Despite being a flourishing field, regarding search, the contemporary online scientific publishing properly exploits mostly raw publication data (bags of words) and shallow meta-data (authors, key words, citations, etc.). The much needed economical mass exploitation of the knowledge implicitly contained in publication texts is still largely an uncharted territory. Within our long-term ambition to quell the lions there, we have made the first step with the CORAAL prototype presented in this report. The tool essentially extracts asserted publication meta-data together with the knowledge implicitly present in the respective text, integrates the emergent content and exposes it via a multiple-perspective search&browse interface.
Digital content processing has no doubt introduced a whole lot of new possibilities of dealing with scientific publications. It makes knowledge much more open and exploitable than in the old “paper times”. However, one still needs to go manually through a lot of possibly irrelevant content very often before actually finding the right answers. If we are to make the next step, it is necessary to process knowledge (i.e., concepts and their mutual relations), and not just data or shallow meta-data (i.e., chunks of free text, titles or authors).
Substantial automation of such meaning-intensive information processing is hardly possible with the current industry-strength technologies (e.g., full-text search), since they lack proper support for extraction, representation and processing of knowledge implicitly present in texts. As an illustration, imagine for instance finding a support of the claim that acute granulocytic leukemia is different from T-cell leukemia. With the current solutions, it is easy to find articles that contain both or either of the terms, however, the number of results may be quite high (e.g., 593 on PubMed). It is tedious or even impossible to go through all of them in order to find out which of them actually mention the two leukemias being different.
Methods for automated knowledge extraction than can dig more than mere
key words from text exist, however, their results are deemed to be to too noisy
and sparse to be exploited by the current state of the art without significant
manual post-processing [
The decision to apply CORAAL to the life sciences domain was motivated by our cooperation with clinicians and researchers at Masaryk Oncology Institute in Brno, Czech Republic. When figuring out how to apply our text mining and light-weight knowledge representation expertise to their needs, we articulated a particular “translational biomedical knowledge integration” use case. Essentially, the oncologists were interested in how the information in patient records links to relevant statements in the biomedical literature, and vice versa. The critical requirements are, firstly, lack or minimisation of manual efforts necessary for the knowledge extraction and integration, and, secondly, efficient, exhaustive and intuitive querying of large amounts of such integrated knowledge. CORAAL presents a first step towards a practical realisation of the use case, tackling the extraction, integration and exploitation of the publication knowledge. 1.2
The state-of-the-art applications like ScienceDirect or PubMed Central require almost no effort in order to expose arbitrary life science publications for search. However, the benefit they provide is rather limited when compared to cuttingedge approaches aimed at utilising also the publication knowledge within the query construction and/or result visualisation. Such innovative solutions may require much more a priori effort in order to work properly, though.
GoPubMed [
Textpresso [
Regarding more recent and/or scalable tools for expressive search in
biomedical literature, EBIMed [
From the overview of the state of the art in the field, it is obvious that the two biggest challenges are: 1. reliable automation of more expressive content acquisition; 2. representation of the acquired content allowing for exhaustive and expressive querying. None of the related systems addresses both problems appropriately. On the other hand, CORAAL, the system we present in this paper, can easily employ legacy domain resources or even work without any human intervention. Apart of full-text indexing, the extracted statements are integrated into a light-weight knowledge base with similarity-based inference capabilities that allow for refinement, augmentation and querying of the stored facts. In effect, CORAAL delivers a comprehensive combination of full-text and knowledgebased search in life science publications that is to large extent complementary to the related state of the art. 1.3
The rest of the paper is organised as follows. Section 2 presents the essentials of the processing pipeline in CORAAL and outlines the relevant technological aspects of the tool. The deployment and user perspectives of CORAAL are covered in Section 3. We summarise the delivered work and outline future directions in Section 4. 2
In order to provide comprehensive search capabilities in CORAAL, we decided
to complement a standard (full-text) publication search approach with advanced
services catering for knowledge-based search. By knowledge-based search we
mean querying for and browsing of expressive statements capturing relations
between concepts in the respective source articles. CORAAL is built on the top
of two substantial research outputs of our group at DERI – the KONNEX [
EUREEKA provides for knowledge extraction from text and other knowledge
resources (e.g., ontologies or machine readable thesauri). It employs two data
structures. Firstly, a lexicon takes care of mapping extracted lexical labels to
unique entity identifiers and vice versa (i.e., dealing with synonymy resolution
and disambiguation). The extraction process continually updates the second
data structure in EUREEKA – a light-weight knowledge base supporting
arbitrary extensions by custom rules, as well as similarity-based reasoning and
querying of the stored data. The knowledge bases are exposed to consumers
via a query answering interface. The interface allows for evaluating conjunctive
argument-relation-argument queries with negation and variables. It returns
extracted and/or inferred statements relevant to the query, sorted according to
their intrinsic relevance degree (derived from all particular relationships in the
knowledge base using a generalised IR ranking algorithm). Theoretical and
technical details of EUREEKA and its deployment in CORAAL are provided in [
KONNEX tackles the integration of the extracted publication text and
metadata, represented as RDF graphs in a triple store. Operations related to data
registration (inclusion and integration with the stored content), repository
maintenance, full-text query processing and indices are handled by respective manager
KONNEX modules, possibly composed of sub-modules handling particular data
or query types. Details are given in [
There are several conceptually separate modules in CORAAL, moreover, EUREEKA is written in the Python programming language, while KONNEX in Java. Therefore we utilise an inter-process communication layer implemented using the D-BUS framework (cf. http://en.wikipedia.org/wiki/D-Bus). On the top of the core-level EUREEKA and KONNEX APIs, a set of helper web services rests. These manage the user requests and forward the data returned by the core APIs to the web hub, which is a set of Java servlets handling particular types of search. The servlets produce machine-readable RDF representing answers to user queries. The RDF has XSL style sheets attached in order to render the results in a human-readable form via the Exhibit faceted browsing web frontend (cf. http://www.simile-widgets.org/exhibit/). Such a solution results in CORAAL being a pure Semantic Web application, as the data-flow between the core infrastructure and the other modules is strictly based on RDF graphs. While being presented in a human-readable form in the browser, the produced data can be directly analyzed by an application or fetched by a crawler. 3 3.1
We have processed 11,761 Elsevier journal articles from the provided XML repositories that were related to cancer research and treatment. The access to the articles was kindly provided within the Elsevier Grand Challenge competition we participated in (cf. http://www.elseviergrandchallenge.com). The domain was selected so due to the expertise of our sample users and testers from Masaryk Oncology Institute in Brno, Czech Republic. We processed articles evenly distributed across journals dedicated to the following topics: oncology, genetics, pharmacology, biochemistry, general biology, cell research, and clinical medicine. From the article repository, we extracted the knowledge and publication metadata for further processing by CORAAL. Besides the publications themselves, we employed legacy machine-readable vocabularies for the refinement and extension of the extracted knowledge (currently, we use the NCI and EMTREE thesauri – see http://www.cancer.gov/cancertopics/ terminologyresources and http://www.embase.com/emtree/, respectively).
CORAAL exposes two data-sets as an output of the publication processing: 1. We used a triple store containing publication meta-data (citations, their contexts, structural annotations, titles, authors and affiliations) associated with respective full-text indices. The resulting store contained 7, 608, 532 of RDF subject-predicate-object statements describing the input articles. This included 247, 392 publication titles and 374, 553 authors (both from full-texts and references processed). 2. We employed a custom EUREEKA knowledge base with facts of various certainty extracted and inferred from the article texts and the seed life science thesauri. Directly from the articles, 215, 645 concepts were extracted (and analogically extended). Together with the data from the initial thesauri, the domain lexicon contained 622, 611 terms, referring to 347, 613 unique concepts. The size of the emergent knowledge base was 4, 715, 992 weighed statements (ca. 99 and 334 extracted and inferred statements per publication in average, respectively). This number is significantly smaller than in the case of the semifinal prototype. However, this is due to a full integration of the knowledge from formerly separate contexts, the data themselves are still the same. The contextual meta-knowledge related to the statements (like provenance information) amounts to more than 10, 000, 000 additional statements should it be expressed in RDF triples. 3.2
CORAAL can be accessed at http://coraal.deri.ie:8080/coraal/. The following browsers have been tested with CORAAL and are known to work on most desktop configurations and operating systems: (1) Firefox (versions 2.x, 3.x and newer); (2) Internet Explorer (versions 7.x and newer; in most cases only on Windows Vista, though); (3) Opera (versions 9.6 and newer); (4) Safari (versions 3.1 and newer); (5) Google Chrome (all versions).
For the user interface of our system, we employed the MIT’s state-of-the-art Exhibit framework (cf. http://www.simile-widgets.org/exhibit/). It supports faceted browsing (cf. http://en.wikipedia.org/wiki/Faceted_browser) of the knowledge-based search results, letting users to conveniently focus on the relevant answers. Similarly, we allow for faceted browsing of the classical fulltext search results that are tightly integrated with the knowledge extracted or inferred from the respective articles. Both types of search (knowledge-based and full-text) are illustrated in the rest of the section.
Knowledge-Based Search We expose the content of the CORAAL
knowledge base via a query-answering module. It returns answer statements sorted
according to their relevance scores and similarity to the query (see [
Statement queries can be asked either directly in the CORAAL Knowledge tab, or one can use an interface for assisted query construction with autocompletion capability, as seen in Figure 1. The auto-completion is contextsensitive – if you settle, e.g., on a subject, only relations (properties) actually associated with the particular subjects in the knowledge base are offered. Negative or conjunctive statements are created using the NOT, AND check-box or button, respectively. This allows for very convenient query construction even for users who are not sure how exactly should they ask the system.
The knowledge search results (Figure 2) are displayed as particular statements provided with several types of meta-information: – source provenance – articles relevant to the statement – context provenance – sub-domain of life sciences the statement relates to (determined according to the main topic of the journal that contained the articles the statement was extracted from) – certainty – how certain the system is that the statement holds and is relevant to the query (values between 0 and 1) – inferred – whether the statement was inferred or not (i.e., directly extracted) One can filter the answer statements based on their particular elements (subjects, properties and objects), associated meta-information and their negativity. The filtering boxes have a contextual help associated with them – after pressing the (?) symbol, an inline hint on the usage of the respective box is opened. Using a particular filtering, one can conveniently focus only on statements of interest. Article provenance summaries of particular statements can be displayed in-line as shown in Figure 3.
After clicking on the concept names, statements related to it are displayed (similarly to the previously shown answer display). Additionally, we provide a box with further information on the concept (Figure 4). Currently we automatically fetch an abstract, image and biomedical classification details from the respective Wikipedia entries (looking up the synonyms of the concept in question). However, other online sources, such as the KEGG service (cf. http: //www.genome.jp/kegg/kegg2.html) can be easily incorporated. Full-Text Search Besides the knowledge-based search, CORAAL supports also classical full-text search, which results in article paragraphs containing the queried terms. Links to the detailed view of the respective publication are provided, as well as citation contexts for the queried terms (i.e., publications referenced from the adjacent text). The full-text search results can be filtered, for instance by authors or citation contexts. One can also filter publications according to the concepts they contain, but also according to the associated indirect (inferred) general topics or specific instances. Besides publications, one can search for titles and authors. The respective results can be filtered, too. Figure 5 shows list of authors corresponding to the “Lin” name filtered only to those who have written an article concerned with the “gene amplification abnormality” topic. This feature of CORAAL can be used when looking for candidate experts on certain topics of interest. 4
With CORAAL, we started to implement the requirements of the use case outlined in Section 1.1. We have processed non-trivial amount of publication data purely automatically, extracted and processed the knowledge hidden in the articles, and exposed it conveniently to the users. The next and rather straightforward step is to integrate the publication knowledge with knowledge in patient records. For that, an appropriate parser for the patient records is essentially enough, as the EUREEKA engine employed in CORAAL is agnostic w.r.t. the origin of the input data. In the long-term perspective, one major task lies ahead of us, though. To make CORAAL really useful in practice, we still need to reduce the noise level in the exposed knowledge. Perhaps the easiest way is to utilise the wisdom of the crowds by supporting secure, error-prone and unobtrusive user involvement in the knowledge base updates (namely by (in)validation of existing statements, introduction of new statements and submission of new rules refining the domain semantics).
Acknowledgments This work has been supported by the ‘L´ıon’, ‘L´ıon II’ projects funded by Science Foundation Ireland under Grants No. SFI/02/CE1/ I131, SFI/08/CE/I1380, respectively. We acknowledge much appreciated help from Ioana Hulpus, who developed the initial user interface for CORAAL. Very special thanks goes to the people who have actively participated in the continuous prototype evaluation and testing, namely to (in an alphabetical order): Doug Foxvog, Peter Gr´ell, MD, Miloˇs Hol´anek, MD, Matthias Samwald, Holger Stenzhorn and Jiˇr´ı Vyskoˇcil, MD.