It is common to organize information in a categorization scheme known as a “taxonomy”. Taxonomies ordinarily are hierarchal constructs of categories as nodes. Information is assigned to a node of the taxonomy based on the content of the information. For example www.yahoo.com organizes documents of various domains in a taxonomy to allow users to search and browse the information by category. taxonomies are somewhat arbitrary in that each may use a different set of categories and may organize those categories in various ways. It is often desirable to “align” two or more taxonomies to create a single browsable document collection. When attempting to align document taxonomies, there are frequently “isolated nodes”, i.e. categories of documents in one taxonomy seeming to have no correlate in the other taxonomies. An example is in the Archery category on Yahoo (www.yahoo.com), the sub-category of “Kyudo” (traditional Japanese archery). Unfortunately, at the time of testing, there was no equivalent to this category on DMOZ (Open Directory Project) or About.com. Previously, in order to map such taxonomies to one another, it was necessary to manually edit the taxonomies. Manual editing is cumbersome and not pragmatic on a large scale.
The inventor has created a soft ontology expansion to assist in the alignment of taxonomies which can produce numerous candidate ontology nodes, such as “coaching/training” and “competitions/tournaments” and so on. The invention is a methodology that can be implemented on a computing device, such as a general purpose programmable computer that is programmed to accomplish the functional steps of the invention.
In a preferred embodiment, the archery taxonomies noted above are used as an example. In this example one candidate ontology node of particular interest is “traditional archery.” While not a node in any of three reference taxonomies, “traditional archery” nonetheless applied to a great number of documents in all three, and especially in Yahoo's apparently unique “Kyudo” category. Having used DMOZ as a “master taxonomy”, we then tried adding not only “traditional archery”, but every other similar example, with the result that, in the archery topic tree, we ended up adding three new nodes:
The first of these, “traditional archery”, included (as a child node) all the Kyudo documents, plus numerous documents from the other two indices, all of which were pertaining to traditional forms of archery. Since not all traditional forms of archery (such as medieval European forms) are the same as Kyudo, it made sense that Kyudo be subsumed in the new node, rather than the other way around.
The new node, “Coaching & Training” emerged because of the failure of Yahoo's “Gear and Instruction” node to align with any nodes of the other indices. A near match was the “Equipment Manufacturers” node, but this match failed to score high enough on the resemblance test (set forth below), essentially because Yahoo's node had the problem of mixing gear and instruction, whereas the other indices separated archery “gear” in its own category. Meanwhile the other indices did not have a dedicated “instruction” category. But since “training/coaching” had been produced by the soft ontology expansion and was linkable to “instruction”, it was therefore created as a new node.
Finally, “Stories and Discussion” was devised as a means of uniting several unclassified documents (i.e. those simply categorized under the main “archery” node and not assigned a specific topic) with the “Chats & Forums” of DMOZ.A heuristic can be used for automating this addition of nodes according to the following:
1. Find an expanded concept that is instantiated disproportionately in the document glosses of an unmapped node.
2. Test if that node is instantiated also in numerous documents not classified at a leaf node in a plurality of taxonomies.
3. If such a node is found, then create a new node with that concept and place the relevant documents under it.
As illustrated in
The modules of the preferred embodiment can be software code and/or computer hardware that serves to accomplish the functional steps described herein. The modules can reside on a single computing device or multiple computing devices communicating with one another, such as over a network. The modules can be operative to receive and read the relevant taxonomies and to write data thereto.
The overall approach to taxonomy alignment is to combine an intensional and extensional analysis so as to arrive at a mapping between nodes in divergent taxonomies. The intensional method examines the semantics of words in the names of the nodes, and in the titles of documents classified in it, as well as in the glosses applied to those documents by the taxonomy editors. The extensional method need not examine semantics, but only checks the cross-classification of documents that happen to be included in more than one of the taxonomies, coextensive classification as a clue to alignment.
An embodiment of the invention is applied to human-crafted document taxonomies bearing short glosses. These glosses are, by definition, meant to summarize in a few words what the documents are primarily about and what differentiates each one from others in the same topic. Hence they are valuable input material for semantic resemblance analysis. The content words of the document titles and glosses, as well as bi-grams containing a topic word in any derived form can be used. For example, in the archery category “field archery” and “archer's union”, in addition to single words such as “arrows” and “bows” could be used. Then it is determined which of these may be closely related by semantic resemblance. For measuring semantic resemblance, one can test for “semantic proximity” as disclosed in WordNet (Miller 1995), which we define as ‘having a maximum distance of 2 in the WordNet hierarchy, with the additional limitations:
1. Only synonyms, hyponyms, hypernyms, and sister-terms are to be considered;
2. Sister-terms are considered proximate only if they share multiple content words in their glosses and/or example sentences in WordNet; and
3. Hypernyms are included only if they are at least 4 levels down in the WordNet hierarchy from the root.
In applications for Web documents or other documents having excessively frequent words, certain words that are excessively frequent across all categories, and hence not useful, can be ignored. In an example of Web documents, these words are included in a preponderance of glosses in DMOZ, such as “photos”, “contact details”, “site map”, etc.) Table 1 shows an outline of a case study that is an embodiment of the invention.
The result of applying the method of the embodiment is, for example, that “clubs” and “organizations” are treated as equivalent terms. Word sense disambiguation is rarely an issue (and so was neglected in this embodiment), as the example is confined to a particular domain. In other words, since we are starting with purely “Archery” documents, there is little need to disambiguate, say, “club” or “bow”, which could have completely different senses in other topics besides Archery.
This allows implementation of a simple percentage match scoring of the content words in node names. For example, “Equipment Manufacturers” and “Archery and Bowhunting Gear Manufacturers” receive a score of 0.80, owing to the following facts. First, “Archery” is omitted because it is the same as the overarching topic of “Archery” and hence implicit in all node names. Second, the stop word “and” is discarded. Third, “gear” is matched to “equipment” as a hypernym. That leaves five words total, with only one of them (“bowhunting”) lacking a match: hence the score of 415=0.80. In the embodiment, any match greater than 0.66 was deemed sufficient for alignment. The virtue of this node name resemblance test is that it facilitates alignment of, for example, “Clubs and Organizations” with “Clubs and Associations” in two different taxonomies.
However, this still leaves us well short of aligning all the nodes. For example, we still have “Kyudo” and “Gear and Instruction” and “Chat and Forums” in the Archery domain, which either find no semantic resemblance to the names of other nodes, or they find poorly scoring ones. In the case of “Kyudo”, the word is not even in WordNet. In the case of “Chat and Forums” in DMOZ, there is simply no correlate in either Yahoo's or About.com's archery listings. And in the case of Yahoo's “Gear and Instruction”, the problem is one of having only a small, partial match: “gear” gets related to “equipment,” but “instruction” is not like anything at all in the other taxonomy node names, and thus the score never comes close to passing our threshold of 0.66.
Another, issue is that of the numerous documents not assigned a leaf node. In other words, in all three indices, many documents were simply classified in “Archery” without being assigned to a sub-category. In some cases, this seems correct, in that the documents in question were very general archery documents (or websites) not belonging to any particular sub-class. But in many other cases, it seemed that a node in a different taxonomy was a natural place for such documents.
In other words, while a website of personal stories about archery was classified, in one taxonomy, simply as an “archery” document, if it were viewed against a different taxonomy, it would have found a perfect home in ‘National Teams”. This defeats the taxonomy alignment, in that by merely aligning “Archery” with “Archery” in the two respective taxonomies, it is implied that none of the documents in the one taxonomy belong in “National Teams” of the other—and yet many of them did. This much was obvious from a human point of view, but the embodiment requires an algorithm for handling it.
This predicament led to a soft ontology expansion of all three taxonomies. In other words, it is desirable to enrich the ontological characterization of each specific leaf node, so that it can be aligned with an appropriate subset of the documents lumped together in a more general topic category of a different taxonomy. In order to accomplish this, the method of the preferred embodiment first combine all the documents of all nodes in each taxonomy in a single set. Then, going back to our extracted words and bi-grams (e.g. “calendar” and “field archery”, etc.), the words were related in WordNet to look for patterns. To do this, the WordNet glosses and example sentences are examined and compared with collocations and phrases in the document glosses. It was found that, if two words were frequently paired (collocated after skipping non-content words) in the taxonomy document glosses and also were found in each other's WordNet glosses, they were, without exception (in our case studies), genuinely related and of ontological import in the category.
An operational definition of “frequent” is having at least one occurrence in all three taxonomies and having multiple occurrences (2 or more) in at least two of three taxonomies. However, other thresholds can be used to define “frequent.” This technique employs WordNet to enrich or develop one's taxonomy and/or ontology. The embodiment drives the process by reference to the glosses already created by editors of the various taxonomies. As a result, the following examples were derived in Archery:
[calendar,schedule] having a relation to [event] [tournament,competition] having relations to both [results] and [standings],[outdoor] having a relation to [ranges], [bow] having relations to [crossbow], [compound bow], and [long bow].
In the embodiment, these are referred to as “concepts” rather than merely “word occurrences” because each is based on a small web of similar words, (e.g. “calendar”˜“schedule”) rather than a single word and each has an additional small web of word relations (“events,” etc.), all of which are contextualized to the overall local topic of Archery. The totality of all such extracted concepts we call a “soft ontology,” in that it delineates the raw materials, as it were, of the local ontology, but obviously falls short of a formal representation of the relations between the concept.
Next, when checking the non-leaf-node documents' glosses for the presence of these concepts, and when finding them, then employing the same percentage matching algorithm as above, the embodiment connects some non-leaf-node documents of one taxonomy to leaf nodes of a different taxonomy. For example, several documents containing “U.S.” in the DMOZ Archery node, found their way into ‘National Teams” on the Yahoo taxonomy. In the end, 37 of 189 documents were thus “migrated downward” to a leaf node, with the result that, on inspection, it seemed the alignment between taxonomies was more complete and intensionally unequivocal.
This illustrates that taxonomy alignment cannot be divorced from issues of taxonomical scope and adequacy. In other words, if one taxonomy lacks the scope or granularity of another, then the only way to achieve proper alignment is to sort through some of the items in the less granular taxonomy so as to “multiply align” it to other nodes. It is important to note why this is not the same thing as document classification. First, the body content of the documents need not be examined. Secondly, all the documents were already classified; it is just that sometimes they were not completely classified in one taxonomy when seen from the point of view of a different taxonomy. In the face of this fact, by leveraging the very substantive and relevant glosses, the embodiment can further granularize a taxonomy by reference to its items' glosses having strong semantic resemblance to those of another taxonomy's leaf nodes, in respect of instantiating the same concepts as derived from the ontology expansion. The inventors conducted several other case studies in the “Sports” domain of all three indices noted above (DMOZ, Yahoo, and About.com), including topics such as soccer, basketball, and sports injuries, with similar results.
Certain of the soft ontology concepts seem to embrace otherwise isolated nodes of one taxonomy, together with non-leaf node documents of another. A very clear example was that mentioned above, namely, “'Kyudo” in the Yahoo directory. The soft ontology expansion had derived “traditional archery” as a bi-gram, with a relation to “history”. This was very dense in the Kyudo category (occurring in all but one of its items), and meanwhile applied to 16 non-leaf-node documents from DMOZ, including these three examples:
Donadoni Archery—Supplier of traditional archery equipment in Italy. Lists a wide range of European and US made target and bowhunting products;
The Archery Centre—Specialists in field, traditional, and re-enactment archery equipment. Includes down-load catalogue and hints; and
Perris Archery—Recurve, compound and traditional archery equipment suppliers.
The procedure was, in such cases, to use the concept string as a new node name, and to include as a child node the originally isolated node. In other words, the master taxonomy now included “Archery/Traditional Archery/Kyudo,” with the three documents above from DMOZ placed in “traditional archery” (a node that had not existed before) and Yahoo's Kyudo documents taking their place in the “Kyudo” node, but now having as its parent the new node, “Traditional Archery.” This resulted in a far better alignment than the previous condition of an isolated “Kyudo” category; Kyudo documents now had a closer parent than just being a direct child of “Archery.” A new interstitial node of “traditional archery” functions to explain where “Kyudo” belongs in view of all three overlapping taxonomies.
The same is true of “Stories and Discussion” being introduced as a parent of “Chat and Forums”, and of “Coaching and Training” as a parent for the “Instruction” documents that Yahoo had mixed in with “Gear”. Table 2 shows the overall alignment results. Regarding accuracy, the introduction of new nodes carried just one misclassified document, the key point being that it was “carried”, the document had been misclassified already on one of the third party indices. In general, the accuracy of this method will be as good as the accuracy of the classification of the participant taxonomies. In the Archery case above, all the names read nicely, but when we did Soccer, one node received the name “instructing” when “Instruction” arguably is more suitable. Heuristics can be used for selecting a word form or derivation most likely to be preferred by humans as a node name.
The availability of editorially created glosses constitutes a rich resource to guide semantic resemblance analysis, and has the added bonus, when soft ontology expansion is applied via WordNet, to create new interstitial nodes for a more complete and unequivocal alignment of taxonomies. The invention has been described through a preferred embodiment. However, various modifications can be made without departing from the scope of the invention as defined by the appended claims. The invention can be effected with various computing devices and systems programmed in a known manner to carry out the novel functions described herein.
This application claims priority to U.S. patent application Ser. No. 12/025,200 filed Feb. 2, 2008, titled METHOD AND APPARATUS FOR ALIGNING MULTIPLE TAXONOMIES which claims priority to U.S. Provisional Application No. 60/899,040 filed Feb. 2, 2007, the contents of which are incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
60899040 | Feb 2007 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 12025200 | Feb 2008 | US |
Child | 14243812 | US |