=Paper=
{{Paper
|id=Vol-1458/G04_CRC30_Blank
|storemode=property
|title=Company Search - When Documents are Only Second Class Citizens
|pdfUrl=https://ceur-ws.org/Vol-1458/G04_CRC30_Blank.pdf
|volume=Vol-1458
|dblpUrl=https://dblp.org/rec/conf/lwa/BlankBH15
}}
==Company Search - When Documents are Only Second Class Citizens==
https://ceur-ws.org/Vol-1458/G04_CRC30_Blank.pdf
Company Search — When Documents are only
Second Class Citizens
Daniel Blank, Sebastian Boosz, and Andreas Henrich
University of Bamberg, D-96047 Bamberg, Germany,
firstname.lastname@uni-bamberg.de,
WWW home page: http://www.uni-bamberg.de/minf/team
Abstract. Usually retrieval systems search for documents relevant to
a certain query or—more general—information need. However, in some
situations the user is not interested in documents but other types of
entities. In the paper at hand, we will propose a system searching for
companies with expertise in a given field sketched by a keyword query.
The system covers all aspects: determining and representing the expertise
of the companies, query processing and retrieval models, as well as query
formulation and result presentation.
Keywords: Domain specific search solutions, expertise retrieval
1 Motivation
The idea for the presented search engine came up when the association of IT
companies in Upper Franconia tried to figure out the opportunities to set up a
register of competencies for their members and associated companies. The basic
idea was: Whenever a company with a demand for IT solutions is looking for
a potential provider of services or solutions the “register of competencies” will
easily expose good matches.
The first idea was to conduct a survey among the companies and derive a
brochure or website. However, the foreseeable problems with low return rates
as well as update and maintenance problems led to a broader consideration of
potential approaches. The next thing that came into mind was “search engines”.
A closer look at commercial and freely available candidates made clear that
these approaches were not convincing for the intended target group. The main
reason based on small experiments was the inappropriate result presentation.
The results consisted of documents and in many situations it was rather unclear
what qualified the documents for top ranks in the result list and—even more
important—which company was associated with the document.
Consequently, the next step was the design of a special purpose search engine
optimized for the “company search task”. In fact, this scenario can be envisaged
Copyright c 2015 by the paper’s authors. Copying permitted only for private and
academic purposes. In: R. Bergmann, S. Görg, G. Müller (Eds.): Proceedings of
the LWA 2015 Workshops: KDML, FGWM, IR, and FGDB. Trier, Germany, 7.-9.
October 2015, published at http://ceur-ws.org
361
List with
• Name
Web-Site
• URL WWW
Crawling
• Address
Index Time
Text
LDA
Processing
Indexing
Company Web Page Auxiliary
Index Index Data
Query Time
Query Result
Processing Preparation
Fig. 1. System overview
as a special case of expertise retrieval—a topic intensely considered in literature
[3]. The contribution of the paper at hand in this context is the design and
reflection of a system based on state-of-the-art models and components adapted
for a real world application scenario. As we will see this requires some peculiar
design decisions and user interface aspects.
The remainder of the paper is organized as follows: In the two subsections
of section 1 below we present a rough overview of the proposed system and we
shortly address related work. Thereafter we discuss the four components identi-
fied to make up an expertise retrieval system according to Balog et al. [3, p. 145]:
modeling and retrieval (section 2), data acquisition (section 3), preprocessing and
indexing (section 4), as well as interaction design (section 5). Finally section 6
concludes the paper.
System Overview The basic assumption of the system is that a manually
defined set of companies should be searchable in the system. These are the
members and associated companies of the IT-Cluster Upper Franconia. A further
assumption is that all these companies have a more or less expressive website.
Hence, the starting point for the system is a list of companies, consisting of
the name of the company, the URL of the corresponding website and the office
address (represented in the upper left corner of Fig. 1).
362
The URLs are used to crawl the websites of the companies. Roughly spoken,
each company c is represented by the concatenation of the content blocks of its
web pages, called dc . Of course, some text processing is necessary here for noise
elimination, tokenizing, stemming, and so forth.
Since the corpus (consisting of about 700 companies at present) is rather
small and queries might be specific (for example a search for “Typo3”) we in-
corporated topic models (using Latent Dirichlet Allocation currently) to boost
companies with a broader background in the topic of the query. Using the terms
and LDA-based boost-factors two indexes are build: In the first index the com-
panies are indexed based on the pseudo documents dc . In the second index the
single web pages are indexed because we also want to deliver the best landing
pages for the query within the websites of the ranked companies in the result.
Finally some auxiliary data (for example the topic models generated via LDA)
is stored since it is needed during query processing or result presentation.
When a query is issued the query is processed on the company index and on
the web page index. Then a result page is generated which represents companies
as first class citizens. For each company a home page thumbnail, a characteriza-
tion of its competencies and its relationship to the query, as well as up to three
query related landing pages are presented. All these aspects will be considered
in more detail in the remainder of this paper, but beforehand we want to shortly
address related work.
Related Work To our best knowledge, there is no directly related work on
company search. The two most related areas are expertise retrieval [3] and en-
tity search [10]. Many lessons can be learned from these areas. Nevertheless,
there are some peculiarities with our company search scenario. In expert finding
scenarios the identification of the experts is often a hard problem (see the expert
search in the TREC 2007 enterprise track as an example [1]). Another aspect
is the ambiguity of names or the vague relationship between persons and docu-
ments. On the other hand, representing experts by pseudo documents is also an
established approach in expert search [2] and an elaborate result presentation is
important here as well.
2 Modeling and retrieval
When thinking about the retrieval model for the given scenario on a higher level,
a model of the competencies of a company has to be matched with the query
representing the user’s information need. From the requirements it was defined
that a keyword query should be used. With respect to the representation of
the company profiles interviews showed that an automatic extraction process is
preferable to the manual definition of profiles because of the sheer creation effort
and update problems. Due to the addressed domain of IT companies it can be
assumed that all companies worth to be found maintain a website depicting their
competencies. Of course other sources of evidence could also be addressed—such
as newspaper reports, business reports, or mentions of the companies on other
363
pages in the Internet. These approaches are surely worth consideration in the
future. Nevertheless, the concentration and restriction to the company’s own
website also has the advantage of predictability and clear responsibilities. Put
simply, if a company complains that it is not among the best matches for a
particular query, we can pass the buck back and encourage them to improve
their website—what they should do anyway because of SEO considerations.
To avoid our arguments eventually turning against us, we have to exploit
the information on the websites as good as we can. Besides crawling and pre-
processing aspects addressed in the following sections 3 and 4, in particular we
have to use an appropriate retrieval model. As a first decision we have to choose
between a company-based approach (each company is represented by a pseudo
document used directly for ranking) and a document-based approach (the docu-
ments are ranked and from this ranking a ranking of the companies is derived).
A comparison of these approaches can, for instance, be found in [3], however not
showing a clear winner. We plan to test both approaches in the future but we
started with the company-based approach where each company c is represented
by a pseudo document dc generated as the concatenation of the content blocks
of the web pages crawled from the respective website. In the future, we plan to
test weighting schemes based on the markup information, the depth of the single
pages, and other parameters.
For a more formal look we use the following definitions:
– q = {w1 , w2 , . . . wn } is the query submitted by the user (set of words)
– C = {c1 , c2 , . . . cm } is the set of companies
– dc is the concatenation of the documents representing company c
– fw,dc is the number of times w appears in dc
– cfw is the number of companies for which dc contains w
– λ and µ are design parameters
Following [6] we use a candidate generation model and try to rank the com-
panies by P (c|q), the likelihood of company c to be competent for query q. As
usual, by invoking Bayes’ Theorem, this probability can be refactored as follows
[3]:
P (q|c)P (c) rank
P (c|q) = = P (q|c)P (c) ≈ P (q|dc )P (c)
P (q)
Currently, we use a constant for the company prior P (c). However, it turned out
that this will be an interesting point for future optimizations highly correlated
with aspects of document length normalization for the pseudo documents dc . For
test purposes we inserted big national and international IT companies in the list.
In our current implementation these companies did not make it to the absolute
top ranks even for queries simply consisting of a registered product name of the
company. Instead, small service providers which have specialized in support for
this product were ranked higher. Interestingly, this problem is already an issue
with expert finding, but an even bigger challenge in company search because of
the heterogeneous company sizes.
Another point which became obvious in first experiments was the well known
vocabulary mismatch. For example with the query “Typo3” the ranking did
364
not consider the broader competence of the companies in the topics of web
applications or content management systems. As proposed in [4, 5] we decided to
use Latent Dirichlet Allocation (LDA) to address this problem. AnQ independence
assumption to calculate the probabilities wordwise by P (q|dc ) = w∈q P (w|dc )
and a combination of the word-based perspective with a topic-based one would
then lead to:
fw,dc + µ cf w
!
|C|
P (w|dc ) = λ + (1 − λ)Plda (w|dc )
|dc | + µ
Plda (w|dc ) stands for the probability that a word w is generated by a topic
which is generated by dc (see [5, 8] for more details). To simplify things further,
we employed an idea presented in [9]. Here the Lucene Payloads are used to boost
terms via LDA. The payload lda(w, dc ) assigned to word w is determined as the
weight of w according to the topic distribution of dc . This means that lda(w, dc )
is high when w fits well with the broader topics dealt with in dc . Combining this
boost factor with term frequency and inverse document frequency information
we yield the following scoring function:
X
score(c, q) = tf (w, dc ) · idf (w) · lda(w, dc )
w∈q
Of course, this is only a first pragmatic starting point and the above consid-
erations point out various interesting aspects for future comparisons and opti-
mizations.
3 Data acquisition
As a prerequisite documents representing companies have to be obtained first.
For crawling company websites we chose to employ crawler4j, a lightweight Java
web crawler (https://github.com/yasserg/crawler4j).
The crawling of each company is an individual process which allows us to
crawl multiple companies at once. We start with the company’s home URL as
a seed and use a truncated form of that URL as a pattern to discard all links
to external domains found during the process. For our first investigation, we
crawled a maximum amount of 2000 documents per company in a breadth first
manner, where each document is a web page. We plan to leverage additional
document types, such as PDF, in the future.
For each page the corresponding company, page title and full URL are stored
in a database. This information is reused later when creating the web page
indexes. To obtain dc (the pseudo document describing company c) the contents
of all crawled pages of the company are concatenated. To reduce noise, we apply
Boilerpipe [7] (https://code.google.com/p/boilerpipe) to all documents in
order to extract the main textual content from those pages first. This step aims
to eliminate those page elements which do not contribute to the actual content
of a page and are repeated very often: navigation menus, footer information, etc.
365
4 Preprocessing and indexing
Early experiments have shown that data quality plays a seminal role for the
quality of a topic model learned. That is why we utilize a customized Lucene An-
alyzer before applying the LDA to dc or indexing the company documents. The
analyzer filters German stop words, applies a Porter Stemmer for the German
language and uses a series of regular expressions to remove or modify tokens. As
an example, digit-only tokens are removed, while tokens of the form word1:word2
are split into two tokens, word1 and word2. Consistently, incoming user queries
are processed by the same analyzer.
After the analyzing step, an LDA topic model of all company representations
dc is created, utilizing the jgibbsLDA (http://jgibblda.sourceforge.net/)
implementation. The resulting model is represented in a Java class hierarchy,
which enables us to directly access the distribution of topics for each company,
as well as the word probability distributions within topics. Therefore the payload
function lda(w|dc ) for each word w in dc can be computed immediately. Another
representation of dc is created, where each term is enriched with its determined
LDA payload. The generated LDA model is reused for result preparation.
The company index is created from all pseudo documents dc enriched with
payloads. When executing a query it is examined by the index searcher and
consequently determines the ranks of the companies in the result set. To be able
to show a query’s top documents for a given company, we also create an index
for the companies’ web pages. All crawled web pages are considered and for each
page we also preserve the information of the corresponding company. Both types
of indexes are based on Lucene (https://lucene.apache.org/core/). Prior to
indexing we apply the analyzing process described above.
With the creation of a company index representing companies and their com-
petencies, a web page index for the companies as well as the overall topic model,
all steps necessary to enable searching are completed.
5 Interaction design
As usual the query is given as a simple keyword query. In the future more so-
phisticated variants are conceivable, for example allowing for geographic filter
constraints. Nevertheless, the simple and familiar keyword solution has its ad-
vantages and the use of geographic filter constraints is debatable as long as only
companies located in Upper Franconia are listed, anyway.
With respect to the result presentation the situation is more demanding.
Discussions with potential users disclosed the following requirements: (1) Com-
panies are the main objects of interest. (2) Address information, a first overview,
and a visual clue would be nice. (3) The general company profile as well as the re-
lationship to the query should become obvious. (4) Entry points (landing pages)
for the query within the website of the company are desired.
The result page depicted in Fig. 2 directly implements these requirements.
Companies are ranked with respect to the retrieval model described in section 2.
366
Fig. 2. Result page for the query “Typo3”
For each company in the result list a row with the name and three information
blocks is shown. The company name is directly taken from the input data as well
as the address information (Fig. 1 upper left corner). A screenshot of the home-
page (captured with Selenium; http://www.seleniumhq.org/) and a prepared
link to OpenStreetMap complete the left overview block for each company. In
the middle block a word cloud is given. Here the size of the terms represents the
importance of the terms for the company profile (based on tf · idf information).
The color represents the relationship of the terms to the query. Orange repre-
sents a strong relationship. The relationship is calculated based on a company’s
prevalent LDA topics. Currently, we consider the five top terms of the five topics
with the highest correlation to the query. At most thirty terms are shown in the
word cloud taking terms important for the company profile and important for
the relationship to the query in a round robin procedure. Finally, the right block
consists of up to three most relevant landing pages within the company website
represented by their title, the URL, and a query-dependent snippet.
367
6 Conclusion
In this paper we have described the company search problem and presented
a solution based on pseudo document ranking, the use of LDA to incorporate
topical relevance, and a suitable result presentation. Currently the prototype im-
plementation is tested. It turned out that the effectiveness and the efficiency are
promising in preliminary interviews with representatives of local IT companies.
Current response times of the system are below two seconds. The most obvi-
ous challenges are the appropriate ranking of companies with different sizes, the
visualization of the company profiles in the result page, and a reasonable mod-
eling and presentation of topics (number of topics in LDA and also alternative
approaches). The current prototype is available on the project web page1 .
References
1. Bailey, P., De Vries, A.P., Craswell, N., Soboroff, I.: Overview of the TREC-2007
enterprise track. In: The Sixteenth Text REtrieval Conference (TREC 2007) Pro-
ceedings. NIST Special Publication: SP 500-274 (2007)
2. Balog, K., Azzopardi, L., de Rijke, M.: Formal models for expert finding in en-
terprise corpora. In: Proceedings of the 29th Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, SIGIR ’06, pp.
43–50. ACM, New York, NY, USA (2006). DOI 10.1145/1148170.1148181
3. Balog, K., Fang, Y., de Rijke, M., Serdyukov, P., Si, L.: Expertise retrieval. Found.
Trends Inf. Retr. 6(2–3), 127–256 (2012). DOI 10.1561/1500000024
4. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. The Journal of
Machine Learning Research 3, 993–1022 (2003)
5. Croft, W.B., Metzler, D., Strohman, T.: Search Engines: Information Retrieval in
Practice. Pearson Education (2009)
6. Fang, H., Zhai, C.: Probabilistic models for expert finding. In: Pro-
ceedings of the 29th European Conference on IR Research, ECIR’07,
pp. 418–430. Springer-Verlag, Berlin, Heidelberg (2007). URL
http://dl.acm.org/citation.cfm?id=1763653.1763703
7. Kohlschütter, C., Fankhauser, P., Nejdl, W.: Boilerplate detection using shallow
text features. In: Proceedings of the Third ACM International Conference on Web
Search and Data Mining, WSDM ’10, pp. 441–450. ACM, New York, NY, USA
(2010). DOI 10.1145/1718487.1718542
8. Wei, X., Croft, W.B.: LDA-based document models for ad-hoc retrieval. In: Pro-
ceedings of the 29th annual international ACM SIGIR conference on Research and
development in information retrieval, pp. 178–185. ACM (2006)
9. Zhang, M., Luo, C.: A new ranking method based on latent dirichlet allocation.
Journal of Computational Information Systems 8(24), 10,141–10,148 (2012)
10. Zhou, M.: Entity-centric search: querying by entities and for entities.
Dissertation, University of Illinois at Urbana-Champaign (2014). URL
http://hdl.handle.net/2142/72748
1
http://www.uni-bamberg.de/minf/forschung/firmensuche
368