A Flexible System for Text Analysis with Semantic Networks

1. Introduction
The explosive growth of digital and digitized text creates
opportunities for scholars and students to conduct new
analyses and develop unique insights about our written culture
and heritage. To effectively use large collections of textual data,
scholars and students need flexible, easy to use tools that
provide powerful analysis and visualization.
1.1 Goals
The Automated Learning Group is developing interactive tools
for mining large, complex semantic networks, which were
automatically extracted from a text corpus. The corpus could
be a collection of documents or a stream of messages. The semantic network represents the concepts in the collection as
entities and relations. The visual environment allows the user
to query the semantic network to retrieve portions of the graph.
The display allows the user to view and navigate these networks
to discover patterns in the collection.
1.2 Technical Approach
Our approach to the visual analysis of text documents
implements extraction and visualization of important
entities and the relations among them. This is accomplished in
the following high-level steps:
1. Perform lexical analysis of textual document set
2. Extract key entities and relations from the text
3. Compose entity-relation triples into semantic network model
summarizing key information in document corpus
4. “Publish” the semantic network in a repository
5. Visualize semantic network model as graph
6. Search and prune visual graph through appropriately
structured queries against semantic network model.
2 Text Analysis
Textual processing is accomplished by passing documents
through a series of syntactic, semantic, and functional
analysis tools. These tools primarily consist of D2K 1and T2K,
our own analytic software, working in concert with GATE 2and
MontyLingua 3. These tools are seamlessly connected together
into the D2K visual programming environment
The GATE toolkit is utilized for standard syntactic processes
including tokenization, sentence splitting, and part-of-speech
tagging. Syntactically annotated documents are passed to
GATE’s Named Entity (NE) tagger. The GATE NE tagger
identifies proper nouns in the text that belong to relevant
categories such as “Person”, “Organization”, and “Location.”
NE tagging passes through three modules that perform different
kinds of co-reference2. The final output of the NE tagging and
co-referencing processes are a set of annotations that identify
references to “Person”, “Organization”, and “Location” entities.
After NE extraction, we identify key relations in the text using
a tool within MontyLingua. This Jister tool extracts sentence
structures called “jists,” which carry thematic-role information
such as verb, subject, and object. This is similar to semantic
role labeling 2, but less formally structured. An example jist:
Tiger Woods wrapped up the tournament at noon.
(Verb: ‘wrap up’, Subj: ‘Tiger Woods’,
Obj1: ‘tournament’,Obj2: ‘at noon’)
The next step involves the normalization of references in the
MontyLingua jists with entities identified in the analysis. For
example, “Tiger Woods” may have been identified as a
entity:PERSON, and “tournament” referenced back to “Masters
Golf Tournament” appearing earlier in the text. Tagging the
verb “wrap up” as an entity:ACTION, and “at noon” as a TIME
relation to the verb, we can generate triples as follows: <Tiger
Woods> <is-a> <entity:PERSON> <Masters Golf
Tournament> <is-a> <entity:EVENT> <wrap
up> <is-a> <entity:ACTION> <Tiger Woods>
<actor> <wrap up>
Figure 1 shows the D2K toolkit with a view of the text
processing itinerary. The itinerary is a dataflow graph: the nodes
in the figure are D2K modules (major processing blocks),
connected by edges representing the flow of data during
execution. The itinerary can exploit both data parallelism
(multiple documents at the same time) and task parallelism
(different modules in parallel). The D2K itinerary can be run
on a desktop or scaled up to multiple computers or High
Performance Computer systems.Figure 1 shows the D2K toolkit
with a view of the text processing itinerary. The itinerary is a
dataflow graph: the nodes in the figure are D2K modules (major
processing blocks), connected by edges representing the flow
of data during execution. The itinerary can exploit both data
parallelism (multiple documents at the same time) and task
parallelism (different modules in parallel). The D2K itinerary
can be run on a desktop or scaled up to multiple computers or
High Performance Computer systems.
3 Semantic Network Storage and
Retrieval
The triples generated from the semantic extraction process
are stored in an RDF 5metadata store. We use Kowari,
developed by Tucana Technologies 6. Additional triples are
generated to represent metadata in conformance with a common
vocabulary, and user annotations can be included as well.
Queries are coded in an SQL-like query language called iTQL.
The result is a set of triples, which represents a semantic graph
6.
This architecture demonstrates a key design principle for robust
Cyberinfrastructure. The analysis is decoupled from the
visualization, so that a large scale analysis can asynchronously
update the triples as new results are computed, while interactive
tools will automatically pick up the new data by refreshing the
query. The triples generated from the semantic extraction
process can be combined with many other similar relation triples
from many sources, and additional triples are generated to
represent a common vocabulary. 4 Visual Investigation of Semantic
Networks
Using the visual environment, investigators can perform
searches over the semantic networks extracted from a
text corpus. The familiar web browser paradigm was employed
in the user interface design. The user interface allows one to
construct more complex queries by incorporating multiple rules
and filters. Each user query is converted into iTQL and executed
by the Kowari server6 . The query history is available as a
pull-down menu, just as query histories are in a web browser.
Investigators can directly observe semantic relationships
between entities in an interactive link-node graph visualization.
(Figure 2)
Relations between entities in the resulting semantic network
graph are displayed as a link-node graph visualized using
Prefuse 7, and also as a hierarchical tree of entities conforming
to the common vocabulary. The subject and object entities in
the relations are displayed as nodes in the graph visualization,
predicates are displayed as links. This simplification of the
more complex semantic network, stored in Kowari, provides a
compact and usable abstraction of the important relations
extracted from the text streams.
The Entity pane at the upper left displays lists of entities. The
Relations pane at the lower left displays a list of relations, with
additional options to highlight synonyms, antonyms, hyponyms
or hypernyms based on WordNet. Selecting an entity in the left
pane also highlights the corresponding node or edge in the
visualization.
Every entity in the graph maintains a link back to the original
text document from which it was extracted. By right-clicking
on a node, and selecting View Source Documents, the text of
the original document will be retrieved from the repository and
displayed.
5 Collections
This tool can be applied for many different types of text,
across one or many collections. In addition, it can analyze
evolving collections of text, such as documents from one or
more RSS feed.
This tool has been used as part of the Nora project, a
multi-institution collaboration to produce software for
discovering, visualizing, and exploring significant patterns
across large collections of full-text humanities resources in
existing digital libraries.8
For example, we are experimenting with the digitized text of
the novel Uncle Tom’s Cabin, available as part of the Early
American Fiction Collection from the University of Virginia9.
The system performs feature extraction in order to determine
shared characteristics of the selected documents, such as
chapters of the novel. The resulting arc-node graph can be
viewed and navigated to discover patterns and relations within
and among the texts. Figure 2 illustrates an example analysis
of text from nineteenth century novels, showing the use of
concepts related to “Mother” and “Man”.10
6 Conclusion
This paper described interactive tools for mining large,
complex semantic networks automatically extracted from
a text corpus. These approaches can be applied to news feeds,
technical literature, or literary collections.
We believe this is a promising approach, although we are only
beginning to develop these tools for use by humanists, who
will be the ultimate judges of the utility and validity of this
approach. The general purpose semantic analysis is widely
applicable to many types of text, though it is difficult to predict
the impact of such analysis on our understanding of texts.
The query interface and graphical displays are still under
development. The entity-relationship graphs may be quite
complicated, so we must find new visual methods and
metaphors to enable scholars and students to understand the
information in the graphs, and to use them formulate
hypotheses, and answer questions.
7 Acknowledgements
The Nora project is funded in part by the Mellon
Foundation. This work was funded in part by the National
Center for Advanced Secure Systems Research (NCASSR) at
the University of Illinois at Urbana-Champaign (UIUC), a
multi-institutional cybersecurity research team. NCASSR is
led by the National Center for Supercomputing Applications
(NCSA) and supported by funding from the Office of Naval
Research (ONR). Substantial portions of the code were
implemented by David Clutter and Fang Guo. Thanks to Patricia
S. Taylor. Figure 1: D2K Itinerary showing text processing.
Figure 2: Visual interface showing result of a query.
1. <http://alg.ncsa.uiuc.edu>
2. <http://gate.ac.uk>
3. <http://web.media.mit.edu/~hugo/montyli
ngua>
2. <http://gate.ac.uk>
2. <http://gate.ac.uk>
5. <http://www.w3.org/RDF>
6. <http://kowari.org>
6. <http://kowari.org>
6. <http://kowari.org>
7. <http://prefuse.org>
8. <http://www.noraproject.org>
9. <http://etext.virginia.edu/eaf/overview
.html>
10. Tom Horton, Kristen Taylor, Bei Yu and Xin Xiang, "Quite Right,
Dear and Interesting”: Seeking the Sentimental in Nineteenth
Century American Fiction"Digital Humanities 2006, pp. 81-82