Biology Reference
In-Depth Information
to incorporate lexical level features such as parts of speech or orthographic
labels (if greater rule generality is required and resources exist). The limita-
tion of the language is in the modeling of long-distance syntactic relations,
but in practice, we have not found this to be a problem in news reports.
The English SRL rule topic for biohazard events incorporates approximately
110 rules for entity detection, 12 major word lists containing 870 terms, and
more than 2800 template rules for detecting direct signals (e.g., international
travel, zoonosis, category A agents, novel diseases, and malformed blood
products), as well as indirect signals (distal indicators, e.g., infrastructure
strain in social services). Template rules are also used to detect temporal
aspects of events such as the historical or hypothetical. One important point
to note is that much of the biohazard rule topic, such as lists of countries or
victim expressions, can be reused in a modular fashion for the detection of
other public health events planned for future development in BioCaster.
After extracting the basic event frame, objects still require disambiguation. It
is often the case that country names are not given explicitly within the article.
Instead, this information can be inferred at times from role titles or organiza-
tion names. For example, “Victoria chief health officer” could potentially imply
countries such as Canada, Australia, or the Seychelles. Country, province, dis-
ease, and agent names are disambiguated statistically, and then geopositioned
to a latitude and longitude using lookup in the BCO. The final event frame is
then stored in the knowledge base for exploitation tools to access.
15.3.4 Data Dissemination
BioCaster provides output in a variety of forms. The open access public por-
tal provides users with 30 days of access to the Global Health Monitor, a
news collection site with biogeographic mapping capability based on Google
Maps
4
(Figure 15.2). News articles can be searched by time, text type, syn-
drome, or disease, and further sources of information can be found through
dynamically generated links to existing biomedical ontologies and academic
database. News headlines are ranked on the site according to an algorithm
that calculates the novelty of the disease and location based on the previ-
ous 30 day's data, with special provision for potentially fast-spreading events
and highly stressed health systems.
A login-restricted alerting interface is used by a small test community at
the Ministry of Health in Japan and the Health Protection Agency in the
United Kingdom. Here, users can define targeted rules that let them receive
e-mail alerts whenever news comes into the system on topics of interest. For
example, a user could define a rule asking to be alerted on the topic of novel
influenza in Europe involving a case of international travel. More advanced
news search and analytics are also incorporated into the login site allow-
ing users access to aggregated data on events. It is expected that this ser-
vice will soon be opened up to more users within the global public health
community.
