Towards Semantic Recommendation of Biodiversity Datasets based on Linked Open Data

Abstract

Conventional content-based ltering methods recommend documents based on extracted keywords. They calculate the similarity between keywords and user interests and return a list of matching documents. In the long run, this approach often leads to overspecialization and fewer new entries with respect to a user's preferences. Here, we propose a semantic recommender system using Linked Open Data for the user pro le and adding semantic annotations to the index. Linked Open Data allows recommendations beyond the content domain and supports the detection of new information. One research area with a strong need for the discovery of new information is biodiversity. Due to their heterogeneity, the exploration of biodiversity data requires interdisciplinary collaboration. Personalization, in particular in recommender systems, can help to link the individual disciplines in biodiversity research and to discover relevant documents and datasets from various sources. We developed a rst prototype for our semantic recommender system in this eld, where a multitude of existing vocabularies facilitate our approach.

Keywords

• content ltering
• diversity
• Linked Open Data
• recommender systems
• semantic indexing
• semantic recommendation

Categories and Subject Descriptors

1. INTRODUCTION

Content-based recommender systems observe a user's browsing behaviour and record the interests [ 1 ]. By means of natural language processing and machine learning techniques, the user's preferences are extracted and stored in a user pro le. The same methods are utilized to obtain suitable content keywords to establish a content pro le. Based on previously seen documents, the system attempts to recommend similar content. Therefore, a mathematical representation of the user and content pro le is needed. A widely used scheme are TFIDF (term frequency-inverse document frequency) weights [ 19 ]. Computed from the frequency of keywords appearing in a document, these term vectors capture the in uence of keywords in a document or preferences in a user pro le. The angle between these vectors describes the distance or the closeness of the pro les and is calculated with similarity measures, like the cosine similarity. The recommendation lists of these traditional, keyword-based recommender systems often contain very similar results to those already seen, leading to overspecialization [ 11 ] and the \Filter-Bubble"-e ect [ 17 ]: The user obtains only content according to the stored preferences, other related documents not perfectly matching the stored interests are not displayed. Thus, increasing diversity in recommendations has become an own research area [ 21, 25, 24, 18, 3, 6, 23 ], mainly used to improve the recommendation results in news or movie portals.

One eld where content recommender systems could enhance daily work is research. Scientists need to be aware of relevant research in their own but also neighboring elds. Increasingly, in addition to literature, the underlying data itself and even data that has not been used in publications are being made publicly available. An important example for such a discipline is biodiversity research, which explores the variety of species and their genetic and characteristic diversity [ 12 ]. The morphological and genetic information of an organism, together with the ecological and geographical context, forms a highly diverse structure. Collected and stored in di erent data formats, the datasets often contain or link to spatial, temporal and environmental data [ 22 ]. Many important research questions cannot be answered by working with individual datasets or data collected by one group, but require meta-analysis across a wide range of data. Since the analysis of biodiversity data is quite time-consuming, there is a strong need for personalization and new ltering techniques in this research area. Ordinary search functions in relevant data portals or databases, e.g., the Global Biodiversity Information Facility (GBIF)1 and the Catalog of Life,2 only return data that match the user's query exactly and fail at nding more diverse and semantically related content. Also, user interests are not taken into account in the result list. We believe our semantic-based content recommender system could facilitate the di cult and time-consuming research process in this domain.

Here, we propose a new semantic-based content recommender system that represents the user pro le as Linked Open Data (LOD) [ 9 ] and incorporates semantic annotations into the recommendation process. Additionally, the search engine is connected to a terminology server and utilizes the provided vocabularies for a recommendation. The result list contains more diverse predictions and includes hierarchical concepts or individuals.

The structure of this paper is as follows: Next, we describe related work. Section 3 presents the architecture of our semantic recommender system and some implementation details. In Section 4, an application scenario is discussed. Finally, conclusions and future work are presented in Section 5.

RELATED WORK

The major goal of diversity research in recommender systems is to counteract overspecialization [ 11 ] and to recommend related products, articles or documents. More books of an author or di erent movies of a genre are the classical applications, mainly used in recommender systems based on collaborative ltering methods. In order to enhance the variety in book recommendations, Ziegler et al. [ 25 ] enrich user pro les with taxonomical super-topics. The recommendation list generated by this extended pro le is merged with a rank in reverse order, called dissimilarity rank. Depending on a certain diversi cation factor, this merging process supports more or less diverse recommendations. Larger diversi cation factors lead to more diverse products beyond user interests. Zhang and Hurley [ 24 ] favor another mathematical solution and describe the balance between diversity and similarity as a constrained optimization problem. They compute a dissimilarity matrix according to applied criterias, e.g., movie genres, and assign a matching function to nd a subset of products that are diverse as well as similar. One hybrid approach by van Setten [ 21 ] combines the results of several conventional algorithms, e.g., collaborative and case-based, to improve movie recommendations. Mainly focused on news or social media, approaches using content-based ltering methods try to present di erent viewpoints on an event to decrease the media bias in news portals [ 18, 3 ] or to facilitate the ltering of comments [ 6, 23 ].

Apart from Ziegler et al., none of the presented approaches have considered semantic technologies. However, utilizing ontologies and storing user or document pro les in triple stores represents a large potential for diversity research in recommender systems. Frasincar et al. [ 7 ] de ne semantically enhanced recommenders as systems with an underlying knowledge base. This can either be linguistic-based [ 8 ], where only linguistic relations (e.g., synonymy, hypernomy, meronymy, antonymy) are considered, or ontology-based. In the latter case, the content and the user pro le are represented with concepts of an ontology. This has the advantage

1GBIF, http://www.gbif.org

2Catalog of Life, http://www.catalogueoflife.org/col/ search/all/ that several types of relations can be taken into account. For instance, for a user interested in \geology", the pro le contains the concept \geology" that also permits the recommendation of inferred concepts, e.g., \fossil". The idea of recommending related concepts was rst introduced by Middelton et al. [ 15 ]. They developed Quickstep, a recommender system for research papers with ontological terms in the user pro le and for paper categories. The ontology only considers is-a relationships and omits other relation types (e.g., partof). Another simple hierarchical approach from Shoval et al. [ 13 ] calculates the distance among concepts in a pro le hierarchy. They distinguish between perfect, close and weak match. When the concept appears in both a user's and document's pro le, it is called a perfect match. In a close match, the concept emerges only in one of the pro les and a child or parent concept appears in the other. The largest distance is called a weak match, where only one of the pro les contains a grandchild or grandparent concept. Finally, a weighted sum over all matching categories leads to the recommendation list. This ontological ltering method was integrated into the news recommender system epaper. Another semantically enhanced recommender system is Athena [ 10 ]. The underlying ontology is used to explore the semantic neighborhood in the news domain. The authors compared several ontology-based similarity measures with the traditional TF-IDF approach. However, this system lacks of a connection to a search engine that allows to query large datasets.

All presented systems use manually established vocabularies with a limited number of classes. None of them utilize a generic user pro le to store the preferences in a semantic format (RDF/XML or OWL). The FOAF (Friend Of A Friend) project3 provides a vocabulary for describing and connecting people, e.g., demographic information (name, address, age) or interests. As one of the rst, in 2006 Celma [ 2 ] leveraged FOAF in his music recommender system to store users' preferences. Our approach goes beyond the FOAF interests, by incorporating another generic user model vocabulary, the Intelleo User Modelling Ontology (IUMO).4 Besides user interests, IUMO o ers elements to store learning goals, competences and recommendation preferences. This allows to adapt the results to a user's previous knowledge or to recommend only documents for a speci c task. 3.

DESIGN AND IMPLEMENTATION

In this section, we describe the architecture and some implementation details of our semantic-based recommender system (Figure 1). The user model component, described in Section 3.1, contains all user information. The source les, described in Section 3.2, are analyzed with GATE [ 5 ], as described in Section 3.3. Additionally, GATE is connected with a terminology server (Section 3.2) to annotate documents with concepts from the provided biodiversity vocabularies. In Section 3.4, we explain how the annotated documents are indexed with GATE M mir [ 4 ]. The nal recommendation list is generated in the recommender component (Section 3.5). 3.1

User profile

The user interests are stored in an RDF/XML format utilizing the FOAF vocabulary for general user information. In

3FOAF, http://xmlns.com/foaf/spec/

4IUMO, http://intelleo.eu/ontologies/user-model/ spec/ order to improve the recommendations regarding a user's previous knowledge and to distinguish between learning goals, interests and recommendation preferences, we incorporate the Intelleo User Modelling Ontology for an extended pro le description. Recommendation preferences will contain settings in respect of visualization, e.g., highlighting of interests, and recommender control options, e.g., keyword-search or more diverse results. Another adjustment will adapt the result set according to a user's previous knowledge. In order to enhance the comprehensibility for a beginner, the system could provide synonyms; and for an expert the recommender could include more speci c documents.

The interests are stored in form of links to LOD resources. For instance, in our example pro le in Listing 1, a user is interested in \biotic mesoscopic physical object", which is a concept from the ENVO5 ontology. Note that the interest entry in the RDF le does not contain the textual description, but the link to the concept in the ontology, i.e., http://purl. obolibrary.org/obo/ENVO_01000009. Currently, we only support explicit user modelling. Thus, the user information has to be added manually to the RDF/XML le. Later, we intend to develop a user pro ling component, which gathers a user's interests automatically. The pro le is accessible via an Apache Fuseki6 server.

Listing 1: User pro le with interests stored as Linked Open Data URIs <rdf:Description rdf:about="http://www.semanticsoftware.info/person /felicitasloeffler"> <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Person"/> <foaf: rstName>Felicitas</foaf: rstName> <foaf:lastName>Loe er</foaf:lastName> <foaf:name>Felicitas Loe er</foaf:name> <foaf:gender>Female</foaf:gender> <foaf:workplaceHomepage rdf:resource="http://dbpedia.org/page/

University_of_Jena"/> <foaf:organization>Friedrich Schiller University Jena </foaf:organization> <foaf:mbox>felicitas.loe er@uni jena.de</foaf:mbox> <um:TopicPreference rdf:resource="http://purl.obolibrary.org/obo/

ENVO_01000009"/> </rdf:Description> 3.2

Source files and terminology server

The content provided by our recommender comes from the biodiversity domain. This research area o ers a wide range of 5ENVO, http://purl.obolibrary.org/obo/envo.owl 6Apache Fuseki, http://jena.apache.org/documentation/ serving_data/ existing vocabularies. Furthermore, biodiversity is an interdisciplinary eld, where the results from several sources have to be linked to gain new knowledge. A recommender system for this domain needs to support scientists by improving this linking process and helping them nding relevant content in an acceptable time.

Researchers in the biodiversity domain are advised to store their datasets together with metadata, describing information about their collected data. A very common metadata format is ABCD.7 This XML-based standard provides elements for general information (e.g., author, title, address), as well as additional biodiversity related metadata, like information about taxonomy, scienti c name, units or gathering. Very often, each taxon needs speci c ABCD elds, e.g., fossil datasets include data about the geological era. Therefore, several additional ABCD-related metadata standards have emerged (e.g., ABCDEFG8, ABCDDNA9). One document may contain the metadata of one or more species observations in a textual description. This provides for annotation and indexing for a semantic search. For our prototype, we use the ABCDEFG metadata les provided by the GFBio10 project; speci cally, metadata les from the Museum fur Naturkunde (MfN).11 An example for an ABCDEFG metadata le is presented in Listing 2, containing the core ABCD structure as well as additional information about the geological era. The terminology server supplied by the GFBio project offers access to several biodiversity vocabularies, e.g., ENVO, BEFDATA, TDWGREGION. It also provides a SPARQL endpoint12 for querying the ontologies. 3.3

Semantic annotation

The source documents are analyzed and annotated according to the vocabularies provided by the terminology server. For this process, we use GATE, an open source framework that o ers several standard language engineering components [ 5 ]. We developed a custom GATE pipeline (Figure 2) that analyzes the documents: First, the documents are split into tokens and sentences, using the existing NLP components included in the GATE distribution. Afterwards, an `Annotation Set Transfer' processing resource adds the original 7ABCD, http://www.tdwg.org/standards/115/ 8ABCDEFG, http://www.geocase.eu/efg 9ABCDDNA, http://www.tdwg.org/standards/640/ 10GFBio, http://www.gfbio.org 11MfN, http://www.naturkundemuseum-berlin.de/ 12GFBio terminology server, http://terminologies.gfbio. org/sparql/ markups of the ABCDEFG les to the annotation set, e.g., abcd:HigherTaxon. The following ontology-aware `Large KB Gazetteer' is connected to the terminology server. For each document, all occurring ontology classes are added as speci c \gfbioAnnot" annotations that have both instance (link to the concrete source document) and class URI. At the end, a `GATE M mir Processing Resource' submits the annotated documents to the semantic search engine.

3.4 Semantic indexing

For semantic indexing, we are using GATE M mir:13 \M mir is a multi-paradigm information management index and repository which can be used to index and search over text, annotations, semantic schemas (ontologies), and semantic metadata (instance data)" [ 4 ]. Besides ordinary keywordbased search, M mir incorporates the previously generated semantic annotations from GATE to the index. Additionally, it can be connected to the terminology server, allowing queries over the ontologies. All index relevant annotations and the connection to the terminology server are speci ed in an index template.

3.5 Content recommender

The Java-based content recommender sends a SPARQL query to the Fuseki Server and obtains the interests and preferred recommendation techniques from the user pro le as a list of (LOD) URIs. This list is utilized for a second SPARQL query to the M mir server. Presently, this query asks only for child nodes (Figure 3). The result set contains ABCDEFG metadata les related to a user's interests. We intend to experiment with further semantic relations in the future, e.g., object properties. Assuming that a speci c fossil used to live in rocks, it might be interesting to know if other species, living in this geological era, occured in rocks. Another ltering method would be to use parent or grandparent nodes from the vocabularies to broaden the search. We will provide control options and feedback mechanisms to support 13GATE M mir, https://gate.ac.uk/mimir/ the user in steering the recommendation process actively. The recommender component is still under development and has not been added to the implementation yet.

Listing 2: Excerpt from a biodiversity metadata in ABCDEFG format [ 20 ] le

APPLICATION

The semantic content recommender system allows the recommendation of more speci c and diverse ABCDEFG metadata les with respect to the stored user interests. Listing 3 shows the query to obtain the interests from a user pro le, introduced in Listing 1. The result contains a list of (LOD) URIs to concepts in an ontology.

Listing 3: SPARQL query to retrieve user interests SELECT ?label ?interest ?syn WHERE f ?s foaf: rstName "Felicitas" . ?s um:TopicPreference ?interest . ?interest rdfs:label ?label .

?interest oboInOwl:hasRelatedSynonym ?syn g

In this example, the user would like to obtain biodiversity datasets about a \biotic mesoscopic physical object", which is the textual description of http://purl.obolibrary.org/ obo/ENVO_01000009. This technical term might be incomprehensible for a beginner, e.g., a student, who would prefer a description like \organic material feature". Thus, for a later adjustment of the result according to a user's previous knowledge, the system additionally returns synonyms.

The returned interest (LOD) URI is utilized for a second query to the search engine (Figure 3). The connection to the terminology server allows M mir to search within the ENVO ontology (Figure 4) and to include related child concepts as well as their children and individuals. Since there is no metadata le containing the exact term \biotic mesoscopic physical object", a simple keyword-based search would fail. However, M mir can retrieve more speci c information than stored in the user pro le and is returning biodiversity metadata les about \fossil". That ontology class is a child node of \biotic mesoscopic physical object" and represents a semantic relation. Due to a high similarity regarding the content of the metadata les, the result set in Figure 3 contains only documents which closely resemble each other.

We introduced our new semantically enhanced content recommender system for the biodiversity domain. Its main bene t lays in the connection to a search engine supporting integrated textual, linguistic and ontological queries. We are using existing vocabularies from the terminology server of the GFBio project. The recommendation list contains not only classical keyword-based results, but documents including semantically related concepts.

In future work, we intend to integrate semantic-based recommender algorithms to obtain further diverse results and to support the interdisciplinary linking process in biodiversity research. We will set up an experiment to evaluate the algorithms in large datasets with the established classi cation metrics Precision and Recall [ 14 ]. Additionally, we would like to extend the recommender component with control options for the user [ 1 ]. Integrated into a portal, the result list should be adapted according to a user's recommendation settings or adjusted to previous knowledge. These control functions allow the user to actively steer the recommendation process. We are planning to utilize the new layered evaluation approach for interactive adaptive systems from Paramythis, Weibelzahl and Mastho [ 16 ]. Since adaptive systems present di erent results to each user, ordinary evaluation metrics are not appropriate. Thus, accuracy, validity, usability, scrutability and transparency will be assessed in several layers, e.g., the collection of input data and their interpretation or the decision upon the adaptation strategy. This should lead to an improved consideration of adaptivity in the evaluation process.

ACKNOWLEDGMENTS

This work was supported by DAAD (German Academic Exchange Service)14 through the PPP Canada program and by DFG (German Research Foundation)15 within the GFBio project. 14DAAD, https://www.daad.de/de/ 15DFG, http://www.dfg.de

References

[1] F. Bakalov , M.-J. Meurs , B. Konig- Ries , B. Sateli , R. Witte , G. Butler, and A. Tsang . An approach to controlling user models and personalization e ects in recommender systems . In Proceedings of the 2013 international conference on Intelligent User Interfaces , IUI '13 , pages 49 { 56 , New York, NY, USA, 2013 . ACM.

[2] O. Celma. FOAFing the music: Bridging the semantic gap in music recommendation . In Proceedings of 5th International Semantic Web Conference , pages 927 { 934 , Athens, GA, USA, 2006 .

[3] S. Chhabra and P. Resnick . Cubethat: News article recommender . In Proceedings of the sixth ACM conference on Recommender systems, RecSys '12 , pages 295 { 296 , New York, NY, USA, 2012 . ACM.

[4] H. Cunningham , V. Tablan , I. Roberts, M. Greenwood , and N. Aswani . Information extraction and semantic annotation for multi-paradigm information management . In M. Lupu, K. Mayer , J. Tait , and A. J . Trippe, editors, Current Challenges in Patent Information Retrieval , volume 29 of The Information Retrieval Series , pages 307 { 327 . Springer Berlin Heidelberg, 2011 .

[5] H. Cunningham et al. Text Processing with GATE (Version 6 ). University of She eld, Dept. of Computer Science, 2011 .

[6] S. Faridani , E. Bitton , K. Ryokai , and K. Goldberg . Opinion space: A scalable tool for browsing online comments . In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI '10 , pages 1175 { 1184 , New York, NY, USA, 2010 . ACM.

[7] F. Frasincar , W. IJntema, F. Goossen , and F. Hogenboom . A semantic approach for news recommendation . Business Intelligence Applications and the Web: Models, Systems and Technologies, IGI Global , pages 102 { 121 , 2011 .

[8] F. Getahun , J. Tekli , R. Chbeir , M. Viviani , and K. Yetongnon . Relating RSS News/Items. In M. Gaedke, M. Grossniklaus , and O. D az, editors, ICWE , volume 5648 of Lecture Notes in Computer Science, pages 442 { 452 . Springer, 2009 .

[9] T. Health and C. Bizer . Linked Data: Evolving the Web into a Global Data Space . Synthesis Lectures on the Semantic Web: Theory and Technology . Morgan & Claypool, 2011 .

[10] W. IJntema , F. Goossen , F. Frasincar , and F. Hogenboom . Ontology-based news recommendation . In Proceedings of the 2010 EDBT/ICDT Workshops, EDBT '10 , pages 16:1 { 16 : 6 , New York, NY, USA, 2010 . ACM.

[11] P. Lops , M. de Gemmis, and G. Semeraro . Content-based recommender systems: State of the art and trends . In F. Ricci, L. Rokach , B. Shapira , and P. B. Kantor, editors, Recommender Systems Handbook , pages 73 { 105 . Springer, 2011 .

[12] M. Loreau . Excellence in ecology. International Ecology Institute, Oldendorf, Germany, 2010 .

[13] V. Maidel , P. Shoval , B. Shapira , and M. Taieb-Maimon . Ontological content-based ltering for personalised newspapers: A method and its evaluation . Online Information Review , 34 Issue 5 : 729 { 756 , 2010 .

[14] C. D. Manning , P. Raghavan , and H. Schutze. Introduction to Information Retrieval. Cambridge University Press, 2008 .

[15] S. E. Middleton , N. R. Shadbolt , and D. C. D. Roure . Ontological user pro ling in recommender systems . ACM Trans. Inf . Syst., 22 ( 1 ): 54 { 88 , Jan . 2004 .

[16] A. Paramythis , S. Weibelzahl , and J. Mastho . Layered evaluation of interactive adaptive systems: Framework and formative methods. User Modeling and User-Adapted Interaction , 20 ( 5 ): 383 { 453 , Dec . 2010 .

[17] E. Pariser. The Filter Bubble - What the internet is hiding from you . Viking , 2011 .

[18] S. Park , S. Kang , S. Chung , and J. Song . Newscube: delivering multiple aspects of news to mitigate media bias . In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI '09 , pages 443 { 452 , New York, NY, USA, 2009 . ACM.

[19] G. Salton and C. Buckley . Term-weighting approaches in automatic text retrieval . Information Processing and Management , 24 : 513 { 523 , 1988 .

[20] Museum fu r Naturkunde Berlin. Fossil invertebrates, UnitID:MB.Ga . 3895 . http://coll.mfn-berlin.de/u/MB_Ga_3895.html.

[21] M. van Setten . Supporting people in nding information: hybrid recommender systems and goal-based structuring . PhD thesis , Telematica Instituut, University of Twente, The Netherlands, 2005 .

[22] R. Walls , J. Deck , R. Guralnick , S. Baskauf , R. Beaman , and et al. Semantics in Support of Biodiversity Knowledge Discovery: An Introduction to the Biological Collections Ontology and Related Ontologies . PLoS ONE 9 ( 3 ): e89606, 2014 .

[23] D. Wong , S. Faridani , E. Bitton , B. Hartmann , and K. Goldberg . The diversity donut: enabling participant control over the diversity of recommended responses . In CHI '11 Extended Abstracts on Human Factors in Computing Systems, CHI EA '11 , pages 1471 { 1476 , New York, NY, USA, 2011 . ACM.

[24] M. Zhang and N. Hurley . Avoiding monotony: Improving the diversity of recommendation lists . In Proceedings of the 2008 ACM Conference on Recommender Systems, RecSys '08 , pages 123 { 130 , New York, NY, USA, 2008 . ACM.

[25] C.-N. Ziegler , G. Lausen, and L. Schmidt-Thieme . Taxonomy-driven computation of product recommendations . In Proceedings of the Thirteenth ACM International Conference on Information and Knowledge Management , CIKM '04 , pages 406 { 415 , New York, NY, USA, 2004 . ACM.