Nowadays, research group members are often overwhelmed by the large amount of relevant literature for any given topic. The abundance of publications leads to bottlenecks in curating and organizing literature, and important knowledge can be easily missed. While a number of tools exist for managing publications, they generally only deal with bibliographical metadata and their support for further content analysis is limited to simple manual annotation or tagging. Here, we investigate how we can go beyond these approaches by combining semantic technologies, including natural language processing, within a user-friendly wiki systems to create an easy-to-use, collaborative space that facilitates the semantic analysis and management of literature in a research group. We present the Zeeva system as a rst prototype that demonstrates how we can turn existing papers into a queryable knowledge base.
The increasing growth of scienti c publications has become the centre of attention of researchers from various domains, ranging from cognitive studies to computational linguistics. The apparent disproportion of human capabilities versus the pace of information generation has encouraged researchers to look for new approaches that can help to extract, organize, and manage knowledge from the immense amount of publications available in ever-growing repositories. To overcome this bottleneck, we envision a collaborative, wiki-based solution for the semantic management of research literature that integrates (i) a web-based interface, (ii) semantic knowledge representation, and (iii) text mining for automatic content analysis. Here, we report on the feasibility and usability of such an approach, based on a rst prototype called Zeeva.
As a running example, consider a research group where various members work collaboratively on a speci c topic. An ongoing task is the curation of relevant existing research publications, including background and related work. These need to be systematically organized, stored with bibliographical metadata, and shared in a way that allows team members to search, annotate, and comment on speci c works. There exists a multitude of bibliographical management systems that o er basic functionality for this, both as web-based solutions (e.g., BibSonomy1) or
a self-hosted system (e.g., Aigaion2). However, none of these tools provide for semantic management of information that goes beyond managing bibliographical metadata or simple tagging, for example, to list the contributions, claims, hypotheses, or results stated in a paper. Representing these concepts explicitly would facilitate semantic linking, querying, and analyzing a body of research, and allow to relate the ndings with a research topic under investigation (e.g., \semantic publishing") within a research group. By using semantic standards like RDF3 we can open up knowledge \bottled up" in a paper to tools and methods from the semantic web, including semantic browsing, search, and visualization.
Ideally, information such as claims or contributions would be explicitly marked
up in published research. However, this is unfortunately not the case for nearly
all existing papers. Instead of relying on our research team members to provide
these semantic annotations manually, we aim to support them with text mining
pipelines that can automatically analyze and extract structural and semantical
information from research papers. While a number of text mining tools and
pipelines exist, none of them have so far been seamlessly integrated into a
research literature management platform suitable for a research group. In our
work, we apply the \Semantic Assistants" approach [
The rest of this paper is organized as follows: The fundamentals of the techniques used in the Zeeva system are iterated in Section 2. Section 3 provides a high-level description of the Zeeva infrastructure followed by its implementation details in Section 4. The Zeeva wiki user interface is shown in Section 5, where we describe how users can use the various components in Zeeva to e ectively curate scienti c publications. 2
In this section, we brie y introduce two fundamental concepts for our work: semantic wikis and natural language processing. 2.1
Wikis are web applications that allow people to add, modify or delete content
in a collaborative manner. Using a web browser interface and a simple markup
language, wiki users can create and hyperlink wiki pages, making wikis \quick"
and easy-to-use for authoring documents. Semantic wikis extend the idea of a
collaborative authoring environment, where the content that is written for human
reading purposes is combined with an underlying knowledge model that describes
wiki content in a formal language suitable for automatic machine processing
techniques. Among the existing semantic wikis, Semantic MediaWiki (SMW) [
a notable example. It extends the MediaWiki4 engine functionality by introducing a special markup that can be used to create semantic triples from the knowledge contained in a wiki.
In SMW, the underlying ontology is formed as semantic metadata and inserted into wiki pages by human users. A speci c markup notation is manually typed into wiki pages to describe a property and its related value. Internally, SMW creates a semantic triple from the existing markup in the page and stores it in a relational database that can be queried from within the wiki pages using so-called inline queries. For example, users can dynamically create lists of cities located in a speci c country with a population of over a million. Such queries directly make use of the semantic metadata available in the wiki repository to create and update such lists, thereby removing the overhead of manually maintaining the results by human users. While such capabilities indeed increase the usefulness of wikis in di erent applications, the downside of this approach is that the semantic markup has to be manually provided and maintained by human users. 2.2
Natural Language Processing (NLP) is a fast-moving domain of research that
uses various techniques from the Arti cial Intelligence and Computational
Linguistics areas to process text written in natural languages. NLP is a broad term
encompassing general-purpose text processing techniques like text segmentation,
to domain-speci c applications, such as question-answering systems. One
application from the NLP domain is text mining. Text mining aims at extracting
high-quality structured information from usually free form text and represent
it in a (semi-)structured format. As an essential part of the literature curation
process, text mining techniques proved to be e ective in terms of the time needed
to extract and formalize the knowledge contained within a document. As the
use of NLP techniques in software is being gradually adopted by developers,
various applications have emerged to enable software engineers to integrate NLP
capabilities in their applications, e.g., based on web services. To facilitate the
development of re-useable components and their con guration into NLP pipelines,
frameworks, such as the General Architecture for Text Engineering (GATE) [
We now list the requirements for our collaborative semantic literature analysis platform and then describe the design that we derived from these requirements.
:hasAuthor
:hasTitle #JohnDoe "Towards a semantic ..."
SeDmataanbtaicsWeiki Coming back to our scenario from the introduction, we can identify three major requirements for semantic literature management: Centralized Repository of Knowledge (R1). Scientists need a tool that can manage di erent types of artifacts generated throughout an analysis task (e.g., articles, bibliography data, metadata, personal notes) in a centralized repository. Therefore, the proposed system must provide users with the ability to store raw data (the original articles), as well as any information generated by users (e.g., textual annotations) and analysis tools (e.g., automated extraction of contributions). Automatic Text Analysis Support (R2). Di erent tasks in literature analysis can be supported with various automatic text processing techniques. These techniques are themselves diverse in their concrete implementation and the resources they use. Hence, the proposed system must provide access to various NLP pipelines in a uni ed manner.
Collaborative Analysis Environment (R3). Many scienti c articles are the result of collaborative studies between two or more researchers. Therefore, the proposed system shall provide an environment where all researchers have access to the most up-to-date information and can keep track of content modi cations. 3.2
Based on these requirements, we made three fundamental design decisions: Provide a wiki-based interface to support collaboration (R3), use a semantic engine for knowledge representation (R1) and integrate both with text mining services (R2).
oped a wiki-based application as an evaluation platform for experimenting how various semantic services can support di erent user groups in concrete literature analysis tasks like reviewing papers. Wiki systems like MediaWiki are lightweight, collaborative authoring environments that are easy-to-use and highly scalable.
the Zeeva system is to formalize the knowledge contained in scienti c publications (R1). Fig. 1 shows the work ow of the system in an example scenario where a researcher wants to store some bibliographic metadata as well as a list of contributions of a paper into the wiki.
To provide for a semantic representation of a publication, semantic markup needs to be created (manually or automatically) and entered into a repository. To support the semantic representation, we selected a semantic wiki engine, as it can be seamlessly integrated with the wiki user interface (R3) on the one hand and the text mining pipelines on the other hand (R2). In a semantic wiki, semantic markup for the content that was found is saved in a wiki page. Upon saving the page, the underlying wiki engine processes the semantic markup present in the page and transforms the natural language representation of the semantic metadata into RDF triples using its custom markup parsers. The semantic triples are then stored in the wiki repository where they can be queried directly within the wiki environment. They also become accessible from external applications through an RDF feed.
Text Mining Pipelines for Literature Analysis. The described scenario above heavily relies on human capabilities in reading text and formalizing the extracted knowledge in a semantic wiki. Luckily, a multitude of NLP techniques, such as Information Extraction (IE) already exist that can support researchers in automatically extracting entities of interest from text (R2). Consequently, we designed the Zeeva platform in a way that arbitrary NLP pipelines can be seamlessly made available to researchers within the wiki interface, eliminating any unnecessary context switching to an external NLP applications during a researcher's work ow.
Be leveraging the service-oriented architecture of the Semantic Assistants [
Readability Metrics: This pipeline calculates standard readability metrics,
like Flesch and Kincaid [
Claims and Contribution Extraction: A custom pipeline designed speci cally for the purpose of literature analysis. It targets the particular need of researchers to automatically extract claims and contributions of a given paper in a verbatim format. Such extracted metadata can be used to nd related work in their domain of interest.
info/munpex
MediaWiki Engine
M e d i a W ii k A P I
Z xvsaeeneno E t i Web Server Database Zeeva Wiki
W ii k − N L P C o m p o n e n t
W e b S e r v e r
NLP Service Connector Service Invocation
Service Information Semantic Assistants
Wiki
Ontologies
Language
Service
Descriptions
User
These NLP pipelines are seamlessly integrated with the wiki user interface, based
on our Wiki-NLP integration architecture [
In this section, we explain the technical details of the Zeeva system implementation and illustrate how NLP results are transformed into semantic metadata. 4.1
The front-end of the Zeeva system is a wiki application that provides users with a platform to store and record their ndings with versioning mechanisms. Users interact with the Zeeva wiki using their Web browser and can view and edit the wiki content using a simple markup language. The core functionality of the MediaWiki engine powering our Zeeva system can be extended through installing extensions. The Semantic MediaWiki7 (SMW) extension allows us to use a special markup in wiki pages to annotate parts of their content with formal descriptions. Each semantic markup translates into a semantic triple with the wiki page as the subject, the declared property as the predicate and the given value as the object. SMW stores the generated triples in the wiki repository that can be later queried both within the wiki, as well as from external applications through an RDF feed.
In addition, we have designed Zeeva Pubs, a custom extension for MediaWiki that allows wiki users to invoke arbitrary NLP pipelines on a given article. The Zeeva Pubs extension uses the MediaWiki API to communicate with users through the wiki interface. Through a special page provided by the extension, users can specify a document URL for an analysis task, along with a list of NLP pipelines that may aid them at their task at hand. The extension then sends the user-provided content to the Semantic Assistants server via a web service call, where it is received by the Wiki-NLP component. The Semantic Assistants framework then takes care of executing the designated pipelines on the paper text and writing the results back to the Zeeva wiki.
The Wiki-NLP component in the Semantic Assistants framework also bears the responsibility of transforming NLP pipelines output to semantic metadata. For example, when an NLP pipeline generates a readability score for a given paper, in addition to writing the score into the wiki page for displaying to human users, it generates a semantic triple with a hasReadabilityScore predicate and makes it persistent in the wiki database. This way, papers analyzed by the NLP pipelines will be implicitly enriched with metadata automatically extracted from their content and formalized for machine processing purposes. Fig. 2 provides a high-level overview of the Zeeva system architecture. 4.2
Having described the Zeeva system architecture, we now have to speci cally de ne how we make use of the NLP pipelines results in order to generate semantic metadata in the wiki. In this paper, we treat the NLP pipelines as a black box, i.e., we do not describe the text mining techniques used in them. Rather, we are interested in seeing how the collaboration between the users and the Zeeva text mining pipelines can help with a given analysis task.
The input and output type of each NLP pipeline in the Semantic Assistants framework is precisely de ned using the Semantic Assistants ontology. Once a user invokes a pipeline through the Zeeva wiki interface, a RESTful service request with the URL of the paper and a list of pipeline names is sent to the Semantic Assistants server. The Semantic Assistants server then fetches the content of the paper and executes all the user-selected pipelines one by one on the provided content. When the execution is nished, an XML document containing the analysis results is generated (Fig. 3) and sent to the Wiki-NLP component. <saResponse> <annotation type="Title"> <document url="http://www.semanticsoftware.info/.../mobiwis13_android.pdf"> <annotationInstance content="Smarter Mobile Apps through Integrated ..."/> ...
</document> </annotation> </saResponse>
Since the resulting XML document cannot be written directly to the wiki database, nor it is suitable for human reading, the Wiki-NLP component has to parse the XML document into another representation format. In Zeeva, we make use of the MediaWiki templating mechanism. The Wiki-NLP component transforms the pipelines' results to wiki-speci c markup by parsing the XML document and embedding the output in pre-de ned templates in the wiki. These templates are provided through installing our Zeeva Pubs extension, and de ne (i) the look and feel of the results when embedded in wiki pages and (ii) the semantic metadata that should be attached to each pipeline output instance, since every eld can be associated with a semantic property. Fig. 4 shows a Publication template with NLP pipeline results embedded in it. {{Publication |Title = [[hasTitle :: Smarter Mobile Apps through Integrated Natural Language ...]] |Author = [[hasAuthor :: Bahar Sateli, Gina Cook, and Rene Witte]] ... }}
The template markup is then written to the Zeeva database, where it can be accessed by users, either through a browser (Fig. 7) or exported by the SMW engine as an RDF document (Fig. 5). <rdf:RDF> <owl:Ontology rdf:about="http://localhost/Zeeva/.../Sateli-MOBIWIS2013"> <owl:imports rdf:resource="http://semantic-mediawiki.org/swivt/1.0"/> </owl:Ontology> <swivt:Subject rdf:about="http://localhost/Zeeva/.../Sateli-2DMOBIWIS2013"> <rdf:type rdf:resource="http://localhost/Zeeva/.../Category-3APublication"/> <property:HasTitle rdf:datatype="http://www.w3.org/2001/XMLSchema#string">
Smarter Mobile Apps through Integrated Natural Language Processing Services </property:HasTitle> ...
<!-- Created by Semantic MediaWiki, http://semantic-mediawiki.org/ --> </rdf:RDF> We refer back to our running example from the introduction to show how a group of researchers can use the Zeeva system. The task under study is to obtain an overview of the research in a particular domain through curating relevant publications. More speci cally, our researchers need to systematically organize the papers that they have found on the web and for each article extract (i) the bibliographical metadata, and (ii) a list of claims and contributions of that paper.
A special page in Zeeva wiki, shown in Fig. 6, allows users to provide a URL and a desired page name for the paper to be analyzed, as well as selecting one or multiple NLP assistants for the analysis task. Provided that the Wiki-NLP integration has adequate permissions to retrieve the article (e.g., from an open access repository or through an institutional license), the article is then passed on to all of the NLP pipelines chosen by the user in the Semantic Assistants server. Once all the pipelines are executed, the user is automatically redirected to the newly created page with the analysis results transformed into user-friendly representations, like lists or graphs. Fig. 7 shows the wiki page created with bibliographical metadata, like title and author names, extracted from a paper.
As for the semantic entities, Fig. 8 shows two rhetorical entities, namely, claims and contributions, automatically extracted by the \Claims and Contribution Extraction" text mining pipeline. Since Zeeva's underlying wiki engine revisions all changes to wiki pages, users can review the pipelines' output and modify them in case of erroneous results.
In our example, multiple users can follow the same process to automatically extract bibliographic and rhetorical entities from their designated papers. By exploiting the metadata from analyzed papers, the research team can now obtain an overview of the existing papers in the wiki by looking at their bibliographical data and their contained claims and contributions. Fig. 9 shows the results of an example query asking for all contributions of a speci c author from the papers in the wiki, using the SMW inline query syntax: {{#ask: [[Category: Publication]] [[hasAuthor:: Bahar Sateli]] | ?hasTitle = Title | ?hasContribution = Contribution}}
Tables, such as the one shown in Fig. 9, are created automatically by querying the semantic metadata that researchers generate in the wiki together with the intelligent NLP assistants. The advantage of this approach is that not only these tables are dynamically created and kept up-to-date by the wiki system, they also allow researchers to discover related ndings present in the wiki which may have been imported and analyzed by other users of the system. Such semantic support in a collaborative environment can improve the productivity of researchers in tasks like literature reviews or nding experts. WikiPapers8 is a semantic wiki with the goal of creating \the most comprehensive compilation" of literature focused on research of wikis through a community of volunteers. Users can create new entries for publications, journals, authors, event and datasets using semantic forms. Thereby, dynamic lists of publications by category, author or keywords can be generated and maintained online for researchers. AcaWiki9 is another semantic wiki system designed to \collect summaries and literature reviews of peer-reviewed academic research" and make them available to the general public. Any user can post a summary about an article on AcaWiki and provide additional bibliographic and practical relevance data (e.g., links to related news articles or blog posts) using the provided semantic forms. Although pursuing a similar goal of formalizing the body of knowledge contained in scienti c publications, our approach does not rely solely on human users as primary providers of semantic metadata, but o ers an innovative approach where users can bene t from state-of-the-art techniques from the natural language processing domain in the metadata generation process.
Orthogonal to the development of literature analysis tools, frameworks like
SALT [
writing a document. While the development of such ontologies are e ective steps towards providing interoperability of the extracted semantic entities, there is still no tool support that can help authors automatically enrich their documents with such markup. The focus of our approach in this research work, is to assess the feasibility of a semantic wiki-based literature analysis environment with integrated text mining support, rather than constructing a new ontology. 7
Research groups are in dire need of novel methods and tools for managing scienti c publications that go beyond simply storing bibliographical data. We propose Zeeva, a proof-of-concept system that demonstrates how the next generation of literature management tools can support research groups by transforming publications into an active knowledge base. With the notion of embedded \Semantic Assistants" performing text analysis, Zeeva plays the role of an intelligent agent within a collaborative research task.
There are a number of further steps, both in ongoing research and implementation. We plan to develop and integrate additional NLP pipelines that further automate the analysis of research publications. We are investigating the re-use of existing RDF Schemas and OWL ontologies for research publications and their annotation, in order to integrate them for a Linked Data context. Finally, we will perform a user study on a large group, to investigate what are currently the most time-consuming tasks and how much precisely a tool like Zeeva can help in terms of both e ort and accuracy.