In this paper, we propose a new approach to entity-based information retrieval by exploiting semantic annotations of documents. With the increased availability of structured knowledge bases and semantic annotation techniques, we can capture documents and queries at their semantic level to avoid the high semantic ambiguity of terms and to bridge the language barrier between queries and documents. Based on various semantic interpretations, users can re ne the queries to match their intents. By exploiting the semantics of entities and their relations in knowledge bases, we propose a novel ranking scheme to address the information needs of users.
The ever-increasing amount of semantic data on the Web pose new challenges but at the same time open up new opportunities for information access. With the advancement of semantic annotation technologies, the semantic data can be employed to signi cantly enhance information access by increasing the depth of analysis of current systems, while traditional document search excels at the shallow information needs expressed by keyword queries and the meaningful semantic annotations contribute very little. There is an impending need to exploit the currently emerging knowledge bases (KBs), such as DBpedia and Freebase, as underlying semantic model and make use of semantic annotations that contain vital cues for matching the speci c information needs of users.
There is a large body of work that automatically analyzes documents and the analysis results, such as part-of-speech tags, syntactic parses, word senses, named entity and relation information, are leveraged to improve the search performance. A study [1] investigates the impact of named entity and relation recognition on search performance. However, this kind of work is based on natural language processing (NLP) techniques to extract linguistic information from documents, where the rich semantic data on the Web has not been utilized. In [2], an ontology-based scheme for semi-automatic annotation of documents and a retrieval system is presented, where the ranking is based on an adaptation of the traditional vector space model taking into account adapted TF-IDF weights.
This work can be dedicated to research in this area. Nevertheless, it provides a
signi cantly new search paradigm. The main contributions include: (
In this section, we present our document retrieval process, which consists of ve steps. While lexica extraction and text annotation are performed o ine, entity matching, query re nement and document ranking are handled online based on the index generated by o ine processing.
Lexica Extraction. In this step, we constructed the cross-lingual lexica by exploiting the multilingual Wikipedia to extract the cross-lingual groundings of entities in KBs, also called surface forms, i.e., words and phrases in di erent languages that can be used to refer to entities [3]. Besides the extracted surface forms, we also exploit statistics of the cross-lingual groundings to measure the association strength between the surface forms and the referent entities.
Text Annotation. The next step is performed to enrich documents with entities in KBs to help to bridge the ambiguity of natural language text and precise formal semantics captured by KBs as well as to transform documents in di erent languages into a language independent representation. For this purpose, we employ our cross-lingual semantic annotation system [4] and the resulting annotated documents are indexed to make them searchable with KB entities.
Entity Matching. Our online search process starts with the keyword query in a speci c language. Instead of retrieving documents, our approach rst nds entities from KBs matching the query based on the index constructed in the lexica extraction step. These entities represent di erent semantic interpretations of the query and thus are employed in the following steps to help users to re ne the search and in uence document ranking according to their intents.
Query Re nement. Di erent interpretations of the query are presented for users to select the intended ones. Since interpretations correspond to entities in this step, users can choose the intended entity for re nement of their information needs. We also enable users to adjust the weights of entity relations to in uence the document ranking for a personalized document retrieval. For this, the chosen entity is shown and extended with relations to other entities retrieved from KBs.
Document Ranking. After query re nement by users, the documents in di erent languages containing the chosen entity are retrieved from the index constructed by text annotation. Then, we exploit the semantics of entities and relations for ranking. We observe that annotated documents generally share the following structure pattern: every document is linked to a set of entities, where where LCCde is the largest connected component of d containing e and jLCCdej represents the number of entities in LCCe.
d
Relation-Based Ranking: Given the chosen entity e, the users can weight both the existence and the occurrence frequency of its relations to in uence the document ranking. This di erentiation separates the one scenario where users are interested in obtaining more detailed information about the relationship (qualitative information) from the other, where users are interested in the quantity. Let Re be the set of relations of chosen entity e. We de ne xr = 1 jrdj if r 2 Re, otherwise 0, and yr = log(avgr) , where jrdj denotes the occurrence frequency of r in d and avgr is the average occurrence frequency of r. Then, we propose ScoreRelation(d; e) between document d and entity e to capture the relevance of d to the weighted relations in Re as follows:
ScoreRelation(d; e) = X xr wrexistence + yr wrfrequency
(
By taking into account both focus-based and relation-based ranking, we present the nal function for scoring the documents as given in Eq 3. a subset (several subsets) of these entities are connected via relations in the KB, forming a graph (graphs). In this regard, a document can be conceived as a graph containing several connected components. Leveraging this pattern, we propose a novel ranking scheme based on the focus on the chosen entity and the relevance to the weighted relations.
Focus-Based Ranking: Intuitively, given two documents d1 and d2 retrieved for the chosen entity e, d1 is more relevant than d2 if it focuses more on e than d2 does, i.e., when the largest connected component of d1 containing e is larger than that of d2. Based on this rationale, we propose ScoreF ocus(d; e) between document d and entity e to capture the focus of d on e as follows:
ScoreF ocus(d; e) = jLCCdej where efd denotes the total number of entities in d, avgef is the average number of entities in the document collection, and s is a parameter taken from IR literature, which has been typically set to 0:2. Score(d; e) =
ScoreF ocus(d; e) ScoreRelation(d; e)
ndlde
where ndlde is the normalized document length of d w.r.t. annotations, i.e. the
number of entities contained in d, which is used to penalize documents in
accordance with their lengths because a document containing more entities has
a higher likelihood to be retrieved. The e ect of this component is similar to
that of normalized document length w.r.t. terms in IR. We can compute it as
ndlde = (1
s) + s
efd
avgef
(
We now discuss our preliminary evaluation results. In the experiment, we use DBpedia [5] as the KB and Reuters Corpus Volume 1 (RCV1) as the document corpus containing about 810,000 English news articles. To assess the e ectiveness of our approach, we investigate the normalized discounted cumulative gain (nDCG) measure of the top-k results instead of the common measures like precision and recall, which are not suitable to our scenario because the results can be di erent in relevance for each query and di er for each facet or weight used. We asked volunteers to provide keyword queries in Chinese (17 in total) along with descriptions of the intents used to set the weight for the relations, which yield the average nDCG of 0:87 and the average number of results of 612. 4
In this paper, we show that the semantics captured in KBs can be exploited to allow the information needs to be speci ed and addressed on the semantic level, resulting in the semantic representations of documents and queries, which are language independent. The user feedback on our demo system [6] suggests that the proposed approach enables more precise re nement of the queries and is also valuable in terms of the cross-linguality. In the future, we plan to advance the query capability to support keyword queries involving several entities and conduct more comprehensive experiments to evaluate our system. Acknowledgments. This work is supported by the European Community's Seventh Framework Programme FP7-ICT-2011-7 (XLike, Grant 288342) and FP7-ICT-2013-10 (XLiMe, Grant 611346). It is also partially supported by the German Federal Ministry of Education and Research (BMBF) within the SyncTech project (Grant 02PJ1002) and the Software-Campus project \SUITE" (Grant 01IS12051).