This paper presents our attempt to build a text mining tool for collecting specific words - verbs in our case - that usually occur together with particular product category as support for ontology designers. As the ontologies are headstone for the success of the semantic web, our effort is focused on building small and specialized ontologies concerning one product category and describing its frequent relations in common text. We describe the way we use web directories to obtain suitable information about the products from UNSPSC taxonomy and we propose the method how the extracted information could be further processed.
Information Extraction (IE) and Ontology (OL) learning are frequently discussed
issues in the field of Semantic Web. The problems of information extraction using
hand-crafted patterns have been addressed in many papers and it is obvious that the
most promising way is automated or semi-automated ontology-based extraction of
information. Since the results of IE from rigidly structured and semi-structured texts
are already quite satisfying, the problems remain in field of unstructured free text
processing. Large amount of knowledge-sparse text with full linguistic analysis
would be too demanding. Shallow linguistic methods typically rely on POS tagging
and/or shallow parsing. In our work we focus on finding verbs as simple POS
category (in [
extension of a preceding concept in the hierarchy. Consider for example .../Industries/Construction and Maintenance/Materials and Supplies/ /Masonry_and_Stone/Natural Stone/International Sources/Mexico.
Semantic interpretation of a sample of DMOZ paths revealed that: – Terms in the headings belong to quite a small set of classes, such as ‘Object’ (i.e. product such as ‘Car’), ‘Subject’ (e.g. ‘Manufacturer’ or ‘Dealer’), ‘Domain’ (of competence of company, such as ‘Transport’ or ‘Insurance’), ‘Location’ (e.g. ‘Mexico’) etc. – Surface ‘parent-child’ arrangement of headings belonging to particular classes corresponds (with some ambiguity) to ‘deep’ ontological relations.
The idea of closed loop between IE and OL bootstrapped with web directory
headings was first formulated in [
The reason why we use UNSPSC is that we would like to join this taxonomy and
list of products with content of company websites to gain valuable information about
verbs that usually occur in one sentence with some product category from the
taxonomy. We build a tool that collects these verbs from given web pages. Presented text
mining tool is based on combination of catalogue and fulltext search engine. Our
approach is exploits redundancy of data on large data repositories like World Wide
Web. We are exploiting the knowledge stored in hand classified web directories like
DMOZ and we use their ability to provide web sites relevant to term we have chosen.
The problem that had been already discussed in [
In this paper, we first describe the reason why UNSPSC was chosen, and why we use directories as source for our data. In next section we introduce our method to identify verbs related to products and in third section we describe experiments and the results. At the end of the paper related work and our future plans are discussed. 2
2.1
As suggested in [
product category which can hardly be further divided. On the other hand this raise problem that UNSPSC tree leaves (product categories) varies from the directory headings in commonly used directory structure including DMOZ and Google directory. At current time there aren’t any tools to automate the process of assigning right UNSPSC category to relevant DMOZ category so it must be done manually by choosing product from taxonomy and then finding appropriate category in directory. There are either a lot of categories describing our term or none. In the first case we focus on Business branch where we expect that the manufacturers and the company offering our products will be stated. The latter case – where no category is found – is worse and we have to find similar category, or find similar product. These two issues disallow this part of our work to be done automatically. In our test we found 7 nodes corresponding to the same number of products from UNSPSC from “Material handling” field. 2.2
We take advantage of human-classified web page links stored in web directories. As stated above their structure is not always valid taxonomy. Subheadings are often not specialization of headings; some of them are even not concepts (names of entities) but properties that implicitly restrict the extension of a preceding concept in the hierarchy. This is reason why we make use of UNSPSC classification in our paper. We would like to obtain so called „indicator verbs” that characterize particular term (product category in our case) in UNSPSC. Particular terms will be then generalized and may mine verbs that are indicative for the upper level of these terms. The trial was only made on one category and several terms, which limits the representativeness of results. Only several common verbs were obtained and they had to be classified manually, as we don’t have any other categories to be compared with results from this. Next paragraph describes the text mining tool that collects data from selected directory category.
Table 1 depicts sub-tasks of the tool. The input data for this tool are the URL of directory in DMOZ containing links relevant to chosen term and the product category chosen from UNSPSC. When we have chosen the right category the script can be run. The first part uses link extractor to obtain all company’s web sites URLs. The list of extracted links is stored in file for further processing. Every URL from the list is then inserted into Yahoo Search Engine with the term we are currently exploiting and the parameter “site” is added. This ensures that the particular term is only searched on the selected web site. This process is repeated until all URLs from the list have been processed. We only store first 10 links from every domain, but it is only matter of setting of script and here we see a possibility of extracting more data. Up to 100 links from every company URL can be stored.
Now we have several hundreds links (depending on number of links on list) to
sites where our desired term occurs in the page full text. Next part task is to extract
every sentence from this set of links where the terms occur. The task is carried out by
means of regular expressions and finding occurrence of the term in set of documents.
As the sentences are discovered and saved into file we need to carry out some
syntactical analysis to discriminate verbs from other lexical units. It is done by Adwait
Ratnaparkhi's Java based Maximum Entropy POS Tagger (MXPOST) [
word include lemma include announced
announce are
be providing feature following leading products provide feature follow lead are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are, are providing, providing, providing, providing, providing, providing, providing featured, features, features, features, features, features, features, features, features, feature, feature following, following, following, following, following, following, following leading, leading, leading, leading, leading, leading, leading, leading, leading products is products, products, products, products, products, products not a verb including include
including, including, including, including, including, including
We store all extracted verbs as matrix in relational database, where the discovered verbs represents rows and the desired terms represents columns. The intersection of row and column is the number of verb occurrences in all web pages of the directory category. We can use these sets of verbs and number of their occurrences for further examination on how the verbs characterize some broader term and if a small ontology could be build for every term. The goal of effort is to build separate lists of verbs such that some only characterize specific product types, while some other characterize whole product areas. We started with node Handling materials and products from this category. 3
In extracting selected information we currently face the problem of how to set proper amount of web sites to be analyzed and what is the right amount of words taken in surroundings of the word. As stated above in this particular experiment, 10 web pages from one company website and one sentence from each web page containing examined word is taken. Our research shown that this can be sufficient amount of data for extraction of verbs for some common product types, but other product types from the same category of UNSPSC suffer from lack of data for extraction because they don't have appropriate category (node) in DMOZ.
There are two possible ways to overcome this problem. The first one is to concede
the given constrains and allow our script to crawl more pages from one website and
also allow to extract more sentences from one page. The second approach we have on
mind is take advantage of some news resources like Google News service as there
might appear verbs that characterize some product category. But from previous
experience [
Using the above described tool we have built a database containing 303 verbs for 7 product categories from handling material category. These are only words that have appropriate category in DMOZ and therefore our approach could be used for their extraction. These verbs occurred 7300 times near the selected terms.
Our goal is to find some method that would enable us to categorize verbs as either: - common for most products. - characterizing one branch of products - specific for small group of products, or even only one product.
Even from seven product categories – as expected – some verbs are obvious to be entirely neutral and do not characterize the products at all. According to three methods described later, verbs be, have, provide and use are common for all sentences describing any product. Then we have verbs describing activities connected with manufacturing of any types of products e.g. design, require, offer, make, contact, manufacture, develop, supply, etc. More specific for our branch might be verbs describing activities related to manipulating with material. They are handle, lift, install and move.
We experimented with three different measures that could separate specific verbs
from more general ones. First and second are normalizations of frequencies to
eliminate the influence of very frequent verbs. Normalization based on proportions of
product categories in collection is the first, Croft’s normalization using elimination
of high-frequency terms with a specific constant is the second and TF/IDF [
We tried these three methods to class verbs to their corresponding groups of verbs. All methods provided quite similar results. The first is normalization described by formula (1), where Fij is normalized frequency, fij is the frequency of verb j in product category i, Vtj is sum of all occurrences of product category i in collection and V is total number of collected verbs. Then Vtj / V represents how many per cent has product category i in collection. We recalculate whole matrix to get numbers ranging from 0 to 43 representing the normalized frequencies showing that the verbs with high value (30-43 in our case) are independent on the product category and thus they can be considered as common one. Verbs with values from 10 to 30 are not so often and they could be used as branch descriptors. The rest are with frequency lower than 10 are out of our interest for this moment.
Fij = fij * (Vtj / V) . (1)
Croft’s normalization (2) moderates the effect of high-frequency verbs, where cfij is Croft's normalized frequency, fij is the frequency of verb j in product category i, mi is the maximum frequency of any verb in product category i, K is a constant between 0 and 1 that is adjusted for the collection. K should be set to higher value (higher than 0.5) for collections with short documents. We used 0,3 as there are no different between 0.3 and 0.5 in our table. With this formula we get sum values for every verb ranging from 2.1 (7 product category × 0.3 for zero occurrences) for no occurrences of verb in our database to 8.58 for the most often verbs. Verbs with number above 5 normalized occurrences are significant for us as the common indicator while verbs between 3 and 5 normalized occurrences could be taken as the products representing verbs. The rest, with 3 and lower occurrences is for us as in previous method uninteresting.
cf = K + (1 - K) * fij / mij . (2)
TF/IDF (term frequency / inverse document frequency) (3), where wij is a weight of verb in product category i, fij is the frequency of verb j in product category i, N is number of all verbs in collection and n is sum of verb j occurring in all product categories. TF/IDF is technique that gives verb a high rank in a document if the verb appears frequently in a document or the verb does not appear frequently in other product categories. In other words a verb that occurs in a few product categories is likely to be a better discriminator than a verb that appears in most or all categories. As a result in this test we got values from 0 to 1350. Where as usual, the highest values between 1000 and 1350 are verbs that occur independently on selected product category and we consider them as common verbs. We are much more interested in verbs with value starting around 300 and ending at 1000. As stated above, these could be used as identifiers of the product category.
wij = fij * log2(N / n) (3)
In our trial we only examined 7 product categories from one UNSPSC node and hence we are not able to classify verbs into four categories as we suggested in part 1. We only classified them on common and specialized verbs. The first 15 results with values from each of described method are shown in Table 3.
The reason why this approach cannot be automatically run are mainly the
noncorresponding items from product taxonomy to categories in widely-spread product
catalogues. Our plans and intentions for the development of this tool are stated in
future work section.
The idea of combination information extraction with ontology learning has been
described by Maedche in [
As described in this paper, there are currently some limitations of this approach; they are mainly caused by lack of data to be mined from websites for some specialized terms. We proposed some techniques how to overcome these limitations. One of them is relaxing restrictions of fulltext search engines and the second is searching in all subdirectories for given terms in all whole DMOZ branch tree structure. All our plans for future stem from our effort to obtain as much data as possible and also better automation of whole process. We recently discovered a tool that could help us to use fulltext search on selected nodes from DMOZ web directory. As soon as we obtain verbs for more branches we could try to classify the verbs into four categories as proposed in section 3 and use them for creating ontologies with relations labeled with extracted verbs. 6
The authors would like to thank to Martin Labský and Martin Kavalec for their comments and help.
The research has been partially supported by the grant no. 201/03/1318 of the Grant Agency of the Czech Republic „Intelligent analysis of the WWW content and structure“. 7