.. -*- mode: rst -*-
.. include:: ../definitions.rst

========================================================
?. Natural Language Processing and Information Retrieval
========================================================

-------------------------
?. Information Extraction
-------------------------

For many NLP applications, the priority is to extract meaning from
written text. This must be done *robustly*; processing should not fail
when unexpected or ill-formed input is received.  Today, robust
semantic interpretation is easiest when shallow processing methods are
used.  As we saw in the previous section, these methods severely limit
(or even omit) syntactic analysis and restrict the complexity of the
target semantic representations.

One well-known instance of shallow semantics is Information Extraction
(IE). In contrast to full text understanding, IE systems only try to
recognize a limited number of pre-specified semantic topics and use a
constrained semantic representation.  The initial phase of information
extraction involves the detection of named entities, expressions that
denote locations, people, companies, times, and monetary amounts.  (In
fact, these named entities are nothing other than the low-level
categories of the semantic grammars we saw in the previous section,
but under a new guise.)  To illustrate, the following text contains
several expressions which have been tagged as entities of types
``time``, ``company`` and ``location``:

  | The incident occurred <time>around 5.30pm</time> when a man walked into
  | <company>William Hill bookkeepers</company> on <location>Ware Road</location>,
  | and threatened a member of staff with a small black handgun. 

The final output of an information extraction system, usually referred
to as a template, consists of fixed number of slots that must be
filled. More semantically, we can think of the template as denoting an
event in which the participating entities play specific roles.  Here
is an example corresponding to the above text:

  | Crime event
  |   Time: around 5.30pm
  |   Location: William Hill bookkeepers on Ware Road
  |   Suspect: man
  |   Weapon: small black handgun

Early work in IE demonstrated that hand-built rules could achieve good
accuracy in the tasks defined by the DARPA Message Understanding
Conferences (MUC).  In general, highest accuracy has been achieved in
identifying so-called named entities, such as ``William Hill bookkeepers``.
However, it is harder to identify which slots the entities should
fill; and harder still to accurately identify which event the entities
are participating in.  In recent years attention has shifted to making
IE systems more portable to new domains by using some form of machine
learning.  These systems work moderately well on marked-up text
(e.g. HTML-formatted text), but their performance on free text still
needs improvement before they can be adopted more widely.

IE systems commonly use external knowledge sources in addition to the
input text.  For example, named entity recognizers often use name or
location gazetteers.  In the bioinformatics domain
it is common to draw on external knowledge bases for standard names of
genes and proteins.  A general purpose lexical ontology called WordNet
is often used in IE systems.  Sometimes formal domain ontologies are
used as a knowledge source for IE, permitting systems to perform
simple taxonomic inferences and resolve ambiguities in the
classification of slot-fillers.

For example, an IE system for extracting basketball statistics from
news reports on the web might have to interpret the sentence *Quincy
played three games last weekend*, where the ambiguous subject *Quincy*
might be either a player or a team.  An ontology in which *Quincy* is
identified as the name of a college basketball team would permit the
ambiguity to be resolved.  In other cases the template disambiguates
the slot filler.  Thus, in the basketball domain, *Washington* might
refer to one of many possible players, teams or locations.  If we know
that the verb *won* requires an entity whose role in the taxonomy is a
team, then we can resolve the ambiguity in the sentence *Washington won*.


Message Understanding
---------------------

A message understanding system
will extract salient chunks of text from a news story and populate a
database.

.. figure:: ../images/chunk-muc.png
   :scale: 60

   The Message Understanding Process (from Abney 1996)

Consider the units that have been selected in this process:
a name (``Garcia Alvarado``), a verb cluster (``was killed``),
a locative prepositional phrase (``on his vehicle``).  These
are examples of ``NP``, ``VP`` and ``PP`` chunks.


MUC-7 Data?


Named Entity Recognition
------------------------


Toponym Resolution
------------------


Recognizing Temporal Expressions
--------------------------------


Semantic Role Labeling
----------------------



---------------------------------
Biomedical Information Extraction
---------------------------------



----------------------
Document Summarization
----------------------


------------
Other Topics
------------

text retrieval

text classification

text segmentation

topic detection

web crawling

saving extracted material into a database


.. include:: footer.txt