Expert nding and the identi cation of similar professionals are important tasks for many services provided by companies and institutions. Nowadays, the rapid growth of web services and social and professional networks, allowed di erent kind of users to share personal data and increased the amount of information available. Most of research works focus on a limited set of users, characterized by the same kind of main activities, e.g., researchers, or exploit external knowledge, such as prede ned ontologies. An heterogeneous environment, with possible lack of information, and not well structured data, puts forward new challenges, to address the problem of adapting user pro ling and consequently expert search. In this paper, we rst provide a general perspective on studies on expert nding, similar people identi cation and social recommender systems, highlighting some critical issues related to information extraction and user pro le de nition. We then present a rst attempt to create an expert search system to support users in Library and Archiving Communities, such as researchers, students, authors, in the eld of Textbook Research, in nding other experts to get in contact with or to start a cooperation. This is organized in three phases: rst, we collect information about users in order to build structured pro les; then, we build a Community Knowledge Graph (CKG) which de nes relationships and weights among terms that occur in the pro le sections, emphasizing information shared by users in the entire analyzed community. As third step, we exploit the CKG structure to de ne similarity values among users based on weights got by their common terms in the CKG, and their distances in the graph. We conjecture that the CKG allows to model users emphasizing new semantic aspects of relationships among pro le elements, and helps to improve the similarity computation and the expert search. We present a preliminar experimental evaluation on real users.
2 1
Marco Pavan and Ernesto William De Luca
In recent years, the increasing pervasiveness of social network platforms has made users able to share personal data, also for professional purposes, in order to nd positions and collaborations o ering competence and knowledge. Also several companies and institutions have strong interest in exploiting that kind of information to nd people with particular skills and expertise that t their needs. With these premises it is clear how important is addressing the expert search problem and exploiting new sources of potentially interesting information, therefore, accurate expert search systems enable users and companies to quickly nd the right desirable people without being overwhelmed by irrelevant information or losing too much time seeking on several web platforms.
There has been an increasing interest among researchers in improving the
search results of expert nding, and several solutions had been proposed, even
recently [
Most of works in expert search focus on a single assumption about expert
similarity (i.e., researchers are similar if they have similar papers, or similar
topics; people on Q&A platforms are similar if their questions and answers cover
similar topics, etc.). While this is working well on speci c platforms for
speci c tasks, a more general and adaptable framework could be needed to take in
care multiple possible sources for user similarity, i.e. research papers if available,
tweets if users use Twitter, etc., and to allow us to build an adaptive system
able to handle di erent search tasks. In real situations users could have di erent
needs and di erent approaches on expert search. Someone could be an expert
who is looking for partners with same skills, thus, she could try to search pro les
similar to herself; on the other hand, an inexperienced user, such as a student,
might need general knowledge about a topic, with no particular requirements on
user expertise, just with the su cient background or interest. Yet others, could
be interested in nding users that have collected a lot of information about a
topic, thus, not related to work or skills, but with great interest on a subject,
such as a particular sport, political party, etc. On this basis, it is clear how a
more general pro ling could help, and how information about users obtained
from heterogeneous information sources could enrich the pro le, providing
different insights on user similarity. Recent studies [
Another critical issue is how to get the correct knowledge base, that helps to connect texts to semantic entities, such as categories or topics, and allows us to properly model users and build their pro les. Research works, focused on a single domain, exploit speci c ontologies or databases to address this problem, but obviously this approach is limited to a set of users of the same type, for instance researchers in a speci c eld, employees in a particular company, or customers interested in speci c products.
One of the major challenges of digital archiving for Textbook Research is how to deal with changing technologies and changing user communities, which necessitate tools to formalize, detect and measure knowledge evolution. Semantic representations of contextual knowledge about cultural heritage objects and users, especially in Textbook Research, will enhance organization and access of data and knowledge, because usually the relationships among them are not emphasized or even identi ed.
In this paper, we focus on a novel semantic-Blser approach for expert search. We create a Community Knowledge Graph (CKG) used for semantic enrichment. The graph is dynamically built during the users set analysis and allows to de ne user similarities by comparing the relationships among terms and entities. In particular, we collected structured information similarly to a CV, and taking into account the terms occuring together in di erent elds to emphasize their relationships to other users sharing the same content. Then we build the knowledge behind their pro les. Indeed, we observed that some users who work on similar projects might have similar biographical data, or share same interests. On the other hand, students with interests on particular topic, could be connected to experts through similar pro le content. With a speci c ontology or database the community analysis is limited and it is not possible to properly model a general user who could seek for experts with di erent pro le but with some intersection.
The novelty of our approach is the semantic enrichment of the pro les based on a community graph in order to improve the results of similarities scores among users. We exploit the CKG structure, with its weighted relationships, and we compare users considering the graph distances of their common terms in the CKG.
The paper is structured in 6 sections. After the introduction, we discuss related work in Section 2. We focus on the problem statement in Section 3 and describe our approach in Section 4. Section 5 describes our preliminary evaluation and Section 6 concludes the paper and shows our future work. 2
The expert search problem has arised great interest among researchers, and
several groups addressed the main problems behind this important topic, such as
lack of information about users, or de ning supporting knowledge needed to
build models and improve the results. Gollapalli et al. [
Marco Pavan and Ernesto William De Luca
that identi es latent content and graph classes for people with similar pro le.
Moreira et al. [
Other researchers focus on speci c platform, such as Twitter, to address
speci c expert search problem. Stankovic at al. [
All of these related works have signi cant value, given the importance of all
the issues addressed to improve expert search systems, and it is clear how one of
the most important issues underlying these systems is the user modeling and the
related eventual enrichment process. In this direction other authors recently
presented their work. Abel et al. [
Another important issue often present in research works related to expert
search is the use of a supporting knowledge base, usually de ned with an
ontology. Some researchers focused on ontology-based user interests modeling and
matching. Cena et al. [
To better understand what are the main problems and di culties emerging with expert search tasks and related user pro ling, we list a set of conceptual problems presented by the current state-of-the-art solutions. Many expert search systems, and even recommender systems, as rst step, need to extract the starting information about users in order to build the pro le. Most of existing systems rely on a single data source, with the risk to have incomplete information about users. Depending of the nature of that source, the obtained information could be related on only one aspect of the users, i.e. only their work expertise or skills, or purchases preferences; or it could be very schematic and represented by very short texts, therefore likely with poor knowledge about them. So two rst conceptual problems are:
the user pro le could be incomplete and focused only on certain aspects, therefore ignoring a more complete user overview.
user pro le could be composed of short texts that make di cult the information extraction, due to their brevity that does not provide su cient word occurrences.
Moreover, the texts that compose the initial user data, most of time are not structured, and users are represented as bag-of-words, with no information about what kind of data could be related to personal data, expertise, or general interests. We can de ne this problem as:
do not have structure that allow us to identify what kind of information we have. This issue makes di cult the process of dividing information in sections to properly model users under several aspects.
Very recent works [
Marco Pavan and Ernesto William De Luca are not always the same, arise the need of having dynamic also the knowledge support, evolving over time. On this basis we de ne other two problems:
semantics. Ontologies and any supporting knowledge bases are usually crafted for speci c domains, therefore they might not be well suited for any set of users.
gies and any supporting knowledge bases are usually xed structures, therefore not able to change over time in adaptable frameworks.
Another important issue in expert search systems is related to user similarity computation, used for de ning the scores between couples of users and obtain a ranked list of suggestions that meet the needs expressed by the requesting user. Most of the state-of-the-art proposed systems use a single function to compare the user models they build, but in heterogeneous environments could be helpful to emphasize only some aspects of users, in order to better t the expressed need, taking in care what kind of user is the requester, or even analyzing for what purpose is the request itself.
computation of only one similarity score between two users is a global value that does not take in care the several aspects of a user pro le. Moreover, it does not give di erent weights to part of pro le that need more or less emphasis, based on the current request.
Recent research works [
In the following we describe our approach organized into three main phases. During the rst phase we collect information about users in order to build structured pro les; the second phase consists in building a graph composed of all texts from all users, with attention on enriching the network with semantic relationships among entities obtained by the pro les structure; in the last phase we compute similarity scores among users exploiting the network structure that allows us to consider distances among terms which represent users.
Figure 1 shows an overall representation of our approach. Each phase is described in full details in the following sections. 4.1
We have chosen a real situation to analyze as case study, in order to then investigate on the e ectiveness of our approach. We started a pilot study at the Semantic-based Expert Search in Textbook Research Archives 7 Georg-Eckert-Institute (GEI), to collect data about people who work in the international textbook research eld, but with di erent roles.
A pilot study is a standard scienti c tool for testing a research question in a \soft" way, allowing scientists to conduct a preliminary analysis before starting a full-blown experiment. In our case, we decided to start a survey in order to analyse how we can help researchers in nding other researchers that are interested in a very speci c research area, namely Textbook Research. We wanted to nd out how we can create semantic user pro les being di erent from other research portals like \research gate" or \academia".
We selected a sample of 32 users distributed as follows: 33% men, 67% women (di erent nationalities); 18% with age under 30, 55% between 31 and 40, 27% more than 40; 70% graduates, 30% doctors; 67% with activities related to research. We collected data related to biographical information, their current position, projects where they are/were involved in, their interests, etc., all for researchers, authors and students/trainees. We then structured information by grouping terms which compose the collected textual descriptions in a tree structure with a set of pre-selected entities. It is clear how an heterogeneous set of users could led us to have di erent pro les with di erent kind of shared interests or expertise have been collected into the community knowledge graph.
Figure 2 shows an example of such a structured user pro le with the related entities and terms.
To test the feasibility, equipment and methods, we started the pilot study for nding out what are the important relations that should be collected for enriching the base pro le within semantic information that can be derived from the needs of the researchers we asked to participate. For this study we focus on digi
Marco Pavan and Ernesto William De Luca tal archives eld and in particular on the case study of Textbook Research, that does not involve only researcher, but, as mentioned before, people with di erent backgrounds and interests. Therefore, this preliminary test is also important for training inexperienced users, explaining them how to interact with the system, what is the knowledge we extract, in order to get useful information to nd out what are the sub-group of experiments we have to take into account after this pilot phase. Pilots are rapidly becoming an essential pre-cursor to many research projects, in order to reduce costs. At the same time we could nd out how to set the next experimental phase, because the users gave us feedback of unclear statements or questions. 4.2
As second step we build a knowledge graph in order to have a structure that represent the entire community. We de ne this structure as Community Knowledge Graph (CKG). Based on the structure de ned for user pro les we keep the same entities as nodes and we add directed edges in order to connect words used by users to those entities which represent the important aspects that characterize those users. We also add a new set of nodes in order to represent users and connect them to the words they used. In this way it is possibile to analyze the network by user. We scan all users' data in order to dynamically build the CKG and semantically enrich it with higher weights on edges where the relationship between a word and an entity is repeated. With this approach we emphasize the important relationships inside a community facing the problems P3 described in Section 3.
More formally, let CKG = (V; E) the Community Knowledge Graph, where V = fv1; v2; :::; vng is the set of all entities and words extracted from user pro les, Semantic-based Expert Search in Textbook Research Archives 9 and E = fe1; e2; :::; emg is the set of undirected edges. We say that there exists evi;vj 2 E , vj isRelatedTo vi. The property \isRelatedTo" could be intended in several ways in order to de ne multiple semantic relationships. The TermEntity relation is used for words belonging to the same concept, the Entity-Entity relation represents concepts related to a more general category of information which characterize users and the User-Term relation connects users with the words they use.
Figure 3 shows an example of CKG where it is possible to see how users are connected to the words they use, words to Entities, and nally to the more general entities that we call Categories. During the building process, if some words or entities are already present in the model the edges weights are increased in order to emphasize those relationships3. With this methodology we can build the real structure that users' aspects have. For instance, if a lot of people with a speci c skill work in a particular company, we can have that strong relationship as high weight for the edge that connect the word with company name and the entity which represent the skills. 4.3
The creation of the CKG lets us comparing di erent users extracting information from their user pro les, de ning similarity scores between them. The network structure allows us to have connections between words and entities based on the actual relationships found in the original texts, therefore we extract only the sharing nodes of two users, and we measure a similarity score based on the weights of the relationships computed as in the following:
Starting from two analyzed nodes vi and vj we follow the path in the graph until they share a common node. If this is the case, then we compute the similarity score based on the valid relationships, otherwise we discard it. In the 3 For an easier reading, in Figure 3 multiple edges are shown.
Marco Pavan and Ernesto William De Luca case there are multiple paths we consider the path with the highest total edges weight, in order to emphasize the strong relationship between the two nodes that CKG has extracted.
In Figure 3, the purple dashed line highlights an example of selected path for the weighted distance. Formally, Let Fi;j E be the set of edges into the selected path from vi to vj de ned with our rule previously described, we de ne w(vi; vj ), the weight of the relationships between vi and vj , as follows: w(vi; vj ) =
w(Fi;j ) plen(vi; vj ) where w(Fi;j ) is sum of edges weights into the respective set of edges, and plen(vi; vj ) is the length of the path from vi to vj .
By repeating this process for all couples of nodes we can get a global score for user similarity.
Let Vua V and Vub V be respectively the sets of nodes of hypothetical user ua and user ub. We de ne the similarity score as follow: sim(ua; ub) =
X w(vi; vj ) with vi; vj 2 Vua _ vi; vj 2 Vub . 5
In order to study the e ectiveness of our approach, we have set up a preliminary experimental evaluation. We have run three di erent algorithms on the dataset composed of the 32 users described in section 4.1. The rst one, called bows, is a classic bag of words approach which exploit only the information users provide about their biography, therefore with no additional information about skills, current projects, interests. It simply counts the term frequency in order to build the user pro le. The second one, called bow, is the same bag of words approach but using all information available about users, not only the \short bio". The third one, called pei, is our approach based on the semantic-based enrichment process that uses our CKG. We have run all algorithms 32 times in order to get the top ten suggested users for each one of the analyzed user. For the evaluation we have chosen a sample of 6 testers: 3 internals, therefore people who are working at the GEI and who know all users pro les and activities inside the institute; and 3 external, people who do not know them, but with access to the dataset with all text inserted.
Table 1 shows rst results obtained by each approach, using the Precision metric at several levels, with resulting ranked lists with 3, 5 and 10 elements.
Table 2 displays the Discounted Cumulative Gain scores, using the Precision metric at several levels, with the same granularity levels.
Analyzing the results, we can notice that bow is less e ective than bow-s, even if it exploit more text and information. The pei approach outperforms the Semantic-based Expert Search in Textbook Research Archives 11 other approaches and con rms the assumption that semantic enrichment can increase the performance of the retrieval system personalizing the results.
By looking at the Precision scores on Table 1 it is possible to notice how bow-s tends to keep the scores around a certain value, di erently than bow that increases value as the resulting ranked list get longer. This issue highlights how the use of additional information could help on systems who provide longer lists of results but with no high precision on top ranked elements. Our proposed approach pei overcomes this issue by providing good top ranked results using more semantic information.
Table 2, with DCG scores, shows how all algorithms provide more gain for users, as the resulting ranked list get longer, and it is possibile to see how our technique pei outperforms the others at all levels. 6
In this paper, we provided a rst perspective on studies on expert nding, similar people identi cation and social recommender systems, highlighting some critical issues related to information extraction and user pro le de nition. We presented our semantic-based enrichment approach for expert search that bases on a Community Knowledge Graph, which de nes relationships and weights among textbook researchers. The semantic relations help in nding the di erent experts that could be of interest and could be recommended.
For future work, we plan to expand our approach and investigate how the network structure could be exploited to improve the results, and to run the next evaluation with detailed user tests. Moreover, we want to explore the possibility to extend our approach to resources that users interact with, such as textbooks, manuscripts, or other cultural heritage objects, in order to improve semantic search and information retrieval tasks in digital archives and libraries, considering the relationships among them and with users.
Marco Pavan and Ernesto William De Luca