The paper describes lexitags, a new approach to social semantic tagging whose goal is to allow users to easily enrich resources with semantic metadata from WordNet. This is a paradigm example of the Social Web and the Semantic Web working together: ordinary users help create the metadata so needed by the Semantic Web and in turn, Semantic Web technologies help those users get a richer experience from the Social Web. A family of simple user interfaces for lexitagging is described, as are some methods for the subsequent, automatic generation of lightweight ontologies. These ontologies are presented as an ideal interlingua for the Social Semantic Web.
Two of the most exciting innovations for transforming the World Wide Web are
“Web2.0” [
In particular, the Information Architecture community has embraced folksonomy1 as a way to enhance information management practices. An analogy is often made with the term desire lines, which comes from landscape architecture. The basic idea originates in the observation that, in spite of the careful planning undertaken by architects to lay out walking tracks in their meticulously designed spaces, one will often find emergent paths that have been forged by people who deviate off the planned tracks onto the grass or gravel of the spaces. The paths become entrenched when particular tracks are found useful by many people. It is similar in information spaces, where folskonomy describes the desire lines, representing informal tag based classification schemes that people find useful. The addition of formalized “desire lines” on the web would benefit emerging semantic platforms that rely on such metadata, especially with respect to querying and mining social semantic data.
This paper describes a set of tools and principles currently under development, which will help formalize folksonomy for the web. In the next section we describe some of the main problems with current tagging practice. Then we describe our approach to cleaning up tags and generating formal rdf based semantic tags. Following this, a method for automatically generating lightweight ontologies from semantic tags is described, and their use as a universal interlingua is explained.
The basic problems with folksonomy terms from user tags are nicely summarized in [
[
[
They also note after the quantitative evaluation that “However, we did find that singleuse tags were less common than we had expected”, suggesting a high degree of consensus in tagging behavior, and a correspondingly low degree of “personalised” tags. They additionally note that the high number of tags that were not words that can be found in a standard dictionary may be artificially high. In many cases the tags were misspelled or creative variants of dictionary words. Many examples of misspelling consisted of the transcription of characters across languages. For example, the Norwegian ae can be written as “ae”. Sometimes the reason was the compounding of words and letters as in “17thjuly”. Another prominent practice was the inclusion of geotagging information (latitude and longitude) in the tag. This was particularly popular in Flickr (perhaps unsurprisingly).
Based on the previous observations, it is fair to say that the predominance of tags are
dictionary words, or compounds formed from dictionary words (or numbers). In support of this
conclusion, [
The problem with recommendations is that they can be difficult to enforce, or even
convince people that they should try to follow them. For example, one could stipulate that tags
should follow NISO recommendations [
The solution which is suggested in this paper is at the outset a simple way to gently enforce these best practices through the tagging interface, by allowing people to simply tag with dictionary words, otherwise known as lexical items. It is for this reason that the tags themselves are called lexitags. Lexitags guarantee that every tag can be unambiguously connected with a known lexical item, while still allowing some flexibility in user behavior. For example, both cat and cats are allowed in the user interface, but both are linked to the lexical item {cat}. By keeping information about the surface form and lexical item separate, no information about user behavior is lost: the underlying semantics of the tag is captured, as well as the potentially significant choice of plural or singular. On the other hand, only acceptable spellings can be used, so the problem of misspellings, idiosyncratic spelling variations and so on, disappears.
The primary lexical database in this project is WordNet. This is supplemented by DBPedia which provides terms missing in WordNet, such as names for emerging technologies and people. WordNet is perhaps the most well established electronic lexical database, whose development at Princeton University dates back to 1985. WordNet represents disambiguated word senses with synonym sets (synsets), which are equivalent terms enclosed in braces. For example some of the unique senses of the word cat are: {cat, true cat}, {guy, cat, hombre, bozo}, {cat, gossip}, {kat, khat, qat, quat, cat, Arabian tea, African tea}, {cat-o'-nine-tails, cat}, and so on. WordNet is a very large database, containing in total 206941 word-sense pairs including nouns, verbs, adjectives and adverbs. In addition, each synset contains lexical pointers to related synsets, where the relations are specific to grammatical category. For example nouns are included in (amongst other things) hyponymy and meronymy relations, but adjectives in antonymy. In summary, WordNet is an extensive database of English words, together with a rich set of lexical and semantic relations defined over the lexical items.
The simple idea, then, is to use WordNet as a source for disambiguating tags that are applied to resources by users, and to provide a simple interface where this can be achieved. The disambiguated tags are referred to as lexitags to honor their origins in the lexicon, or semantic tags to indicate that their interpretation is fixed relative to a semantic resource. In order to realize the interface design in a reference implementation, we designed a platform for social bookmarking, which we will refer to as LexiTags, and which was developed through a commercial startup company called LexiTags D.A. The company was jointly established by the author and his colleague Andreas Opdahl at the University in Bergen, and supported by a seed grant from Innovation Norway. It should be noted that there are at least two existing commercial ventures that are advertised as a “semantic bookmarking service”, making them similar to LexiTags in this regard. These are Faviki and Zigtag. Zigtag, which has been running since early 20092 is the most similar in that it uses its own dictionary as tag definitions. Faviki uses WikiPedia concepts instead. However, there are fundamental differences in the motivation for these services and LexiTags. Zigtag and Faviki are bookmarking services, pure and simple. Their interest in “semantic tags” is to enhance findability on their site by providing equivalences between differently spelled tags (NYC, New York City, Big Apple, …), returning results for only one sense of an ambiguous tag, and so on. On the other hand LexiTags is simply a reference implementation of the lexitagging interface, with the focus being on the generation and exploitation of the metadata itself, rather than the underlying purpose of the reference implementation (which in this case happens to be bookmarking). The principles and algorithms are meant to be portable to any application, using the semantics in the generated metadata to bridge the divide in content across the services and applications. Metadata in the form of lexitags becomes an interlingua between applications.
To demonstrate the idea of a simple portable tagging interface, we will discuss here an iPhone interface which is currently in development. The iPhone interface communicates with the LexiTags service over http and can upload html bookmarks, but also photographs taken with the iPhone camera. The application can therefore serve as a tagging interface for bookmarking as well as a photo upload service.
The design principle is that lexitagging must be no more difficult than ordinary tagging otherwise people will not be inclined to use it. One key research problem is how to rank the possible senses so that the sense intended by the user is immediately available in the interface. The current iPhone tagging interface is shown in figure 1. On the left are two ambiguous tags, and on the right cat has been disambiguated (which is evident from the text below the tag). Notice that both URL entry field and photo choser are both available, and the user must chose which they use. URLs have to be manually entered at the moment, but ideally this will be linked to the web browser. In figure 1 we see that the user has chosen (or snapped) an image of a cat, and has assigned two tags “cat” and “cute”. The tags are simply typed in the “Tags” field, and automatically marked as “undefined”. This allows people to initially add tags freely. Once they have typed a few tags, users must tap each one to define it, which brings up the selection interface in figure 2.
On the left of figure 2 is the initial display, showing 5 possible choices. There are actually 10 senses of the word “cat”, so one must scroll to see the others. It is here where the ranking becomes important, since ideally the desired sense will always appear in the top 5 choices. The iPhone choser currently uses a series of words related to each sense as a 2 http://www.readwriteweb.com/archives/ semantic_tagging_service_zigtag_finally_launches.php disambiguator, rather than the actual synset (in this example {cat, true cat}). This is because not all senses have near equivalent synonyms, so the synset in these cases is simply the word itself. For example the synset for the third sense of cat is simply {cat}, which would not be a useful disambiguator. We are still experimenting with the usefulness of this method for selecting the correct sense. We are also experimenting with an alternative display in the larger, browser based desktop client. The display in figure 3 shows for each sense the synsets from WordNet, as well as the full explanatory gloss. Usually one or other at least is available. In addition, each sense is preceded by a determiner in brackets. This is meant to help people with selecting the grammatical category: “(a, an, the)” are nouns, “(to)” are verbs and “(is)” are adjectives. Our feeling is that the simple iPhone interface is adequate, but we will need to experiment extensively before making any conclusive claims regarding usability, since the selection of the appropriate word sense is the most difficult and important role for the lexitags interface.
Once the process is complete, the entry can be submitted. Any tag that has not been
disambiguated by the user is simply discarded at this stage.
disambiguated key terms from the resource itself, then there are a number of useful algorithms for
computing the similarity between those terms and each candidate sense of the tag [
Currently we are using a much simpler approach, which is to use the tags themselves for disambiguating other tags. That is, once the user has sense selected the appropriate lexitag, then that can be used to rank any successive tags. The more tags that have been selected, the more accurate the algorithm can become. The biggest problem with this simplistic approach is that there is no disambiguating evidence for the first tag. However, in these cases we simply use the relative frequency of use, arguing that people are less likely to use infrequent senses of words as tags. There is no reason why these various techniques could not be used in complementary ways, combining ranking estimates based on the different sources. For example the initial disambiguating context could be a fast analysis of the title and some metadata, which would be replaced as the analysis of the text becomes available. In turn, this could be combined with the disambiguated lexitags as the user works his way through the tagging session. It is of course an empirical question to see which combination of these methods results in the best user experience.
The results of a tagging session are recorded in RDF, using a number of common standards including Dublin Core3, FOAF4 and Common Tag.5 Figure 4 shows the format we have adopted from the Common Tags specification. The representation is straightforward, so we only point out the two relations ctag:label and ctag:means. The former is the word string used by the tagger, and the latter is a “dereferencable Resource that identifies the concept expressed by the Tag”.6 Of course in LexiTags this is a WordNet synset. This allows some separation between the word string and its meaning, accommodating the case where NYC and NewYork can both be used as tags, yet refer to the same concept. Because the application is based on open standards, all web sites which expose their data in the Common Tags format will automatically inter operate. Lexitags give extra value in that they can add semantically rich, disambiguated metadata to a URL that may be recorded on another site without rich metadata.
@prefix ctag: <http://commontag.org/ns#> . @prefix wn: <http://www.w3.org/2006/03/wn/wn20/instances/> . @prefix rdfs: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> . @prefix foaf: <http://xmlns.com/foaf/0.1/> . @prefix xsd: <http://www.w3.org/2001/XMLSchema#> . @prefix dc: <http://purl.org/dc/terms/> . [] rdfs:type ctag:TaggedContent ; ctag:isAbout "http://commontag.org/QuickStartGuide"^^xsd:anyURI ; ctag:tagged [ ctag:means wn:dog-n-1; ctag:label "dog"@en; foaf:maker u1234; dc:created "20.01.2200"^^xsd:Date; ctag:taggingDate "22.22.2200"^^xsd:Date ] . http://dublincore.org/ http://www.foaf-project.org/ http://commontag.org/Home http://commontag.org/Specification#means
We have already mentioned the most straightforward advantages of using semantic tags
for finding content on a bookmarking site. But the use of WordNet as the reference semantics
provides far greater benefits. Lexitagging provides us with collections that are marked up with
semantically disambiguated lexical items, which have rich associations to other lexical items in
WordNet. We have taken advantage of this in developing a method for creating lightweight
ontologies for social media sites. [
We have implemented this algorithm in a web service which generates visualizations for a set of lexitagged resources. Figure 6 shows a typical visualization for a small set of tags.
The figure shows the user assigned lexitags in green (light grey leaf nodes), and their respective hypernym chains. The hypernyms eventually intersect at common nodes. The nodes colored blue (dark grey) are retained as SuperTags because they have three or more children. Light grey nodes (except the leaf nodes) are discarded because they have less than three, and black ones because they are too close to the root nodes. All of these parameters are adjustable in the interface, so it is possible to adjust the generality and inclusiveness of the nodes which are finally retained. Already in this small example we see some useful nodes emerge: group, organization, attribute, state, person. The emerging nodes can tell us about the nature of this particular collection. For example, in the sub tree originating from group we see organization but not social group emerge as an important node. The reason is that most resources in this part of the collection deal with either university, newspaper, or military. It is easy to imagine other collections where the predominant node would be social node.
The real power of using lexitags as a basis for lightweight ontologies becomes apparent when relations from WordNet are added to the inferred SuperNodes. For example, figure 7. shows the meronyms of organization. These can be added to the emerging ontology as properties. Once embellished with properties, the ontology becomes a rich representation of the key concepts in a social site, and can be used for various inference tasks. For example, if someone uses the tag Apple inc., then this will be an example of an organization. Since the ontology tells us that organizations have a quorum, this could prompt an application to automatically fill in the names of the Apple board of directors and even suggest them as contacts in the social site. The result is a rich, dynamically emerging ontology which reflects the users attitudes to the underlying domain, and which can change if the concepts or the tagging behavior of users change.
The results presented in this paper are preliminary, but the way forward is clear. The implementation of portable tagging interfaces will result in a growing number of resources tagged with lexitags. The resources could include traditional http bookmarks, geo tagged photographs, wiki and blog entries, and even local file systems. The automatically generated lightweight ontologies will add unique metadata to each site. However, because each site is marked up with the same lexitags, this will facilitate comparisons and sharing between the sites. In fact, we make the bold claim that lexitags (WordNet synsets) are an ideal interlingua for the social semantic web because it has the expressive power to align concepts between any arbitrary ontologies, yet is intuitive in the most basic sense of the word.
Notice that we are not advocating WordNet as a universal ontology. In fact, we are
sympathetic to [
WordNet may not be a universal ontology, but is powerful as an interlingua precisely for
the same reasons that make language so powerful at communicating concepts. Flexibility allows
one to finesse levels of detail but still communicate, and also allows one to reach arbitrary levels
of precision when needed. When using lexitags as an interlingua, designers of individual
ontologies can map their terms to specific interpretations in WordNet as the requirements demand.
They can chose the mappings that reflect their particular world view: for example domains that
require attribution of blame can map their terms to slip or mistake while everyone else can map to
error. If it is important that people who use cinema are kept away from people who use movie [
Another interesting possibility is that the lexitags interface may help solve another
problem with using WordNet as an ontology: lexical gaps. [
Lexitags can also serve as an interlingua between formal ontologies and the social web.
For example, SUMO [
There is a large body of work whose aim is to exploit folksonomies for more effective
information management. Most of the existing literature concerns the exploitation of statistical
regularities in the way tags are assigned to resources by users. [
One of the earliest demonstrations in the first vein was clustering on Flickr, where
polysemous tags are displayed with co-occurring tags in different sets of images. For example, the
tag apple has the following clusters: <mac, macbook, macintosh, computer, laptop, imac,
keyboard, powerbook, osx, macbookpro>, <fruit, red, green, food, tree, macro, canon, orange,
blossom, apples>, <ipod, iphone, music, nano, touch, shuffle, mp3, black, phone, ipodtouch>, and
<nyc, newyork, manhattan, newyorkcity, ny>. The algorithm can identify photographs tagged
with the different uses of apple: apple the fruit, apple the company, and the “Big Apple”.
However, this form of clustering is not simply lexical disambiguation since the company sense of
apple is listed in two different clusters which reflect different distinguishing product lines for the
company. An additional benefit is that different spelling variations of a tag are bundled into the
same cluster as in nyc, newyork, ny, because these tags tend to co-occur with the same pictures.
While the details of the Flickr algorithm are proprietary, various clustering algorithms were
explored by [
In another interesting use of co-occurrence, [
Clustering algorithms can identify different uses of tags, but they do not provide any
semantics beyond this. [
Researchers have also investigated the possibility of inferring hierarchical relations
between folksonomy terms. [
In terms of semantic enrichment there are several attempts to extend statistical
approaches by extending folksonomies using resources such as Wikipedia, on line ontologies, and
WordNet [
There are also a few studies in which users are expected to contribute semantics at the
time of tagging. [
Finally, the two previously mentioned commercial ventures Faviki and Zigtag should be mentioned as existing bookmarking services which make use of defined tags. Faviki uses Wikipedia concepts as common tags, and is able to aggregate tagged content according to Wikipedia categories. Since the defined tags are Wikipedia concepts, Faviki cannot semantically ground tags like interesting, cool, and useful. Zigtag uses dictionary entries, but also allows undefined tags, which make up a significant proportion of their tags.
This birds eye view of the literature shows that existing work is focused almost exclusively on the problem of extracting latent semantics from naive folksonomies composed of messy vocabularies rife with the problems of ambiguity and indeterminacy. In this respect the work presented here represents a much less well explored effort in eliciting precise semantic tags at the time of tagging. The current work is distinguished from similar research along four major dimensions. First, Lexitags aims to provide a lightweight tagging tool that can be used to tag a wide range of content including html bookmarks, pictures, and local filesystem content. Second, we use WordNet as the primary semantic reference, exploiting the structure of WordNet to construct new relationships and lightweight ontologies. Third, no tags are allowed to be completely undefined, which makes for a more coherent tag collection. Fourth, Lexitag users are not expected to make any complex decisions when assigning semantic tags. They are not expected to contribute relational tags, and so on. They simply chose the sense of the word which they already had in mind when writing the tag.
The paper introduced the lexitags approach to social semantic tagging with simple lightweight tagging interfaces. Lexitags are tags whose semantics are grounded in disambiguated lexical items, and which stand in useful relations to other disambiguated lexical items. These form the basis of automatically generated lightweight ontologies which can take the role of universal interlingua between social applications in any domain, and in many non English languages.
Tags which have rich, unambiguous definitions make some aspects of previous work to make sense of tags, un necessary. There is no need to infer that spelling variations on a term have the same meaning, for example, because the distinction between word form and word meaning in lexitags already accommodates spelling variations. Similarly there is no need for disambiguation or clustering for the purpose of identifying different word senses. However, many of the current ideas can still be used in more refined ways. For example clustering is still useful but now at a more detailed level because we can focus on clusters within each sense. If we ignore the fruit sense of apple, for example, it may be possible to discover interesting clusters in the way the company name is used. Similarly, taxonomy inference for “tags-in-use” with any of the methods mentioned is still possible, but now it can be refined by taking into consideration the semantics of the tags. For example if subsumption can only occur between nouns, then glass will never subsume magnifying.
Semantic enrichment becomes much easier too, because lexitags are primarily WordNet synsets. As an example, WordNet already has a rich mapping to DBPedia, so embellishing the dynamically constructed ontology with Wikipedia facts is much simplified. This is the essence of the linked data movement, removing uncertainty and probability from data integration.
One of the most important claims is that WordNet is the ideal means by which to ground
the semantics of common tags. This differentiates Lexitags from previous efforts such as Faviki,
[
In summary, this paper suggests that the tagging world be turned upside down. Rather than using clever algorithms for making sense of messy user generated tags, the clever algorithms should be used to help users generate tags that make sense in the first place.