cakes, the sixth one was an image of a dog. This way, we wanted to see if the representation of a query in the query bar (Fig. 5) is understandable. Furthermore, it gives insight to a deeper interpretation of the participants: Should the dog be in the result list though it has no relation to the query terms? If yes, should it be placed in front of the un tting results? The results reported in Fig. 7 suggest that the basic principle has been understood as the rightmost images were correctly put top of the results in most cases.
Regarding the dog, the participants agreed on scoring it lower than all cakes. But there is a dissent on whether to include the result or to remove it.
Result representation. In addition to the query formulation, we also evaluated prototypes of the result representation (Fig. 4). Four manually assembled representations of web pages were provided and we asked the study participants for possible search terms, a category of the web page and which traits of the representation were pivotal for that decision. Not all participants lled out all elds. But when they did, they correctly predicted the web page's content, with only one exception. Often, the participants were able to specify even the subtopic. 5.2
The summative evaluation was carried out by giving the participants some task, while observing them and making notices. Afterwards, they were given the opportunity to express feedback.
In general, most users succeeded in working with the search user interface. Minor problems involved confusion between user interface and result representation, colours and interpreting the plus/minus button at the end of the scale as being actually a button. We attribute this to the short time frame the participants had to get used to the prototype and its underlying novel concept. Colours in the result representation indicating page colours are confused with the colour scale for weighting a term. The plus respectively minus icon at the ends of the scale is sometimes mistakenly interpreted as a button. 4Study participants saw copyright protected image. a) b) c) 1 2 3 4 5
Furthermore, we found that the search bar can be seen as a text eld in the user's mental model and could therefore support corresponding interactions (i.e. placing a cursor and editing text). However, all participants considered the term weighting as a useful tool and the majority liked the result representation as well. The ratings shown in Fig. 8 indicate minor problems regarding the transition from common user interfaces while both, result representation and the weighting ability, are for most parts considered as good. 6
We presented a novel search user interface concept for exploratory web search addressing the problem of incorporating uncertainty with respect to user's con dence while searching. The main contributions are a novel query formulation mechanism and a compact visualization. This supports an e cient recognition. It also helps users to concern a web page's topic by linking visual and textual information. The implementation demonstrates the feasibility of the concept and the small evaluation suggests that users are able to properly interact with the interface.
Future work will cover the improvement of the system's usability in practice. For instance, by o ering a function to save interesting web pages and using more elaborative methods for visual and textual information extraction in the crawler. The compact representation of results might also be interesting for mobile use, where screen space is limited.
We thank the ickr users tjstaab, freakgirl, lovebuzz, Kirti Poddar and je ro for releasing their images under a creative commons licence.