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 eld 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.
List with • Name • URL • Address
Web-Site Crawling
Text Processing
LDA
Indexing Company
Index
Web Page
Index
Auxiliary
Data
Query Processing
Result
Preparation
as a special case of expertise retrieval|a topic intensely considered in literature
[
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 identied 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 de ned 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 o ce address (represented in the upper left corner of Fig. 1).
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 speci c (for example a search for \Typo3") we incorporated 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 rst index the companies 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 rst class citizens. For each company a home page thumbnail, a characterization 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 [
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 de ned that a keyword query should be used. With respect to the representation of the company pro les interviews showed that an automatic extraction process is preferable to the manual de nition of pro les because of the sheer creation e ort 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 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
preprocessing aspects addressed in the following sections 3 and 4, in particular we
have to use an appropriate retrieval model. As a rst 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
documents are ranked and from this ranking a ranking of the companies is derived).
A comparison of these approaches can, for instance, be found in [
For a more formal look we use the following de nitions: { q = fw1; w2; : : : wng is the query submitted by the user (set of words) { C = fc1; c2; : : : cmg 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 { and are design parameters
Following [
P (cjq) = P (qjc)P (c) ra=nk P (qjc)P (c)
P (q)
P (qjdc)P (c) 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 nding, but an even bigger challenge in company search because of the heterogeneous company sizes.
Another point which became obvious in rst experiments was the well known
vocabulary mismatch. For example with the query \Typo3" the ranking did
not consider the broader competence of the companies in the topics of web
applications or content management systems. As proposed in [
P (wjdc) = fw;dc + jdcj + cfw ! jCj + (1 )Plda(wjdc)
Plda(wjdc) stands for the probability that a word w is generated by a topic
which is generated by dc (see [
w2q
Of course, this is only a rst pragmatic starting point and the above considerations point out various interesting aspects for future comparisons and optimizations. 3
As a prerequisite documents representing companies have to be obtained rst. 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 rst investigation, we crawled a maximum amount of 2000 documents per company in a breadth rst 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 [
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 Analyzer before applying the LDA to dc or indexing the company documents. The analyzer lters 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(wjdc) 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 competencies, a web page index for the companies as well as the overall topic model, all steps necessary to enable searching are completed. 5
As usual the query is given as a simple keyword query. In the future more sophisticated variants are conceivable, for example allowing for geographic lter constraints. Nevertheless, the simple and familiar keyword solution has its advantages and the use of geographic lter 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) Companies are the main objects of interest. (2) Address information, a rst overview, and a visual clue would be nice. (3) The general company pro le as well as the relationship 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. 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 homepage (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 pro le (based on tf idf information). The color represents the relationship of the terms to the query. Orange represents a strong relationship. The relationship is calculated based on a company's prevalent LDA topics. Currently, we consider the ve top terms of the ve topics with the highest correlation to the query. At most thirty terms are shown in the word cloud taking terms important for the company pro le 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.
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 implementation is tested. It turned out that the e ectiveness and the e ciency are promising in preliminary interviews with representatives of local IT companies. Current response times of the system are below two seconds. The most obvious challenges are the appropriate ranking of companies with di erent sizes, the visualization of the company pro les in the result page, and a reasonable modeling and presentation of topics (number of topics in LDA and also alternative approaches). The current prototype is available on the project web page1.