Databases Reference
In-Depth Information
Named Entity
Recognition
Training
Keyphrase
Extraction
Relation
Extraction
Prediction
URI Lookup
Machine Learning
Tools
Fig. 6.
FOX Architecture
Unsupervised Extraction Example: The FOX Framework
Several frameworks have been developed to implement the functionality above for the
Data Web including OpenCalais
3
and Alchemy
4
. Yet, these tools rely mostly on one
approach to perform the di
erent tasks at hand. In this section, we present the FOX
(Federated knOwledge eXtraction) framework
5
, which makes use of the diversity of
the algorithms available for NER, KE and RE to generate high-quality RDF.
The architecture of FOX consists of
three main layers
as shown in Figure 6. The
machine learning
layer implements interfaces for accommodating ensemble learning
techniques such as simple veto algorithms but also neural networks. It consists of
two
main modules
.The
training module
allows to load training data so as to enable FOX
to learn the best combination of tools and categories for achieving superior recall and
precision on the input training data. Depending on the training algorithm used, the user
can choose to tune the system for either precision or recall. When using neural networks
for example, the user can decide to apply a higher threshold for the output neurons,
thus improving the precision but potentially limiting the recall. The
prediction module
allows to run FOX by loading the result of a training session and processing the input
data according to the tool-category combination learned during the training phase. Note
that the same learning approach can by applied to NER, KE, RE and URI lookup as
they call all be modelled as classification tasks.
ff
