- This paper presents an integrated approach to textual resource retrieval, which combines logical inference services with user profiles, in which a structured representation of the user interests is maintained. Learning is performed on documents which have been disambiguated by exploiting the WordNet lexical database, in an attempt to discover concepts describing user interests. The proposed approach relies on several additional features compared to classical lexical knowledge systems, including: structured user recommendation, numeric value management, definition of strict and negotiable constraints and keywords to retrieve potential interesting resources w.r.t. both user request and profile.
I. INTRODUCTION
The main goal of this paper is to propose a strategy to design advanced semantic search engines based on the idea of combining semantic matchmaking with Bayesian text categorization. By means of formal ontologies, modeled using OWL [24], the knowledge on a specific domain is modeled and exploited in order to make explicit the implicit knowledge, and to reason on it by means of the formal semantics expressed in OWL. On the other hand, a content-based recommender, which is able to learn user profiles from disambiguated documents, is used for customized search. The recommender exploits lexical knowledge in the linguistic ontology WordNet [26].
The success of a retrieval system strongly relies also on query formulation and ranking functions. Especially for an ontology-based system, the query language has to be very simple for the end user but, at the same time, its expressiveness must be able to capture the real user needs and to retrieve only what the user is really looking for. In this paper we present a system able both to help the user during the query formulation process via an intensional navigation of the ontology, and to return relevant resources via a ranking function exploiting both the ontology-related semantics of the query and the user profile managed by the content-based recommender.
Hence, the system suggests interesting items to user by taking into account three elements: user profiles, semantic item descriptions and lexical item descriptions.
The rest of the paper is structured as follows: the next section outlines the work which mainly inspired this paper. In Section III a brief summary of semantic-based matchmaking in Description Logics is presented together with a Na¨ıve Bayes method for user profiling. The description of the framework architecture, a domain reference ontology, together with an example of query example satisfying user needs, is presented in Section IV, while conclusions close the paper.
Recent years have witnessed a growing interest towards profiling based resource retrieval. Among the most relevant systems adopting a bayesian classifier we cite LIBRA [22] which produces content-based book recommendations by exploiting product descriptions obtained from Amazon.com Web pages. Documents are represented by using keywords and are subdivided into slots, each one corresponding to a specific section of the document as authors, title, abstract. SiteIF [20] exploits a sense-based representation to build a user profile as a semantic network whose nodes represent senses of the words in documents requested by the user. OntoSeek [14] explores the role of linguistic ontologies in knowledge-based retrieval systems. AMAYA [27] delivers context-aware recommendations, which are based on provided feedback, context data, and an ontology-based content categorization scheme. Former system deals with user profile and on this basis can provide a prediction/recommendation about interesting items for the end user but w.r.t ITR [11] uses content-based filtering algorithms.
Our reference system is ITem Recommender (ITR) whose strategy is to shift from a keyword-based document representation to a sense-based one in order to integrate lexical knowledge in the indexing step of training documents. Several methods have been proposed to accomplish this task. Inn [21] is proposed to include WordNet information at the feature level by expanding each word in the training set with all the synonyms for it, including those available for each sense, in order to avoid a word sense disambiguation (WSD) process [28]. This approach has shown a decrease of effectiveness in the obtained classifier, mostly due to the word ambiguity problem. In [28] is pointed out that some kind of disambiguation is required in any case. Subsequent works [3], [32] show that embedding WSD in document classification tasks can improve classification accuracy.
Besides, for improving search and visualization various example-based search tools have been developed, such as SmartClient [33]. SmartClient uses constraint satisfaction techniques, allows to refine (critique) preference values specified in the first step of the search and supports trade-off analysis among different attributes, e.g., looking for an apartment a user can make a compromise between distance and rent (more distant less expensive). Also in [19] a candidate/critiques model has been presented, which allows users to refine candidate solutions proposed. Here, preferences are elicited incrementally by analyzing critiques through subsequent iterations. It is an Automated Travel Assistant (ATA) for planning airline travels, and similarly to SmartClient, ATA exploits CSP techniques: preferences are described using soft constraints defined on the values of attributes. AptDecision [31] is a tool supporting elicitation of preferences in the real estate domain: browsing the domain, users can discover new features of interest and through their refinement of apartment features, agents can build a profile of their preferences using learning techniques. FindMe [6] uses case-based reasoning as a way of recommending products in e-commerce catalogs. FindMe, and its enhanced version The Wasabi Personal Shopper [4], combines instance-based browsing and tweaking by difference. Different FindMe-like systems have been developed, in various domains. Among systems based on FindMe the most renowned is Entre´e [5], a restaurant recommender, which allows users to refine a query on the basis of the results displayed, so it is possible to choose a restaurant less expensive or closer than the restaurant shown after the first query.
Recently, there has been a growing interest toward systems supporting semantics exploitation, in different domains. In [15] an application is presented, improving traditional web searching using semantic web technologies: two Semantic Search applications are presented, running on an application framework called TAP, which provides a set of simple mechanisms for sites to publish data onto the Semantic Web and for applications to consume these data via a query interface called GetData. The results provided by the system are then compared with traditional text search results of Google.it Web pages. Story Fountain [23] is an ontology-based tool, which provides a guided exploration of digital stories using a reasoning engine for the selection and organization of resources. Story Fountain provides support for six different exploration facilities to aid users engaged in exploration process. The system is being used by the tour guides at Bletchley Park. The approach has been further investigated in [8]. An intelligent query interface exploiting an ontology-based search engine is presented in [7]; the system enables access to data sources through an integrated ontology and supports a user in formulating a query even in the case of ignorance of the vocabulary of the underlying information system.
We do not present here related work on semantic matchmaking (the interested reader is referred to [9]) but only a framework of semantic-enabled e-marketplaces aimed at fully exploiting semantics of supply/demand descriptions in B2C and C2C e-marketplaces [13]. Main features of this framework are the followings: full exploitation of nons-tandard inferences for explanation services in the query-retrieval-refinement loop; semantic-based ranking in the request answering; fully graphical and usable interface, which requires no prior knowledge of any logic principles, though fully exploiting it in the backoffice.
A close relation exists between OWL and Description Logics. In fact, the formal semantics of OWL DL sub-language is grounded in the Description Logics theoretical studies. We assume the reader be familiar with the basics of Description Logics and with two standard inference services provided by a DL reasoner: Subsumption and Satisfiablity [2].
Given a query Q and an item to be retrieved I the following match classes can be identified with respect to an ontology T (see [12], [18], [25]).
• exact - T |= Q ≡ I. I is semantically equivalent to Q.
All the characteristics expressed in Q are presented in I and I does not expose any additional characteristic with respect to Q. • full - T |= I v Q. I is more specific than Q. All the characteristics expressed in Q are provided by I and I exposes also other characteristics both not required by Q and not in conflict with the ones in Q. • plug-in - T |= Q v I. Q is more specific than I. All the characteristics expressed in I are provided by Q and Q requires also other characteristics both not exposed by I and not in conflict with the ones in I. • potential - T 6|= I u Q v ⊥. Q is compatible with I.
Nothing in Q is logically in conflict with anything in I. • partial - T |= I u Q v ⊥. Q is not compatible with I. Something in Q is logically in conflict with some characteristic in I.
With respect to the above classification, in case of potential match a similarity measure is needed to understand “how potentially” I satisfies Q.
The semantic similarity between a query and an item to be retrieved can be computed with the aid of the algorithm rankPotential [12]. Starting from the unfolded version (i.e., normalized with respect to the reference ontology) of both the query and the item description, the algorithm is able to quantify how many information requested in the query are missing in the item description. In order to understand the approach we consider the following trivial example where the ontology is just a simple taxonomy1.
B = C D
v A
T v B u E
≡ A u F
With respect to the previous T consider the query Q = C u D
and the item I = E u B u G. Referring to the above
classification we see that Q and I are a potential match.
Unfolding T in both Q and I we obtain Q = C uB uAuE uF
and I = E u B u A u G. Since in the ontology the third one
is an equivalence axiom, we rewrite D with A u F instead
of expanding it as for B and C. Once we have the unfolded
version of Q and I we say that two pieces of information
{C,F } are missing in I in order to completely satisfy Q
(and then reach a full match). Since the maximum number
of missing pieces of information is equal to the length of the
unfolded Q, in this case five, we assign a normalized semantic
similarity score of (
rankP otential(I, Q)
rank = 1 − rankP otential(>, Q)
(
On the other hand, we consider the problem of learning user profiles as a binary text categorization task [29]. Each document has to be classified as interesting or not with respect to the user preferences. Therefore, the set of categories is C = {c+, c−}, where c+ is the positive class (user-likes ) and c− the negative one (user-dislikes). We present a method able to learn sense-based profiles by exploiting an indexing procedure based on WordNet.
We extend the classical BOW model [29] to a model in which the senses (meanings) corresponding to the words in the documents are considered as features. The goal of the WSD algorithm is to associate the appropriate sense s to a word w in document d, by exploiting its context C (a set of words that precede and follow w). The sense s is selected from a predefined set of possibilities, usually known as sense inventory, that in our algorithm is obtained from WordNet [26]. The basic building block for WordNet is the SYNSET (SYNonym SET), a set of words with synonymous meanings which represents a specific sense of a word. The text in d is processed by two basic phases: the document is first tokenized and then, after removing stopwords, part of speech (POS) ambiguities are solved for each token. Reduction to lemmas is performed and then synset identification with WSD is performed: w is disambiguated by determining the degree of semantic similarity among candidate synsets for w and those of each word in C. The proper synset assigned to w is that with the highest similarity with respect to its context of use. The 1For the sake of simplicity in this example we do not consider roles even if rankPotential is able to deal with them for ALN ontologies. semantic similarity measure adopted is the Leacock-Chodorow measure [17]. Similarity between synsets a and b is inversely proportional to the distance between them in the WordNet is-a hierarchy, measured by the number of hops in the shortest path from a to b. The algorithm starts by defining the context C of w as the set of words in the same slot of w having the same POS as w, then it identifies both the sense inventory Xw for w and the sense inventory Xj for each word wj in C. The sense inventory T for the whole context C is given by the union of all Xj . After this step, we measure the similarity of each candidate sense si ∈ Xw to that of each sense sh ∈ T and then the sense assigned to w is the one with the highest similarity score. Each document is mapped into a list of WordNet synsets following three steps.
1) each monosemous word w in a slot of d is mapped into the corresponding WordNet synset; 2) for each pair of words hnoun,nouni or hadjective,nouni, a search in WordNet is made to verify if at least one synset exists for the bigram hw1, w2i. In the positive case, the algorithm is applied on the bigram, otherwise it is applied separately on w1 and w2; in both cases all words in the slot are used as the context C of the word(s) to be disambiguated; 3) each polysemous unigram w is disambiguated by the algorithm, using all words in the slot as the context C of w.
A new version of the WSD algorithm has been recently produced [30].
The WSD procedure is used to obtain a synset-based vector space representation that we called Bag-Of-Synsets (BOS). In this model, a synset vector corresponds to a document, instead of a word vector. Each document is represented by a set of slots. Each slot is a textual field corresponding to a specific feature of the document, in an attempt to take into account also the structure of documents.
Formally, assume that we have a collection of N documents, each document being subdivided into M slots. Let m be the index of the slot, n = 1, 2, ..., N , the n-th document is reduced to 3 bags of synsets, one for each slot:
dnm = htnm1, tnm2, . . . , tnmDnm i where tnmk is the k-th synset in slot sm of document dn and Dnm is the total number of synsets appearing in the m-th slot of document dn. For all n, k and m, tnmk ∈ Vm, which is the vocabulary for the slot sm (the set of all different synsets found in slot sm). Document dn is finally represented in the vector space by M synset-frequency vectors:
fnm = hwnm1, wnm2, . . . , wnmDnm i where wnmk is the weight of the synset tnmk in the slot sm of document dn and can be computed in different ways: It can be simply the number of times synset tk appears in slot sm or a more complex TF- IDF score. Our hypothesis is that the proposed indexing procedure helps to obtain profiles able to recommend documents semantically closer to the user interests.
As a strategy to learn user profiles on BOS-indexed documents, ITem Recommender (ITR) uses a Na¨ıve Bayes text categorization algorithm to build profiles as binary classifiers (user-likes vs user-dislikes ). The induced probabilistic model estimates the a posteriori probability, P (cj |di), of document di belonging to class cj as follows:
P (cj |di) = P (cj ) Y
P (tk|cj )N(di,tk)
w∈di
where N (di, tk) is the number of times token tk occurs in
document di. In ITR, each document is encoded as a vector
of BOS, one for each slot. Therefore, equation (
P (di) m=1 k=1 P (cj ) Y|S| |bim|
Y P (tk|cj , sm)nkim
where S= {s1, s2, . . . , s|S|} is the set of slots, bim is the
BOS in the slot sm of di, nkim is the number of occurrences
of token tk in bim. Training is performed on BOS-represented
documents, thus tokens are WordNet synsets, and the induced
model relies on synset frequencies. To calculate (
r − 1 M AX − 1 ; w−i = 1 − w+i
The scores in (
Based on the previous techniques we built a system (see 1)
enabling users to perform:
• semantic searching by selecting ontology classes and
properties;
• personalized searching based on user profiles and item
information;
• semantic-personalized searching obtained by combining
the two types of searching;
Witten-Bell smoothing [34] is adopted to compute
P (tk|cj , sm), by taking into account that documents are
structured into slots and that token occurrences are weighted
using scores in equation (
NP(tk,cj,sm)
i N(ti,cj,sm) if N (tk, cj , sm) 6= 0
Pˆ(tk|cj , sm) =
P Vcj 1
Vcj + i N(ti,cj,sm) V −Vcj
otherwise
(
The first set is composed of intrinsic information. For example, in a bibliographic research scenario, intrinsic information could be: item identification number, authors, title, abstract, slot types. The second set is composed of information produced by the classifier during the training step: the bag of synsets obtained by the WSD process on each slot; • user profile - learned by the ITR system, as described in the previous section; • item semantic description - an OWL description of the reference ontology.
All the above information is stored in a repository, while the adopted reference vocabulary to define semantic profile is the WordNet lexical database. Recommender system architecture is composed of several modules and each one has a specific role and instantiates a part of the repository. The Interface Module allows to define semantic item description and user request. It provides a GUI to browse the hierarchy of concepts and to outline properties of the selected class. This feature of the GUI supports a user to define requests as descriptions which are logically consistent w.r.t. the reference ontology. The user request is split into two main parts: • full : in the full part of the request, the user sets the constraints she wants to be satisfied by (in full match with) the retrieved items. • potential : here the user sets her preferences, i.e., her wishful options. The ontology-based score is computed measuring “how many” of these constraints are satisfied by a retrieved item.
We stress the fact that in order to fulfill user needs the research process could be an iterative one. On the basis of returned results, a user is able to refine the previous request exchanging a full constraint for a potential one and vice versa. Our recommender system considers user query features as full constraints by default and as potential only whenever explicitly stated by the user [10].
The Profile Engine Module prepares the item information by performing WSD on the textual descriptions of the items to be recommended. Mainly built on the ITR system, it performs the training step on the disambiguated text in order to infer the user profiles which will be exploited in the recommendation process. Each inferred user profile is a binary classifier able to categorize an item as interesting or not interesting according to the classification score of the class user-likes .
The Match Engine Module implements matchmaking and ranking algorithms. It allows to compare the query with the description of items referring to the same OWL ontology. The reasoner is not embedded within the application, so the Ontology Manager communicates with the Matchmaker via a DIG 1.1 interface over HTTP. The Match Engine is used to manage full and potential constraints. We think that when the user sets a request characteristic as full , she wants to be sure that full feature are explicitly mentioned in the item description to be retrieved. Matchmaker Engine evaluates full or potential match in the following way: • all potential constraints - if user sets all request features as potential ones then returned items will be those potentially satisfying user request. According to an Open World Assumption, returned items could have additional or missing features w.r.t the request but no features in logical conflict with any in the request. In this case the module runs a potential match; • all full constraints - if user sets all request features as full ones then returned items have to express explicitly at least all requested features. Of course this does not prevent the item description to include also not requested features. In this case the module evaluates a full match; • mixed - in this case the module first compute a full matches considering a temporary request composed of full features only. The set of items returned in the previous step could contain also features in logical conflict with someone in the potential request, so the module runs a potential match with the potential part of the request to discard these results.
Results returned by Profile Engine are defined by the
pair hitem identifier, relevance ratei whilst the ones
returned by Match Engine are defined by the pair
hitem identifier, rank valuei according to equations (
The Recommender Manager allows the communication and synchronization between Profile and Match Engine. This module is designed to deal with two issues: results ordering and synchronization. In web-based retrieval systems, ranked lists are surely preferable to unordered sets of items. Nevertheless they become less effective and usable as the complexity of the item description increases, together with possible user preferences. At least two implementations of the Recommender Engine are possible: 1) Recommender Manager operates in two steps. In the first step, this module receives user requests. They are OWL description translated in DIG; then it activates Profile and Match Engine. In the second step, Recommender Engine fetches user request results by two previous modules, then it computes a unique score adding match rank and classifier relevance rate value. The new ranked results will be sent to the GUI; 2) Match Engine works as a filter to produce a subset of items which is sent to Profile Engine. Profile Engine uses only this subset to answer to the user request.
Our recommender system implements the first approach and computes results score according to the following simple formula: score = α ∗ relevance + β ∗ rank where α and β are numeric coefficients. As an initial attempt we set both of coefficients to 0.5 value. The evaluation of several experimental tests and usage of different reference ontologies could require the change of the previous coefficients values.
As real scenario we propose queries formulated in terms of an ontology which models the bibliographic research domain [16]. The ontology is defined by the following classes (Figure 2): • Item - has subclass as Article, inProceeding, Book; • Event - has subclass as Conference, Workshop,
Meeting; • Topic - topic hierarchy will be defined by specific research topic and we may use Computer Science terms of ACM Topic Hierarchy [1] for example; • Author - author hierarchy will be defined by sev• Project - project hierarchy will be defined by specific research project adding properties as financedBy; Besides, the item concept can be defined by the following properties: aboutProject - has project class as range -, hasPublicationYear, presentedAt - has event class as range -, hasAuthor - has author class as range -, developedBy - has organization class as range -, hasTopic - has topic class as range. Obviously this domain ontology can be extended by specific journal item properties like hasVolume, hasMonth and by specific book item properties like hasPublisher and hasEdition.
Finally, in order to retriev only interesting items several disjoint sets are defined. For example Book is disjoint by Article, inProceeding, inBook, inCollection and Proceedings while inProceeding is disjoint by Book, Thesis, Booklet and Manual.
In this scenario a user can propose queries such as the combination of (a)”I’m looking for an inProceeding item published after 2004, developed in a research project by both Enterprise and University” and (b)”I am interested in items with matchmaking as keyword in title, including inProceeding u (≥ my profile”. In the previous request (a) is the semantic query for matchmaker, (b) is the profile-based query for the classifier. According to domain ontology the previous semantic query (a) is translated in the following DL description
2004 hasPublicationYear) u ∀aboutProject.(ResearchProject u ∀developedBy. (Enterprise u University)).
V. CONCLUSION
In this paper, we have described a strategy to design an advanced semantic search engine able to combine logic based matchmaking with Bayesian text categorization. The use of Wordnet has been exploited in order to enhance a text based categorization exploited by a Bayesian approach for automated user profile learning. Combining probabilistic and logic based similarity measure we have shown how to compute a score representing the match degree between a query and an item description. The score takes into account both the ontologybased query and the user profile.
An initial prototype implementing the proposed approach has been developed and presented in the paper. Currently, we are performing experiments on large datasets in order to validate the proposed approach.