EXPLORATORY SEARCH ON THE MOBILE WEB

G¨unter Neumann

and Sven Schmeier

DFKI - German Research Center for Artiﬁcial Intelligence, Stuhlsatzenhausweg 3, 66119 Saarbr¨ucken, Germany

DFKI - German Research Center for Artiﬁcial Intelligence, Alt-Moabit 91c, 10559 Berlin, Germany

Keywords:

Web mining, Information extraction, Topic graph exploration, Mobile device.

Abstract:

We present a mobile touchable application for online topic graph extraction and exploration of web content.

The system has been implemented for operation on a tablet computer, i.e. an Apple iPad, and on a mobile

device, i.e. Apple iPhone or iPod touch. The topics are extracted from web snippets which are determined by

a standard search engine. We consider the extraction of topics as a speciﬁc empirical collocation extraction task

where collocations are extracted between chunks combined with the cluster descriptions of an online clustering

algorithm. Our measure of association strength is based on the pointwise mutual information between chunk

pairs which explicitly takes their distance into account. These syntactically–oriented chunk pairs are then

semantically ranked and ﬁltered using the cluster descriptions. An initial user evaluation shows that this

system is especially helpful for ﬁnding new interesting information on topics about which the user has only a

vague idea or even no idea at all.

1 INTRODUCTION

Searching the web using standard search engines

is still dominated by a passive one–tracked human-

computer interaction: a user enters one or more key-

words that represent the information of interest and

receives a ranked list of documents. However, if the

user only has a vague idea of the information in ques-

tion or just wants to explore the information space,

the current search engine paradigm does not provide

enough assistance for these kind of searches. The user

has to read through the documents and then eventually

reformulate the query in order to ﬁnd new informa-

tion. This can be a tedious task especially on mobile

devices.

In order to overcome this restricted document per-

spective, and to provide a mobile device searches to

“ﬁnd out about something”, we want to help users

with the web content exploration process in several

ways:

1. We consider a user query as a speciﬁcation of

a topic that the user wants to know and learn

more about. Hence, the search result is basically

a graphical structure of that topic and associated

topics that are found.

2. The user can interactively explore this topic graph

using a simple and intuitive user interface in order

to either learn more about the content of a topic or

to interactively expand a topic with newly com-

puted related topics.

3. Nowadays, the mobile web and mobile touch-

able devices, like smartphones and tablet com-

puters, are getting more and more prominent and

widespread. Thus the user might expect a device-

adaptable touchable handy human–computer in-

teraction.

In this paper, we present an approach of ex-

ploratory web search, that tackles the above men-

tioned requirements in the following way.

In a ﬁrst step, the topic graph is computed on

the ﬂy from a set of web snippets that has been col-

lected by a standard search engine using the initial

user query. Rather than considering each snippet in

isolation, all snippets are collected into one document

from which the topic graph is computed. We consider

each topic as an entity, and the edges are considered as

a kind of (hidden) relationship between the connected

topics. The content of a topic are the set of snippets

it has been extracted from, and the documents retriev-

able via the snippets’ web links.

The topic graph is then displayed either on a tablet

computer (in our case an iPad) as touch–sensitive

graph or displayed as a stack of touchable text on a

smartphone (in our case an iPhone or an iPod touch).

By just selecting a node or a text box, the user can

either inspect the content of a topic (i.e, the snippets

Neumann G. and Schmeier S..

EXPLORATORY SEARCH ON THE MOBILE WEB.

DOI: 10.5220/0003736800820091

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 82-91

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

or web pages) or activate the expansion of the topic

graph through an on the ﬂy computation of new re-

lated topics for the selected node. The user can re-

quest information from new topics on basis of previ-

ously extracted information by selecting a node from

a newly extracted topic graph.

In such a dynamic open–domain information ex-

traction situation, the user expects real–time perfor-

mance from the underlying technology. The re-

quested information cannot simply be pre–computed,

but rather has to be determined in an unsupervised

and on–demand manner relative to the current user

request. This is why we assume that the relevant in-

formation can be extracted from a search engine’sweb

snippets directly, and that we can avoid the costly re-

trieval and processing time for huge amounts of doc-

uments. Of course, direct processing of web snip-

pets also poses certain challenges for the Natural Lan-

guage Processing (NLP) components. Web snippets

are usually small text summaries which are automati-

cally created from parts of the source documents and

are often only in part linguistically well–formed, cf.

(Manning et al., 2008). Thus the NLP components

are required to possess a high degree of robustness

and run–time behavior to process the web snippets in

real–time. Since our approach should also be able to

process web snippets from different languages (our

current application runs for English and German),

the NLP components should be easily adaptable to

many languages. Finally, no restrictions to the do-

main of the topic should be pre–supposed, i.e., the

system should be able to accept topic queries from

arbitrary domains. In order to fulﬁll all these require-

ments, we are favoring and exploring the use of shal-

low and highly data–oriented NLP components. Note

that this is not a trivial or obvious design decision,

since most of the current prominent information ex-

traction methods advocate deeper NLP components

for concept and relation extraction, e.g., syntactic and

semantic dependency analysis of complete sentences

and the integration of rich linguistic knowledge bases

like Word Net.

The paper is organized as follows. In the section 2

we brieﬂy summarize the topic graph extraction pro-

cess.

For the sake of completeness and readability,

we present in section 3 details and examples of the

user interfaces for the iPad and iPhone, respectively.

A major obstacle of the topic graph extraction

process described in section 2 is its purely syntac-

tic nature. Consequently, in section 4, we introduce

a semantic clustering approach that helps to improve

This part of the work has partially been presented in

(Neumann and Schmeier, 2011) and hence will be described

and illustrated compactly.

the quality of the extracted topics. The next sections

then describe details of the evaluation of the improved

topic extraction process (section 5), and present our

current user experience for the iPad and iPhone user

interfaces (section 6). Related work is discussed in

section 7, before we conclude the paper in section 8.

2 TOPIC–DRIVEN

EXPLORATION OF WEB

CONTENT

The core idea is to compute a set of chunk–pair–

distance elements for the N–ﬁrst web snippets re-

turned by a search engine for the topic Q, and to com-

pute the topic graph from these elements.

In gen-

eral for two chunks, a single chunk–pair–distanceele-

ment stores the distance between the chunks by count-

ing the number of chunks in-between them. We dis-

tinguish elements which have the same words in the

same order, but have differentdistances. For example,

(Justin, Selina, 5) is different from (Justin, Selina, 2)

and (Selina, Justin, 7).

Initially, a document is created from selected web

snippets so that each line contains a complete snip-

pet. Each of these lines is then tagged with Part–

of–Speech using the SVMTagger (Gimenez and Mar-

quez., 2004) and chunked in the next step.

The chunker recognizes two types of word chains:

noun chunks and verb chunks. Each recognized word

chain consists of the longest matching sequences of

words with the same PoS class, namely noun chains

or verb chains, where an element of a noun chain be-

longs to one of the predeﬁned extended noun tags. El-

ements of a verb chain only contain verb tags. For En-

glish, “word/PoS” expressions that match the regular

expression “/(N(N|P))|/VB(N|G)|/IN|/DT” are con-

sidered as extended noun tag and for German the ex-

pression “/(N(N|E))|/VVPP|/AP|/ART”. The English

Verbs are those whose PoS tag start with VB (and VV

in case of German). We are using the tag sets from

the Penn treebank (English) and the Negra treebank

(German).

The chunk–pair–distancemodel is computed from

the list of noun group chunks.

This is fulﬁlled

by traversing the chunks from left to right. For

each chunk c

, a set is computed by considering

all remaining chunks and their distance to c

, i.e.,

We are using Bing (http://www.bing.com/) for web

search with N set to max. 1000.

Currently, the mainpurpose of recognizing verb chunks

is to improve proper recognition of noun groups. They are

ignored when building the topic graph, but see sec. 8.

EXPLORATORY SEARCH ON THE MOBILE WEB

i+1

,dist

i(i+1)

), (c

i+2

,dist

i(i+2)

), etc. This is to

be done for each chunk list computed for each web

snippet. The distance dist

of two chunks c

and c

is computed directly from the chunk list, i.e. we do

not count the position of ignored words lying between

two chunks.

Finally, we compute the chunk–pair–distance

model CPD

using the frequencies of each chunk,

each chunk pair, and each chunk pair distance. CPD

is used for constructing the topic graph in the ﬁnal

step. Formally, a topic graph TG = (V,E,A) consists

of a set V of nodes, a set E of edges, and a set A of

node actions. Each node v ∈ V represents a chunk and

is labeled with the corresponding PoS–tagged word

group. Node actions are used to trigger additional

processing, e.g. displaying the snippets, expanding

the graph etc.

The nodes and edges are computed from the

chunk–pair–distance elements. Since the number of

these elements is quite large (up to several thou-

sands), the elements are ranked according to a

weighting scheme which takes into account the fre-

quency information of the chunks and their collo-

cations. More precisely, the weight of a chunk–

pair–distance element cpd = (c

), with D

= {( freq

,dist

),( freq

,dist

),...,( freq

,dist

)}, is

computed based on point–wise mutual information

(PMI, cf. (Turney, 2001)) as follows:

PMI(cpd) = log

((p(c

)/(p(c

) ∗ p(c

)))

= log

(p(c

)) − log

(p(c

) ∗ p(c

))

where relative frequency is used for approximating

the probabilities p(c

) and p(c

). For log

(p(c

))

we took the (unsigned) polynomials of the corre-

sponding Taylor series using ( freq

,dist

) in the k-th

Taylor polynomial and adding them up:

PMI(cpd) = (

∑

k=1

)

) − log

(p(c

) ∗ p(c

))

,where x

freq

∑

k=1

freq

The visualized part of the topic graph is then com-

puted from a subset of CPD

using the m highest

ranked chunk–pair–distance elements for ﬁxed c

. In

other words, we restrict the complexity of a topic

graph by restricting the number of edges connected

to a node.

3 TOUCHABLE INTERFACE FOR

MOBILE DEVICES

Today, it is a standard approach to optimize the pre-

sentation of a web page, depending on the device it is

displayed on, e.g., a standard or mobile web browser.

Obviously, the same should hold for graphical user

interfaces, and in our case, for the user interfaces de-

signed for iPad and iPhone.

More concretely, the usage of a different mode

of presentation and interaction with a topic graph de-

pending on the device at hand, is motivated for the fol-

lowing reasons: For a smartphone the capabilities for

displaying touchable text and graphics on one screen

are limited mainly due to its relatively small screen

size. Our concept for presenting the results consists of

a touchable navigation based user interface which al-

lows us to interact easily by single touch and swiping

gestures. For a tablet computer with larger screens the

intelligent mix of graphics and text makes a software

system most appealing to the user. Hence the result

presentation consists of a touchable topic graph offer-

ing multitouch capabilities like zooming and swiping.

We demonstrate our current solution by the fol-

lowing screenshots which show some results of the

search query “Fukushima” running with the current

iPad and iPhone user interfaces. In section 6 we

present and discuss the outcomes of some user exper-

iments.

3.1 Graph–based User Interface on the

iPad

The screenshot in Fig. 1 shows the topic graph com-

puted from the snippets for the query “Fukushima”.

The user can double touch on a node to display the

associated snippets and web pages. Since a topic

graph can be very large, not all nodes are displayed

(using the technology described in the previous sec-

tion). Nodes which can be expanded are marked by

the number of hidden immediate nodes. A single

touch on such a node expands it, as shown in Fig. 2.

A single touch on a node which cannot be expanded

automatically adds its label to the initial user query

and triggers a new search with that expanded query.

Fig. 2 demonstrates how the topic graph from Fig.

1 has been expanded by a single touch on the node

labeled “earthquake”. Double touching on the node

“fukushima dailchi” triggers the display of associated

web snippets (Fig. 3) and the web pages.

3.2 Text–based User Interface on the

iPhone

The next screenshots (Fig. 4 and 5) show the results

of the same query displayed on the iPhone.

Fig. 4 shows the alternative representation of the

topic graph displayed in Fig. 1. By single touching

an item in the list the next page with associated topics

ICAART 2012 - International Conference on Agents and Artificial Intelligence

Figure 1: The topic graph computed from the snippets for

the query “Fukushima”.

Figure 2: The topic graph from Fig. 1 has been expanded

by a single touch on the node labeled “earthquake”.

to this item is shown. Finally, Fig. 5 presents the

snippets after touching the item “fukushima daiichi”.

Touching one snippet will lead to the corresponding

web page.

Figure 3: The snippets that are associated with the node

label “fukushima dai–ichi” of the topic graph from Fig. 2.

A single touch on this snippet triggers a call to the iPad web

browser in order to display the corresponding web page. In

order to go back to the topic graph, the user simply touches

the button labeled iGNSSMM on the left upper corner of

the iPad screen.

Figure 4: The alternative representation of the topic graph

displayed in Fig. 1 on the iPhone.

EXPLORATORY SEARCH ON THE MOBILE WEB

Figure 5: The snippets after touching the item “fukushima

daiichi”.

4 SEMANTIC–DRIVEN

FILTERING OF EXTRACTED

TOPICS

The motivation for using the chunk–pair–distance

statistics as described in section 2 is the assump-

tion that the strength of hidden relationships between

chunks can be covered by means of their collocation

degree and the frequency of their relative positions in

sentences extracted from web snippets, and as such,

are emphasizing syntactic relationships.

In general, chunking crucially depends on the

quality of the embedded PoS tagger. However, it

is known that PoS tagging performance of even the

best taggers decreases substantially when applied on

web pages (Giesbrecht and Evert, 2009). Web snip-

pets are even harder to process because they are not

necessarily contiguous pieces of texts. For example,

an initial manual analysis of a small sample revealed,

that the extracted chunks sometimes are either incom-

plete or simply wrong. Consequently, this also caused

the “readability” of the resulting topic graph due to

“meaningless” relationships. Note that the decreased

quality of PoS tagging is not only caused by the differ-

ent style of the “snippet language”, but also because

PoS taggers are usually trained on linguistically more

well–formedsources like newspaper articles (which is

also the case for our PoS tagger in use which reports

an F–measure of 97.4% on such text style).

Nevertheless, we want to beneﬁt from PoS tag-

ging during chunk recognition in order to be able to

identify, on the ﬂy, a shallow phrase structure in web

snippets with minimal efforts. In order to tackle this

dilemma, investigations into additional semantical–

based ﬁltering seems to be a plausible way to go.

About the Performance of Chunking Web

Snippets

As an initial phase into this direction we collected

three different corpora of web snippets and analysed

them according to the amount of well–formed sen-

tences and incomplete sentences contained in the web

snippets. Furthermore, we also randomly selected a

subset of 100 snippets from each corpus and manu-

ally evaluated the quality of the PoS tagging result.

The snippet corpora and results of our analysis are

as follows (the shortcuts mean: #s = number of snip-

pets retrieved, #sc = well–formed sentences within the

set of snippets, #si = incomplete sentences within the

snippets, #w = number of words, F(x) = F–measure

achieved by the PoS tagger on a subset of 100 snip-

pets with x words).

Fukushima. This corpus represents snippets

mainly coming from ofﬁcial online news magazines.

The corpus statistics are as follows:

#s #sc #si #w F(2956)

240 195 182 6770 93.20%

Justin Bieber. This corpus represents snippets

coming from celebrity magazines or gossip forums.

The corpus statistics are:

#s #sc #si #w F(3208)

240 250 160 6420 92.08%

New York. This corpus represents snippets coming

from different ofﬁcial and private homepages, as well

as from news magazines. The corpus statistics are:

#s #sc #si #w F(3405)

239 318 129 6441 92.39%

This means that 39% of all tagged sentences have

been incomplete and that the performance of the Pos

tagger decreased by about 5% F–measure (compared

to the reported 97.4%on newspaper). Consequently, a

number of chunks are incorrectly recognized. For ex-

ample, it turned out that date expressions are system-

atically tagged as nouns, so that they will be covered

ICAART 2012 - International Conference on Agents and Artificial Intelligence

by our noun chunk recognizer although they should

not (cf. section 2). Furthermore, the genitive posses-

sive (the “’s” as in “Japan’s president”) was classiﬁed

wrongly in a systematic way which also had a neg-

ative effect on the performance of the noun chunker.

Very often nouns were incorrectly tagged as verbs be-

cause of erroneously identiﬁed punctuation. Thus, we

need a ﬁltering mechanism that is able to identify and

remove the wrongly chunked topic–pairs.

Semantic Filtering of Noisy Chunk Pairs

A promising algorithmic solution to this problem is

provided by the online clustering system Carrot2 (Os-

inski and Weiss, 2008) that computes sensible de-

scriptions of clustered search results (i.e., web doc-

uments). The Carrot2 system is based on the Lingo

(Osinski et al., 2004) algorithm. Most algorithms

for clustering open text follow a kind of “document–

comes–ﬁrst” strategy, where the input documents are

clustered ﬁrst and then, based on these clusters, the

descriptive terms or labels of the clusters are deter-

mined, cf. (Geraci et al., 2006). The Lingo algorithm

actually reverses this strategy by following a three–

step “description–comes–ﬁrst” strategy (cf. (Osinski

et al., 2004) for more details): 1) extraction of fre-

quent terms from the input documents, 2) performing

reduction of the (pre–computed) term–document ma-

trix using Singular Value Decomposition (SVD) for

the identiﬁcation of latent structure in the search re-

sults, and 3) assignment of relevant documents to the

identiﬁed labels.

The speciﬁc strategy behind the Lingo algorithm

matches our needs for ﬁnding meaningful semantic

ﬁlters very well: we basically use step 1) and 2) to

compute a set of meaningful labels from the web snip-

pets determined by a standard search engine as de-

scribed in section 2. According to the underlying la-

tent semantic analysis performed by the Lingo algo-

rithm, we interpret the labels as semantic labels. We

then use these labels and match them against the or-

dered list of chunk–pair–distance elements computed

in the topic extraction step described in section 2.

This means that all chunk–pair–distanceelements that

do not have any match with one of the semantic labels

are deleted.

The idea is that this ﬁlter identiﬁes a semantic

relatedness between the labels and the syntactically

determined chunks. Since we consider the labels as

semantic topics or classes, we assume that the non-

ﬁltered pairs correspond to topic–related (via the user

query) relevantrelationships between semantically re-

lated decriptive terms.

Of course, it actually remains to evaluate the qual-

ity and usefullness of the extracted topics and topic

graph. In the next sections we will discuss two direc-

tions: a) a quantitative evaluation against the recog-

nition of different algorithms for identifying named

entities and other rigid identiﬁers, and b) a qualitative

evaluation by means of the analysis of user experi-

ence.

5 EVALUATION OF THE

EXTRACTED TOPICS

Our topic extraction process is completely unsuper-

vised and web–based, so evaluation against standard

gold corpora is not possible, because they simply do

not yet exist (or at least, we do not know about them).

For that reason we decided to compare the outcome

of our topic extraction process with the outcomes of

a number of different recognizers for named entities

(NEs).

Note that very often the extracted topics corre-

spond to rigid designators or generalized named en-

tities, i.e., instances of proper names (persons, loca-

tions, etc.), as well as instances of more ﬁne grained

subcategories, such as museum, river, airport, prod-

uct, event (cf. (Nadeau and Sekine, 2007)). So seen,

our topic extraction process (abbreviated as TEP) can

also be considered as a query–drivencontext–oriented

named entity extraction process with the notable re-

striction that the recognized entities are unclassiﬁed.

If this perspective makes sense, then it seems plau-

sible to measure the degree of overlap between our

topic extraction process and the recognized set of en-

tities of other named entity components to learn about

the coverage and quality of TEP.

For the evaluation of TEP we compared it to the

results of four different NE recognizers:

1. SProUT(Drozdzynski et al., 2004): The SProUT–

system is a shallow linguistic processor that

comes with a rule–based approach for named en-

tity recognition.

2. AlchemyAPI

: AlchemyAPI–system uses statisti-

cal NLP and machine learning algorithms for per-

forming the NE recognition task.

3. Stanford NER(Dingare et al., 2004): The Stan-

ford NER–system uses a character based Maxi-

mum Entropy Markov model trained on annotated

corpora for extracting NEs.

4. OpenNLP

: A collection of natural language pro-

cessing tools which use the Maxent package to re-

http://www.AlchemyAPI.com

http://incubator.apache.org/opennlp/

EXPLORATORY SEARCH ON THE MOBILE WEB

solve ambiguity, in particular for NE recognition.

We tested all systems with the three snippet corpora

described in section 4.

The tables 1, 2, and 3 show the main results for

the three different corpora; table 4 shows the results

summarised. All numbers denote percentages that

show how many relevant

NEs of the algorithm in

the row could be extracted by the algorithm in the

column. For example, in the dataset “Justin Bieber”

TEP extracted 85.37% of the NEs which have been

extracted by SProUT. AlchemyAPI extracted 75.64%

and StanfordNER extracted 78.95% of the NEs that

have been extracted by SProUT. The numbers with

preceding “#” show the number of extracted NEs.

The following roman numbers are used to denote

the different algorithms: I=SProUT, II=AlchemyAPI,

III=StanfordNER, IV=OpenNLP, and V=TEP.

Table 1: Results for query Justin Bieber.

I II III IV V

I #136 75.64 78.95 78.48 85.37

II 69.01 #143 93.97 86.00 97.17

III 76.71 97.52 #172 92.86 96.09

IV 74.70 89.19 88.52 #196 95.10

V 67.77 79.61 80.66 81.13 #157

Table 2: Results for query Fukushima.

I II III IV V

I #121 81.03 83.61 81.35 87.5

II 80.26 #129 93.46 87.36 98.48

III 85.00 94.59 #131 91.67 92.22

IV 74.65 89.13 85.26 #178 91.58

V 72.93 80.04 83.19 82.26 #132

Table 3: Results for query New York.

I II III IV V

I #175 81.39 88.24 85.15 71.05

II 76.60 #169 93.53 86.51 74.36

III 90.00 95.79 #280 92.35 73.28

IV 84.43 92.72 93.17 #230 83.49

V 81.11 83.90 73.77 79.87 #166

Keeping in mind that our approach always starts

with a topic around which all the NEs are grouped,

i.e. NE recognition is biased or directed, it is hard to

deﬁne a gold standard, i.e. manually annotate all NEs

which are important in a speciﬁc context. In context

of the query “Fukushima” most people would agree

that word groups describing the nuclear power plant

Relevance here means that a NE must occur more than

4 times in the whole dataset. The value has been experi-

mentally determined.

Table 4: Summary for NER Evaluation.

I II III IV V

I #432 79,25 83.6 81.66 81.31

II 75.29 #441 93.65 86.62 90.00

III 83.90 95.97 #583 92.29 87.19

IV 83.90 95.97 583 #604 87.19

V 73.94 81.18 79.21 81.09 #455

disaster clearly are NEs. Some would also agree that

terms like “earthquake” or “tsunami” function as NEs

too in this speciﬁc context. Given a query like “New

York” people probably would not agree that “earth-

quake” should function as a speciﬁc term in this con-

text. Of course there are NEs of generic type like

“persons”, “locations”, or “companies”, but it is ques-

tionable whether they sufﬁce in the context of our

task.

Hence we compared the systems directly with the

results they computed. The main interest in our evalu-

ation was whether the extracted NEs by one algorithm

can also be extracted by the other algorithms. Fur-

thermore, we set a very simple rating scheme telling

us that detected NEs with more occurences are more

important than those with lower frequencies.

The results show that, looking at the numbers and

percentages, no system outperforms the others, which

on the other hand conﬁrms our approach. Please

note that the TEP approach works for query-driven

context-oriented named entity recognition only. This

means that all approaches used in this evaluation

clearly have their beneﬁts in other application areas.

Nevertheless by going into details we saw some

remarkable differences between the results the sys-

tems produced. All systems were able to extract

the main general NEs like locations or persons. For

terms that are important in the context of actuality

and current developments, we saw that the TEP ap-

proach is able to extract more relevant items. In case

of “Fukushima”, the SProUT system did not extract

terms like “eartquake”, “tsunami” or “nuclear power

plant”. Of course this is because the underlying rule-

set has not been developed for covering such types

of terms. The AlchemyAPI and StanfordNER systems

were able to extract these terms but failed in detecting

terms like “accident” or“safety issues”. For “Justin

Bieber” relevant items like “movie”, “tourdates” or

“girlfriend” could not be detected by all systems ex-

cept TEP . For the snippets associated with the query

“New York” all systems identiﬁed the most important

NEs, and differed for less important NEs only.

Last but not least the runtime, which plays an im-

portant role in our system, varied from 0.5 seconds for

Except for the TEP, where we used the PMI as de-

scribed above.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

the SProUT system, to 2 seconds for TEP, 4 seconds

for StanfordNER to 15 seconds for AlchemyAPI.

6 EVALUATION OF THE

TOUCHABLE USER

INTERFACE

For information about the user experience we had 26

testers — 20 for testing the iPad App and 6 for test-

ing the iPhone App: 8 came from our lab and 18 from

non–computer science related ﬁelds. 15 persons had

never used an iPad before, 4 persons have been unfa-

miliar with smartphones. More than 80 searches have

been made with our system and with Google respec-

tively.

After a brief introduction to our system (and the

mobile devices), the testers were asked to perform

three different searches (using our system on the iPad,

iPhone and Google on the iPad/iPhone) by choosing

the queries from a set of ten themes. The queries cov-

ered deﬁnition questions like EEUU and NLF, ques-

tions about persons like Justin Bieber, David Beck-

ham, Pete Best, Clark Kent, and Wendy Carlos , and

general themes like Brisbane, Balancity, and Adidas.

The task was not only to get answers on questions

like “Who is ...” or “What is . ..” but also to ac-

quire knowledge about background facts, news, ru-

mors (gossip) and more interesting facts that come

into mind during the search.

Half of the iPad–testers were asked to ﬁrst use

Google and then our system in order to compare the

results and the usage on the mobile device. We hoped

to get feedback concerning the usability of our ap-

proach compared to the well known internet search

paradigm. The second half of the iPad–testers used

only our system. Here our research focus was to get

information on user satisfaction of the search results.

The iPhone–testers always used Google and our sys-

tem mainly because they were fewer people.

After each task, both testers had to rate several

statements on a Likert scale and a general question-

naire had to be ﬁlled out after completing the entire

test. The tables 5, 6, 7, and 8 show the overall result.

Table 5: System on the iPad.

#Question v.good good avg. poor

results ﬁrst sight 43% 38% 20% -

query answered 65% 20% 15% -

interesting facts 62% 24% 10% 4%

suprising facts 66% 15% 13% 6%

overall feeling 54% 28% 14% 4%

Table 6: Google on the iPad.

#Question v.good good avg. poor

results ﬁrst sight 55% 40% 15% -

query answered 71% 29% - -

interesting facts 33% 33% 33% -

suprising facts 33% - - 66%

overall feeling 33% 50% 17% 4%

Table 7: System on the iPhone.

#Question v.good good avg. poor

results ﬁrst sight 31% 46% 23% -

query answered 70% 20% 10% -

interesting facts 45% 36% 19% -

suprising facts 56% 22% 11% 11%

overall feeling 25% 67% 8% -

Table 8: Google on the iPhone.

#Question v.good good avg. poor

results ﬁrst sight 23% 63% 7% 7%

query answered 70% 20% 10% -

interesting facts 33% 33% 33% -

suprising facts 36% - 27% 37%

overall feeling 25% 33% 33% 9%

The results show that people prefer the result rep-

resentation and accuracy in the Google style when us-

ing the iPad. Especially for the general themes the

presentation of web snippets is more convenient and

easier to understand. The iPhone–testers could be di-

vided into two groups: in case they were unfamiliar

with smartphones the testers preferred our system be-

cause it needs much less user interaction and the re-

sult are more readable. Testers being familiar with

smartphones again prefered the Google style mainly

because they are used to it.

However, when it comes to interesting and supris-

ing facts users enjoyed exploring the results using the

topic graph (iPad) or the navigation based representa-

tion (iPhone/iPod). The overall feeling was in favor

of our system which might also be due to the fact that

it is new and somewhat more playful.

The replies to the ﬁnal questions: How successful

were you from your point of view? What did you like

most/least;? What could be improved? were informa-

tive and contained positive feedback. Users felt they

had been successful using the system. They liked the

paradigm of the explorative search on the iPad and

preferred touching the graph instead of reformulating

their queries. For the iPhone they prefered the result

representation in our system in general and there have

been useful comments for improving it. One main

EXPLORATORY SEARCH ON THE MOBILE WEB

issue is the need of a summary or a more knowledge

based answer to the search query as Google often does

it by offering a direct link to wikipedia as a ﬁrst search

result. This will be part of our future research.

Although all of our test persons make use of stan-

dard search engines, most of them can imagine to use

our system at least in combination with a search en-

gine on their own mobile devices. The iPhone test

group even would use our system as their main search

tool (on the smartphone) when the proposed improve-

ments have been implemented.

7 RELATED WORK

Our approach is unique in the sense that it com-

bines interactive topic graph extraction and explo-

ration on different mobile devices with recently devel-

oped technology from exploratory search, text min-

ing and information extraction methods. As such, it

learns from and shares ideas with other research re-

sults. The most relevant ones are brieﬂy discussed

below.

Exploratory Search. (Marchionini, 2006) distin-

guishes three types of search activities: a) lookup

search, b) searching to learn, and c) investigative

search, where b) and c) are considered as forms of

exploratory search activities. Lookup search corre-

sponds to fact retrieval, where the goal is to ﬁnd pre-

cise results for carefully speciﬁed questions with min-

imal need for examinating and validating the result

set. The learn search activity can be found in situ-

ations where the found material is used to develop

new knowledge and basically involves multiple iter-

ations of search. It is assumed that the returned set

of objects maybe instantiated in various media, e.g.,

graphs, maps or texts. Investigativesearching is a next

level of search activity that supports investigation into

a speciﬁc topic of interest. It also involves multiple

iterations even for very long periods and the results

are usually strictly assessed before they are integrated

into knowledge bases. Our proposed approach of ex-

ploratory search belongs to the searching to learn ac-

tivity. In this spirit, our approach is more concerned

with recall (maximizing the number of possibly rele-

vant associated topics that are determined) than pre-

cision (minimizing the number of possibly irrelevant

associated topics that are determined).

Collocation Extraction. We consider the extrac-

tion of a topic graph as a speciﬁc empirical colloca-

tion extraction task. However, instead of extracting

collocations between words, which is still the dom-

inating approach in collocation extraction research

(e.g., (Baroni and Evert, 2008)), we are extracting col-

locations between chunks, i.e., word sequences. Fur-

thermore, our measure of association strength takes

into account the distance between chunks and com-

bines it with the PMI (pointwise mutual information)

approach (Turney, 2001).

(Geraci et al., 2006) also exploit the beneﬁt of

Web snippets for improved internet search by group-

ing the web snippets returned by auxiliary search en-

gines into disjoint labeled clusters. As we do, they

also consider methods for automatic labeling. How-

ever, their focus is on improving clustering of terms

and not on the extraction of empirical collocations be-

tween individual terms. Furthermore, they advocate

the “document–comes–ﬁrst” approach of clustering

Web snippets which is inappropriate for our method-

ology, cf. sec. 4.

Unsupervised Information Extraction. Web–

based approaches to unsupervised information

extraction have been developed by Oren Etzioni and

colleagues, cf. (Banko et al., 2007); (Etzioni, 2007);

(Yates, 2007). They developed a range of systems

(e.g., KnowItAll, Textrunner, Resolver) aimed at

extracting large collections of facts (e.g., names of

scientists or politicians) from the Web in an unsu-

pervised, domain-independent, and scalable manner.

They also argue for light–weight NLP technologies

and follow a similar approach to chunk extraction

as we do (but not a chunk–pair–distance statistics).

Although we do not yet explicitly extract relations

in the sense of standard relation extraction, our topic

graph extraction process together with the clustering

mechanism can be extended to also support relation

extraction, which will be a focus of our next research.

8 CONCLUSIONS AND

OUTLOOK

We presented an approach of interactive topic graph

extraction for exploration of web content. The initial

information request is issued online by a user to the

system in the form of a query topic description. The

topic query is used for constructing an initial topic

graph from a set of web snippets returned by a stan-

dard search engine. At this point, the topic graph al-

ready displays a graph of strongly correlated relevant

entities and terms. The user can then request further

detailed information through multiple iterations.

A prototype of the system has been realized on the

basis of two specialized mobile touchable user inter-

faces for operation on an iPad and on an iPhone which

receive both the same topic graph data structure as in-

ICAART 2012 - International Conference on Agents and Artificial Intelligence

put. We believe that our approach of interactive topic

graph extraction and exploration, together with its im-

plementation on a mobile device, helps users explore

and ﬁnd new interesting information on topics about

which they have only a vague idea or even no idea at

all.

Our next future work will consider the integration

of open shared knowledge bases into the learn search

activity, e.g., Wikipedia or other similar open web

knowledge sources and the extraction of relations,

and ﬁnally to merge information from these different

resources. We already have embedded Wikipedia’s

infoboxes as background knowledge but not yet in-

tegrated them into the extracted web topic graphs,

cf. (Neumann and Schmeier, 2011) for some more

details. If so done, we will investigate the role of

Wikipedia and the like as a basis for performing dis-

ambiguation of the topic graphs. For example, cur-

rently, we cannot distinguish the associated topics ex-

tracted for a query like “Jim Clark” whether they are

about the famous formula one racer or the Netscape

founder or even about another person.

In this context, the extraction of semantic relations

will be important. Currently, the extracted topic pairs

only express certain semantic relatedness, but the na-

ture and meaning of the underlying relationship is un-

clear. We have begun investigating this problem by

extending our approach of chunk–pair–distance ex-

traction to the extraction of triples of chunks with al-

ready promising initial results.

ACKNOWLEDGEMENTS

The presented work was partially supported by grants

from the German Federal Ministry of Economics and

Technology (BMWi) to the DFKI THESEUS project

(FKZ: 01MQ07016).

REFERENCES

Banko, M., Cafarella, M. J., Soderland, S., Broadhead, M.,

and Etzioni, O. (2007). Open information extraction

from the web. In Proceedings of IJCAI–2007, pp

2670–2676.

Baroni, M. and Evert, S. (2008). Statistical methods for

corpus exploitation. In A. L¨udeling and M. Kyt¨o

(eds.), Corpus Linguistics. An International Hand-

book, Mouton de Gruyter, Berlin.

Dingare, S., Nissim, M., Finkel, J., Grover, C., and Man-

ning, C. D. (2004). A system for identifying named

entities in biomedical text: How results from two eval-

uations reﬂect on both the system and the evaluations.

In Comparative and Functional Genomics 6:pp 77-85.

Drozdzynski, W., Krieger, H.-U., Piskorski, J., Sch¨afer, U.,

and Xu, F. (2004). Shallow processing with uniﬁca-

tion and typed feature structures — foundations and

applications. K¨unstliche Intelligenz, pages 17–23.

Etzioni, O. (2007). Machine reading of web text. In

Proceedings of the 4th international Conference on

Knowledge Capture, Whistler, BC, Canada, pp 1-4.

Geraci, F., Pellegrini, M., Maggini, M., and Sebastiani, F.

(2006). Cluster generation and labeling for web snip-

pets: A fast, accurate hierarchical solution. Journal of

Internet Mathematics, 4(4):413–443.

Giesbrecht, E. and Evert, S. (2009). Part-of-speech tagging

- a solved task? an evaluation of pos taggers for the

web as corpus. In Proceedings of the 5th Web as Cor-

pus Workshop.

Gimenez, J. and Marquez., L. (2004). Svmtool: A gen-

eral pos tagger generator based on support vector ma-

chines. In Proceedings of LREC’04, pp. 43 - 46.

Manning, C. D., Raghavan, P., and Sch¨utze, H. (2008). In-

troduction to information retrieval. In Cambridge Uni-

versity Press.

Marchionini, G. (2006). Exploratory search: from ﬁnding

to understanding. Commun. ACM, 49(4):41–46.

Nadeau, D. and Sekine, S. (2007). A survey of named entity

recognition and classiﬁcation. Journal of Linguisticae

Investigationes, 30(1):1–20.

Neumann, G. and Schmeier, S. (2011). A mobile touchable

application for online topic graph extraction and ex-

ploration of web content. In Proceedings of the ACL-

HLT 2011 System Demonstrations.

Osinski, S., Stefanowski, J., and Weiss, D. (2004). Lingo:

Search results clustering algorithm based on singular

value decomposition. In Proceedings of the Inter-

national IIS: Intelligent Information Processing and

Web Mining Conference. Advances in Soft Computing,

Springer.

Osinski, S. and Weiss, D. (2008). Carrot2: Making sense of

the haystack. In ERCIM News.

Turney, P. (2001). Mining the web for synonyms: PMI-IR

versus LSA on TOEFL. In Proceedings of ECML–

2002. Freiburg, Germany, pp 491-502.

Yates, A. (2007). Information extraction from the web:

Techniques and applications. In Ph.D. Thesis, Uni-

versity of Washington, Computer Science and Engi-

neering.

EXPLORATORY SEARCH ON THE MOBILE WEB