When we want to find experts on the Web we might want to search where the knowledge is created by the users. One of such knowledge repository is Wikipedia. People expertises are described in Wikipedia pages and also the Wikipedia users can be considered experts on the topics they produce content on. In this paper we propose algorithms to find experts in Wikipedia. The two different approaches are finding experts in the Wikipedia content or among the Wikipedia users. We also use semantics from WordNet and Yago in order to disambiguate expertise topics and to improve the retrieval effectiveness. In the end, we show how our methodology can be implemented in a system in order to improve the expert retrieval effectiveness.
The Web can be viewed as a repository of knowledge produced by experts. Of course, there are several issues on the entire Web such as spam or advertisements which are not present in the common Enterprise scenario. However, there are subparts of Web where these problems are somehow mitigated.
In previous works the problem finding experts in online communities has been
already studied [
Two approaches are possible to find experts in Wikipedia (see Figure 1). We describe in Section 2 how it is possible to find experts using the Wikipedia content. Looking in the articles describing people we can build an expert profile for each individual that can be later retrieved using a query describing the topic of expertise in which we are looking for experts. The second option is to search for experts among the Wikipedia users (see Section 3). The assumption in this
case is that the users are somehow experts on the topic they write articles about.
In this way, looking also at the edit history of each user, we can build expert
profiles for them that can be later retrieved by a regular expert search system
(see for example [
The advantage of this scenario is that we can search for expertise also getting
help from a semantic lexical database for the English language such as WordNet4.
Similarly to what the authors of [
Wikipedia articles are manually annotated and grouped in categories. One of such category is “Category:People” and all the articles about single individuals are labelled with this category or with its subcategories. In this way we can refer to a person using his related article (if existing).
We propose to create an expert profile for each person in Wikipedia. That is, each article in the category “People” represents a candidate expert. We build for each of these candidates a profile which will describe all her topics of expertise allowing an automatic system to perform a search within this set of profiles. We propose to build this profile as a vector where for each possible topic of expertise there is a score representing the confidence/amount of the candidate expertise.
These scores are real numbers in the interval [
In this paper we focus more on finding experts among Wikipedia users. We
describe the appropriate algorithms for this task within the subsequent
paragraphs of this section. First, we need to extract expert profiles for the people
who participate in the production of the Wikipedia content. Similarly to the
system SuggestBot [
We need algorithms to perform the task of expert finding within the Wikipedia users. One first na¨ıve approach is to use standard Information Retrieval techniques considering the expert profiles (defined as the titles of the edited articles) as standard documents and indexing them using an inverted index. We can then create a query which represents the topic in which we are looking for experts and query the index in order to find a ranked list of people using a TF-IDF similarity measure.
Some more advanced techniques can be used to improve the search
effectiveness for example using the link structure as done in the Web for document
search. The assumption is that the most linked documents have an higher
authority and they are a better representative of the expertise of its authors than
other less linked documents. That is, the author of a highly linked/cited article
is an expert on the article’s topic. In this way we can give different weights to
the authors according to the authority of the content they produce (see
Algorithm 1). The authority weights of the articles can be computed with standards
algorithms which considers the links structure such as PageRank [
Notation. In the described algorithms we use the following notation: – u.score(i) represents how good is the item i as representative of the expertise of the user u. – u.rsv is the score of a user u as returned by an expert search system for a given query. It is used to rank the retrieved experts in the final results. – hW inN(i) represents the subarticle of the article h composed of a window of text of size N around the link to the article i.
Algorithm 1 Algorithm for defining expert profiles Require: A set of users U 1. Build expert profiles for all users u in U do - fetch the list of edited articles - u.score(a) = 1 {concatenate the articles title in one profile Pu} - store the profile Pu in the repository for all article a in the profile Pu do - compute/load the authority weight of a u.score(a) = u.score(a) · authority(a) {modify the expertise score of the user u} end for end for 2. Index and Search the profiles return list of ranked profiles 3.2
A more sophisticated algorithm would make further use of the link structure present between the Wikipedia documents. If with the methodology described in Section 3.1 we obtain a small expert profile because the user has participated in the creation of only few articles, we then propose to expand this profile using the citation network with the assumption that the users know something about what they cite. In this way we can define an expert profile with topic weights influenced by the link structure of Wikipedia. In Algorithm 2 we present the strategy which makes use of this information. The algorithm starts considering a profile containing the edited articles. Then, for all the items in the profile, it first considers the ingoing link information. This Algorithm adds to the profile topics extracted from a windows of N (e.g. N=10) words before and after the anchor of the citing documents. For example if the citing document has a part like “A good IDE for java is [Eclipse:Eclipse].” where [target:anchor] indicates a link, we can add IDE and java to the profile of the authors of the Eclipse page. The Algorithm also adds the linked articles with a lower weight into the profile until the profile is big enough. 3.3
A third possibility is to use a technique similar to collaborative filtering. Using a measure of similarity between users (based on co-editing of articles) we can alternatively: Algorithm 2 Algorithm for expanding expert profiles using the citation network – extend the expert profile including the topics of expertise of very similar users (see Algorithm 3); – extend the list of retrieved experts with those similar to the retrieved ones in a pseudo-relevance feedback fashion (see Algorithm 4).
The similarity measure between user’s profiles we can use is, for example, the standard Jaccard measure:
J (a, b) := |Pa ∩ Pb| |Pa ∪ Pb| (1) where Pa, Pb are the sets of articles edited by a user a and a user b respectively.
In Algorithm 3 we first initialize all the items in the user profile to score = 1. If needed, we then expand this profile adding the articles edited by the most similar users with a score proportional to the users’ similarity.
In Algorithm 4 we first need to run Algorithm 1 in order to have a first list of experts. Then, for each retrieved expert, we search for similar profiles adding them to the results with a Retrieval Status Value (used by the system to rank the experts) proportional to the users’ similarity. 4
In the scenario we are tackling there are several ways to improve the retrieval effectiveness using different evidences. One of such ways is the use of semantics. Algorithm 3 Algorithm for expanding expert profiles using the co-editing information Require: T is the acceptable number of articles in one user’s profile Require: size(·) function that return the number of articles contained in one profile Require: Z the similarity threshold between two users (e.g. 0.9) Require: run Algorithm 1 for all profiles Pu do if size(Pu) < T then 1. Initialize for all items i in the user profile Pu do
u.score(i) = 1 end for 2. Expand profile while (size(Pu) < T ) do for all profiles Pv who have edited any article in common with Pu do Suv ← J (u, v) if Suv > Z then for all items h in Pv do add h to Pu with u.score(h) = u.score(h) · Suv {weight the authority using the users’ similarity} end for end if end for end while end if end for Algorithm 4 Algorithm for expanding expert search results using relevance feedback Require: Z the similarity threshold between two users (e.g. 0.9) 1. retrievedList ← run Algorithm 1 for all profiles Pu in retrievedList do 2. Find similar profiles for all profiles Pv who have edited any article in common with Pu do Suv ← J (u, v) if Suv > Z then
add v to retrievedList with v.rsv = u.rsv · Suv end if end for end for Annotations can help to identify the correct articles to consider (see Section 2), knowledge taxonomies can help in finding the correct experts, and ontologies can help in disambiguating multi senses topics. 4.1
The expert finding task is usually performed in Enterprises where the significant
knowledge areas are limited. For this reason the expert finding system
usually adopt customized taxonomies to model the organization’s most important
knowledge areas [
On the Web the expertise areas covered are much more wide than in an
Enterprise. For this reason finding expert on the Web will require much more
effort to manually develop a universal expertise taxonomy. We propose to use
the Yago ontology [
Differently from the Enterprise context where the topic is narrow and clearly defined, in such a wide topics set as the one maintained in Wikipedia there is one more problem to take into account: the topic ambiguity. Multi sense terms might represent topics of expertise. For example, an expert on “Bank” might be expert on only one of the several senses of this noun: slope/incline | financial institution/organization | ridge | array | reserve | . . .
Using the algorithm JIGSAW [
It is also possible to use co-occurrence statistics to improve the quality of the profiles. If we take a user profile Pu we can disambiguate the topics looking at the context in the related articles. For example, according to the profile, the user u is an expert on “Jaguar” and we find that in the articles considered in his profile the word “Car” often co-occur with the word “Jaguar”. In this way we can add the topic “Car” to the expertises of u always with the final goal of disambiguation. 5
We propose to build an expert search system which makes use of the knowledge present in Wikipedia. As described in this paper it is possible to implement such a system using several techniques. We now want to discuss about the effectiveness evaluation aspects.
The present ways to evaluate the performances of a retrieval system are
based on the standard proposed for the first time by Cleverdon with the
Cranfield Experiments in 1966 [
The most relevant initiatives to our task are SemEval [
In this paper we proposed algorithms to build expert profiles using Wikipedia. We proposed to search for experts both in the content of Wikipedia and among its users. We also proposed some techniques which make use of semantics to improve the results disambiguating and extending the search.
As future steps we aim at build a systems which makes use of these techniques and test it on Wikipedia also asking some (voluntary) Wikipedia users to evaluate the performances of our system.
Acknowledgments. This work was supported by the Nepomuk project funded by the European Commission under the 6th Framework Programme (IST Contract No. 027705). We thank Paul-Alexandru Chirita for his useful comments and advices on our work.