Text analysis and Semantic Search with GATE

University of Sheffield, NLP
Text Analysis with GATE
Diana Maynard
University of Sheffield
Search Solutions Tutorial 2014
London, UK
© The University of Sheffield, 2014
This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike Licence

Outline of the Tutorial
• Introduction to NLP and Information Extraction
• Introduction to GATE
• Social media analysis
• Text Analysis for Semantic Search
• Semantic Search
• Semantic Annotation
• GATE MIMIR
• Example applications
• ExoPatent, TNA Web Archive, BBC News, Envilod,
Prospector

1. NLP and Information Extraction

Oddly enough, people have successfully combined
information and toast...

The weather-forecasting toaster
● This weather-forecasting toaster, connected to a phone point,
was designed in 2001 by a PhD student
● It accessed the MetOffice website via a modem inside the
toaster and translated the information into a 1, 2 or 3 for rain,
cloud or sun
● The relevant symbol was then branded into the toast in the last
few seconds of toasting

With tools such as these, why do we need text
mining?
● It turns out that toast isn't actually a very good medium
for finding information

7
Information overload or filter failure?
● We all know that there's
masses of information
available online these days,
and it's constantly growing
● You often hear people talk
about “information overload”
● But the real problem is not
the amount of information,
but our inability to filter it
correctly
● Clay Shirky has an excellent
talk on this topic
http://bit.ly/oWJTNZ

It is difficult to access unstructured information
efficiently
Information extraction tools can help you:
● Save time and money on management of text and data from
multiple sources
● Find hidden links scattered across huge volumes of diverse
information
● Integrate structured data from variety of sources
● Interlink text and data
● Collect information and extract new facts

What is information extraction?
● Information Extraction (IE) is the automatic discovery of new,
previously unknown information, by automatically extracting
information from different textual resources.
● A key element is the linking together of the extracted
information together to form new facts or new hypotheses to
be explored further
● In IE, the goal is to discover previously unknown information,
i.e. something that no one yet knows

10
IE is not Data Mining
Data mining is about using analytical techniques to find
interesting patterns from large structured databases
Examples:
• using consumer purchasing
patterns to predict which products
to place close together on shelves
in supermarkets
• analysing spending patterns on
credit cards to detect fraudulent
card use.

11
IE is not Web Search
● IE is also different from traditional web search or IR.
● In search, the user is typically looking for something that is
already known and has been written by someone else.
● The problem lies in sifting through all the material that currently
isn't relevant to your needs, in order to find the information that
is.
● The solution often lies in better ways to ask the right question
● You can't ask Google to tell you about
– all the speeches made by Tony Blair about foot and mouth
disease while he was Prime Minister
– all the documents in which a politician born in Sheffield is
quoted as saying something about hospitals.

More about how we can do this will be revealed later...
GATE Mímir:
Answering Questions
Google Can't

Information Extraction Basics
● Entity recognition
is required for
● Relation extraction
which is required for
● Event recognition
which is required for
● Summarisation tasks

What is Entity Recognition?
● Entity Recognition is about recogising and classifying key Named
Entities and terms in the text
● A Named Entity is a Person, Location, Organisation, Date etc.
● A term is a key concept or phrase that is representative of the text
Mitt Romney, the favorite to win the Republican nomination for president in 2012
Person Term Date
● Entities and terms may be described in different ways but refer to
the same thing. We call this co-reference.
co-reference
The GOP tweeted that they had knocked on 75,000 doors in Ohio the day prior.
Organisation
Location

What is Event Recognition?
● An event is an action or situation relevant to the domain
expressed by some relation between entities or terms.
● It is always grounded in time, e.g. the performance of a
band, an election, the death of a person
Relation Relation
Mitt Romney, the favorite to win the Republican nomination for president in 2012
Person Event Date

Why are Entities and Events Useful?
● They can help answer the “Big 5” journalism questions
(who, what, when, where, why)
● They can be used to categorise the texts in different ways
– look at all texts about Obama.
● They can be used as targets for opinion mining
– find out what people think about President Obama
● When linked to an ontology and/or combined with other
information, they can be used for reasoning about things
not explicit in the text
– seeing how opinions about different American
presidents have changed over the years

17
NLP components for text mining
● A text mining system is usually built up from a number of
different NLP components
● First, you need some Information Extraction tools to do the
donkey work of getting all the relevant pieces of information
and facts.
● Then you need some tools to apply the reasoning to the facts,
e.g. opinion mining, information aggregation, semantic
technologies, dynamics analysis and so on.
● GATE is an example of a tool for text mining which allows you
to combine all the necessary NLP components

Approaches to Information Extraction
Knowledge Engineering
 rule based
 developed by
experienced language
engineers
 make use of human
intuition
 easier to understand
results
 development could be
very time consuming
 some changes may be
hard to accommodate
Learning Systems
 use statistics or other
machine learning
 developers do not need
LE expertise
 requires large amounts of
annotated training data
 some changes may
require re-annotation of
the entire training corpus

2. GATE

What is GATE?
GATE is an NLP toolkit developed at the University of Sheffield
over the last 20 years
It includes:
• components for language processing, e.g. parsers, machine
learning tools, stemmers, IR tools, IE components for various
languages...
• tools for visualising and manipulating text, annotations,
ontologies, parse trees, etc.
• various information extraction tools
• evaluation and benchmarking tools

GATE components
● Language Resources (LRs), e.g. lexicons, corpora,
ontologies
● Processing Resources (PRs), e.g. parsers, generators,
taggers
● Visual Resources (VRs), i.e. visualisation and editing
components
● Algorithms are separated from the data, which means:
– the two can be developed independently by users with
different expertise.
– alternative resources of one type can be used without
affecting the other, e.g. a different visual resource can be
used with the same language resource

ANNIE
• ANNIE is GATE's rule-based IE system
• It uses the language engineering approach (though we also
have tools in GATE for ML)
• Distributed as part of GATE
• Uses a finite-state pattern-action rule language, JAPE
• ANNIE contains a reusable and easily extendable set of
components:
– generic preprocessing components for tokenisation,
sentence splitting etc
– components for performing NE on general open domain text

What's in ANNIE?
• The ANNIE application contains a set of core PRs:
• Tokeniser
• Sentence Splitter
• POS tagger
• Gazetteers
• Named entity tagger (JAPE transducer)
• Orthomatcher (orthographic coreference)
• There are also other PRs available in the ANNIE plugin, which
are not used in the default application, but can be added if
necessary
• NP and VP chunker, morphological analysis

ANNIE Modules

Let's look at the PRs
• Each PR in the ANNIE pipeline creates some new
annotations, or modifies existing ones
• Document Reset → removes annotations
• Tokeniser → Token annotations
• Gazetteer → Lookup annotations
• Sentence Splitter → Sentence, Split annotations
• POS tagger → adds category features to Token
annotations
• NE transducer → Date, Person, Location, Organisation,
Money, Percent annotations
• Orthomatcher → adds match features to NE annotations

Tokeniser
 Tokenisation based on Unicode classes
 Declarative token specification language
 Produces Token and SpaceToken annotations with
features orthography and kind
 Length and string features are also produced
 Rule for a lowercase word with initial uppercase
letter:
"UPPERCASE_LETTER" LOWERCASE_LETTER"* >
Token; orthography=upperInitial; kind=word

27
Document with Tokens

ANNIE English Tokeniser
 The English Tokeniser is a slightly enhanced version of the
Unicode tokeniser
 It comprises an additional JAPE transducer which adapts
the generic tokeniser output for the POS tagger
requirements
 It converts constructs involving apostrophes into more
sensible combinations
 don’t → do + n't
 you've → you + 've

Sentence Splitter
• The default splitter finds sentences based on Tokens
• Creates Sentence annotations and Split annotations on the
sentence delimiters
• Uses a gazetteer of abbreviations etc. and a set of JAPE
grammars which find sentence delimiters and then annotate
sentences and splits
• There are a few sentence splitter variants available which work
in slightly different ways

30
Document with Sentences

POS tagger
 ANNIE POS tagger is a Java implementation of Brill's
transformation based tagger
 Previously known as Hepple Tagger
 Trained on WSJ, uses Penn Treebank tagset
 Default ruleset and lexicon can be modified manually (with a
little deciphering)
 Adds category feature to Token annotations
 Requires Tokeniser and Sentence Splitter to be run first

Morphological analyser
 Not an integral part of ANNIE, but can be found in the
Tools plugin as an “added extra”
 Flex based rules: can be modified by the user
(instructions in the User Guide)
 Generates “root” feature on Token annotations
 Requires Tokeniser to be run first
 Requires POS tagger to be run first if the
considerPOSTag parameter is set to true

Gazetteers
● Gazetteers are plain text files containing lists of names (e.g
rivers, cities, people, …)
● The lists are compiled into Finite State Machines
● Each gazetteer has an index file listing all the lists, plus
features of each list (majorType, minorType and language)
● Lists can be modified either internally using the Gazetteer
Editor, or externally in your favourite editor
● Gazetteers generate Lookup annotations with relevant features
corresponding to the list matched
● You can also add extra features and values to individual lists
● Lookup annotations are used primarily by the NE transducer
● The ANNIE gazetteer has about 60,000 entries arranged in 80
lists

Gazetteer editor
definition file
entries entries for selected list

Named Entity Recognition
• Gazetteers can be used to find terms that suggest entities
• However, the entries can often be ambiguous
– “May Jones” vs “May 2010” vs “May I be excused?”
– “Mr Parkinson” vs “Parkinson's Disease”
– “General Motors” vs. “General Smith”
• Handcrafted grammars are used to define patterns over the
Lookups and other annotations
• These patterns can help disambiguate, and they can combine
different annotations, e.g. Dates can be comprised of day +
number + month

Named Entity Grammars
• Hand-coded rules written in JAPE applied to annotations to
identify NEs
• Phases run sequentially and constitute a cascade of FSTs over
annotations
• Annotations from format analysis, tokeniser. splitter, POS tagger,
morphological analysis, gazetteer etc.
• Because phases are sequential, annotations can be built up over
a period of phases, as new information is gleaned
• Standard named entities: persons, locations, organisations, dates,
addresses, money
• Basic NE grammars can be adapted for new applications, domains
and languages

Example of a JAPE pattern-action rule
Rule: PersonName
(
{Lookup.majorType == firstname}
{Token.category == NNP}
):tag
-->
:tag.Person = {kind = fullName}
Look for an entry in a
gazetter list of first
names
followed by a proper noun
create an annotation of type “Person” give the annotation the feature kind
and value “fullName”

JAPE rules are usually a bit more complex
● They can make use of any annotations already created, e.g.
– strings of text
– POS categories
– gazetteer lookup
– root forms of words
– document metadata, e.g. HTML tags
– annotations “in context”
● They can use regular expression operators, including negative
constraints, “contains”, “within” etc.
● And the RHS of a rule can contain any Java code you like

Document with NEs

Using co-reference
 Different expressions may refer to the same entity
 Orthographic co-reference module (orthomatcher)
matches proper names and their variants in a
document
 [Mr Smith] and [John Smith] will be matched as the
same person
 [International Business Machines Ltd.] will match
[IBM]

Orthomatcher PR
• Performs co-reference resolution based on orthographical
information of entities
• Produces a list of annotation IDs that form a co-reference “chain”
• List of such lists stored as a document feature named
“MatchesAnnots”
• Improves results by assigning entity type to previously unclassified
names, based on relations with classified entities
• May not reclassify already classified entities
• Classification of unknown entities very useful for surnames which
match a full name, or abbreviations,
e.g. “Bonfield” <Unknown> will match “Sir Peter Bonfield” <Person>
• A pronominal PR is also available

Coreference

Other NLP Toolkits
• UIMA
• OpenCalais
• Lingpipe
• OpenNLP
• Stanford Tools
All integrated into GATE as plugins

3. Analysing Social Media

Analysing language in social media is hard
● Grundman:politics makes #climatechange scientific issue,people
don’t like knowitall rational voice tellin em wat 2do
● @adambation Try reading this article , it looks like it would be
really helpful and not obvious at all. http://t.co/mo3vODoX
● Want to solve the problem of #ClimateChange? Just #vote for a
#politician! Poof! Problem gone! #sarcasm #TVP #99%
● Human Caused #ClimateChange is a Monumental Scam!
http://www.youtube.com/watch?v=LiX792kNQeE … F**k yes!!
Lying to us like MOFO's Tax The Air We Breath! F**k Them!

Challenges for NLP
● Noisy language: unusual punctuation, capitalisation, spelling, use
of slang, sarcasm etc.
● Terse nature of microposts such as tweets
● Use of hashtags, @mentions etc causes problems for
tokenisation #thisistricky
● Lack of context gives rise to ambiguities
● NER performs poorly on microposts, mainly because of linguistic
pre-processing failure
– Performance of standard IE tools decreases from ~90% to
~40% when run on tweets rather than news articles

Persons in news articles

Persons in tweets

Lack of context causes ambiguity
Branching out from Lincoln park after dark ... Hello Russian
Navy, it's like the same thing but with glitter!
??

Getting the NEs right is crucial
Branching out from Lincoln park after dark ... Hello Russian
Navy, it's like the same thing but with glitter!

How do we deal with this kind of text?
● Typical NLP pipeline means that degraded performance has a
knock-on effect along the chain
● Short sentences confuse language identification tools
● Linguistic processing tools have to be adapted to the domain
● Retraining individual components on large volumes of data
● Adaptation of techniques from e.g. SMS analysis
● Development of new Twitter-specific tools (e.g. GATE's TwitIE)
● But....lack of standards, easily accessible data, common
evaluation etc. are holding back development

TwitIE to the rescue

4. Text Analytics for Semantic Search

Text Search isn't Enough
“I like the Internet. Really, I do. Any time I
need a piece of shareware or I want to find
out the weather in Bogota... I'm the first guy
to get the modem humming. But as a
source of information, it sucks. You got a
billion pieces of data, struggling to be heard
and seen and downloaded, and anything I
want to know seems to get trampled
underfoot in the crowd."
Michael Marshall, The Straw Men. HarperCollins
Publishers, 2002.

Semantic Queries in Google

Searching for Things, Not Strings
• 500 million entities that Google
“knows” about
• Used to provide more accurate
search results
• Summaries of information about
the entity being searched
http://googleblog.blogspot.it/2012/05/introducing-knowledge-graph-things-not.html

Facebook Graph Search

Semantic Enrichment
● Textual mentions aren't actually that useful in isolation
– knowing that something is a “Person" isn't very helpful
– knowing which Person the mention refers to can be very useful
● Disambiguating mentions against an ontology provides extra context
● This is where semantic enrichment comes in
● The end product is a set of textual mentions linked to an ontology,
otherwise known as semantic annotations
● Annotations on their own can be useful but they can also
– be used to generate corpus level statistIcs
– be used for further ontology population
– form the basis of summaries
– be indexed to provide semantic search

Automatic Semantic Enrichment
• Use Text Mining, e.g.
• Information Extraction – recognise names of people,
organisations, locations, dates, references, etc.
• Term recognition – identify domain-specific terms
• Automatically extend article metadata to improve search quality
• Example: using a customised GATE text mining pipeline to enrich
metadata in the Envia environmental science repository
http://www.bl.uk/reshelp/experthelp/science/eventsandprojects/en
viatbl/index.html

Semantic Enrichment in Envia
60

Mining medical records
● Medical records contain a large amount of unstructured text
– letters between hopitals and GPs
– discharge summaries
● These documents might contain information not recorded
elsewhere
– it turns out doctors don't like forms!
– often information-specific fields are ignored, with everything
put in the free text area

Medical Records at SLAM
● NIHR Biomedical Research Centre at the South London and
Maudsley Hospital are using text mining in a number of their
studies
● They have developed applications to extract:
– the results of mental state tests, and the date the test was
administered
– education level (high school, university, etc.)
– smoking status
– medication history
● They have even had promising results predicting suicides!

Cancer Research
● Genome Wide Association Studies (GWAS) aim to investigate
genetic variants across the whole genome
– With enough cases and controls, this allows them to state
that a given SNP (Single Nucleotide Polymorphism) is
related to a given disease.
– A single study can be very expensive in both time and
money to collect the required samples.
● Can we reduce the costs by analysing published articles to
gnerate prior probabilities for each SNP?

Can Semantic Annotation Cure Cancer?
● In conjunction with IARC (International Agency for Research on
Cancer, part of the WHO) we developed a text analysis
approach to mine PubMed
● We showed retrospectively that our approach would have saved
over a year's worth of work and more than 1.5 million Euros
● We completed a new study which found a new cause for oral
cancer
– Oral cancer is rare enough that traditional methods would
have failed to find enough cases to make the study plausible

Government Web Archive
● We developed a semantic annotation application to process
every crawled page in the archive.
● Entities annotated included; people, companies, locations,
government departments, ministerial positions, social
documents, dates, money....
● Where possible, annotations were linked to an ontology which
– was based on DBpedia
– was extended with UK government-specific concepts
– included the modelling of the evolution of government
● Annotations were indexed to allow for complex semantic
querying of the collection
● Demo coming later.....

5.1 Semantic Annotation

Why ontologies for semantic search?
 Semantic annotation: rather than just annotating the word “Cambridge” as a
location, link it to an ontology instance
 Differentiate between Cambridge, UK and Cambridge, Mass.
 Semantic search via reasoning
 So we can infer that this document mentions a city in Europe.
 Ontologies tell us that this particular Cambridge is part of the country called
the UK, which is part of the continent Europe.
 Knowledge source
 If I want to annotate strikes in baseball reports, the ontology will tell me that
a strike involves a batter who is a person
 In the text “BA went on strike”, using the knowledge that BA is a company
and not a person, the IE system can conclude that this is not the kind of
strike it is interested in

More semantic search examples
Q: {ScalarValue}{MeasurementUnit} ->
A: “12 cm”, “190 g”, “two hours”
---
Q: {Reference} ->
A: JP-A-60-180889
A: Kalderon et al. (1984) Cell 39:499-509
68

Example Semantic Search Architecture
69

What is Semantic Annotation?
Annotation:
The process of adding metadata to [parts of] a document.
Semantic Annotation:
Annotation process where [parts of] the annotation schema
(annotation types, annotation features) are ontological objects.

Semantic Annotation

What is an Ontology?
 Set of concepts (instances and
classes)
 Relationships between them (is-a,
part-of, located-in)
 Multiple inheritance
 Classes can have more than one
parent
 Instances can have more than one
class
 Ontologies are graphs, not trees

URIs, Labels, Comments
 The names of a classes or instance shown are URIs
(Universal Resource Identifier)
 http://dbpedia.org/page/Paris
 The linguistic lexicalisation is typically encoded in the label
property, as a string
 The comment property is often used
for documentation purposes, similarly
a string
 Comments and labels are
annotation properties

Datatype Properties
 Datatype properties link individuals to data values
 Datatype properties can be of type boolean, date, int,
 e.g. a person can have an age property
 Available datatypes taken from XMLSchema
http://dbpedia.org/page/Paris

Object Properties
 Object properties link instances together
 They describe relationships between instances, e.g. people work for
organisations
 Domain is the subject of the relation (the thing it applies to)
 Range is the object of the relation (the possible”values”)
Person work_for Organisation
Domain Property Range
Similar to domains, multiple alternative ranges can be specified by using a
class description of the owl:unionOf, but this raises the complexity of the
ontology and makes reasoning harder

DBpedia
● Machine readable knowledge on various entities and topics,
including:
– 410,000 places/locations,
– 310,000 persons
– 140,000 organisations
● For each entity we have:
– entity name variants (e.g. IBM, Int. Business Machines)
– a textual abstract
– reference(s) to corresponding Wikipedia page(s)
– entity-specific properties (e.g. latitude and longitude for
places)

Example from DBpedia
…
Links to GeoNames
And Freebase
Latitude & Longitude

GeoNames
● 2.8 million populated places
– 5.5 million alternate names
● Knowledge about NUTS country sub-divisions
– use for enrichment of recognised locations with the implied
higher-level country sub-divisions
● However, the sheer size of GeoNames creates a lot of ambiguity
during semantic enrichment
● We use it as an additional knowledge source, but not as a primary
source (DBpedia)

5.2 Semantic Annotation In GATE
(and other tools)

Information Extraction for the Semantic Web
● Traditional IE is based on a flat structure, e.g. recognising
Person, Location, Organisation, Date, Time etc.
● For the Semantic Web, we need information in a hierarchical
structure
● Idea is that we attach semantic metadata to the documents,
pointing to concepts in an ontology
● Information can be exported as an ontology annotated with
instances, or as text annotated with links to the ontology

Traditional NE Recognition
John lives in London . He works there for Polar Bear Design .
PERSON LOCATION ORGANISATION

Co-reference
same_as
PER LOC ORG

Relations
live_in
PER LOC ORG

Relations (2)
employee_of
PER LOC ORG

Relations (3)
based_in
PER LOC ORG

Richer NE Tagging
● Attachment of instances in
the text to concepts in the
domain ontology
● Disambiguation of
instances, e.g. Cambridge,
MA vs Cambridge, UK

Ontology-based IE
John lives in London. He works there for Polar Bear Design.

Ontology-based IE (2)
John lives in London. He works there for Polar Bear Design.

Typical Semantic Annotation pipeline
Analyse document structure
Linguistic Pre-processing
NE recognition
Ontology Lookup
Ontology-based IE
Populate ontology (optional)
Corpus
Export as RDF
RDF

OpenCalais
http://viewer.opencalais.com/
Paste text of http://www.membranes.com/
Not easily customised/extended
Domain-specific coverage varies

Zemanta
● Paste text from
www.membranes.com
● The main entity of
interest “Hydranautics”
is missed
● Common problem with
general purpose, open-domain
semantic
annotation tools
● Best results require
bespoke customisation
http://www.zemanta.com/demo/

Semantic Annotation in GATE
● Insert GATE screenshot

Automatic Semantic Annotation
● Locations (linked to DBpedia and GeoNames)
– Annotate the place name itself (e.g. Norwich) with the
corresponding DBpedia and GeoNames URIs
– Also use knowledge of the implied reference to the levels 1, 2,
and 3 sub-divisions from the Nomenclature of Territorial Units for
Statistics (NUTS).
– For Norwich, these are East of England (UKH – level 1), East
Anglia (UKH1 – level 2), and Norfolk (UKH13 – level 3).
– Similarly use knowledge to retrieve nearby places

“South Gloucestershire” Example

Semantic Annotation (2)
● Organisations (linked to DBpedia)
– Names of companies, government organisations, committees,
agencies, universities, and other organisations
● Dates
– Absolute (e.g. 31/03/2012) and relative (yesterday)
● Measurements and Percentages
– e.g. 8,596 km2 , 1 km, one fifth, 10%

The YODIE Service from GATE
http://demos.gate.ac.uk/trendminer/obie/
96
http://demos.gate.c.uk/trendminer/obie/

Evaluation: YODIE on TAC-KBP’2010
PER LOC ORG TOTAL
DB Spotlight 0.97 / 0.40 0.82 / 0.46 0.86 / 0.31 0.85 / 0.39
Zemanta 0.96 / 0.84 0.89 / 0.62 0.82 / 0.57 0.90 / 0.68
LODIE 0.81 / 0.82 0.73 / 0.76 0.56 / 0.59 0.71 / 0.74
Zemanta ∩ LODIE 1.00 / 0.74 0.95 / 0.45 0.97 / 0.42 0.97 / 0.54
Zemanta U LODIE 0.94 / 0.93 0.77 / 0.76 0.72 / 0.71 0.82 / 0.81
● Precision and Recall figures shown above
● Similar results on a set of metadata records and scientific papers
in environmental science (EnviLOD)
● hard when candidate ambiguity is high

Recognition of term variants in Decarbonet

ClimaTerm Web service
● http://services.gate.ac.uk/decarbonet/term-recognition/
● Input: a document
● Output: the text as an XML file annotated with term and URI
“Cars pollute by emitting carbon
dioxide.”
<Document>
<TermInstance="
http://www.eionet.europa.eu/gemet/concept/1148">cars
</Term> pollute by emitting
<Term Instance="http://reegle.info/glossary/1064">carbon
dioxide</Term>
</Document>

Term recognition demo

Semantic Search: An Overview

GATE Mímir
• can be used to index and search over text, annotations,
semantic metadata (concepts and instances)
• allows queries that arbitrarily mix full-text, structural, linguistic
and semantic annotations
• is open source

What can GATE Mímir do that Google can't?
Show me:
• all documents mentioning a temperature between 30 and 90
degrees F (expressed in any unit)
• all abstracts written in French on patent documents from the last
6 months which mention any form of the word “transistor” in the
English abstract
• the names of the patent inventors of those abstracts
• all documents mentioning steel industries in the UK, along with
their location

Search News Articles for
Politicians born in Sheffield
http://demos.gate.ac.uk/mimir/gpd/search/gus

Easily Create Your Own
Custom GATE Mímir Interfaces
http://demos.gate.ac.uk/pin/

MIMIR: Searching Text Mining Results
● Searching and managing text annotations, semantic
information, and full text documents in one search engine
● Queries over annotation graphs
● Regular expressions, Kleene operators
● Designed to be integrated as a web service in custom end-user
systems with bespoke interfaces
● Demos at http://services.gate.ac.uk/mimir/

Modelling Text Annotations and Additional
Knowledge Sources

MIMIR: the technical bit
● Uses a combination of an FTS engine (MG4J), a database for
indexing the semantic annotations (H2), and a knowledge repository
(OWLIM/Sesame)
● Use database to indexing annotation kinds, then the additional
linguistic data for words/annotations are stored in MG4J (POS,
morphological root, etc)
● Semantic query: against the database/ knowledge base for the
additional knowledge
● A query: decompose into the respective parts, query the appropriate
component, then merge the results

Clustering and Scalability
● All searches are local to a document, so we can use clusters of
semantic repositories:
● Faster indexing (less data in each repository)
● Faster searching (search space is broken into slices that are
searched in parallel – like Google).
● Joining results is trivial: union of result sets.
● Simple scalability: just add more nodes.
● Search speed stays almost constant while the data increases (each
individual repository has the same amount of data; there are just
more repositories)

Ranking of Results
● By default documents are returned in index order
● The following ranking algorithms are also available (others can be
implemented via plugins if required)
– Count
– Hit length
– TF.IDF
– BM25
• These algorithms treat annotations within the queries in the same way
as words, using index level frequencies etc.
• The ranking algorithm can be changed without re-indexing, allowing for
easy experimentation

Scaling Up
● We annotated 1.08 million web pages using a GATE language
analysis pipeline.
– Documents crawled using Heritrix10 , with total content size of
57 GiB or 6.6 billions plain text characters.
– The indexing server has 2 Intel Xeon 2.8GHz CPUs 11 GB of
RAM, and runs 64 bit Ubuntu Linux. Indexing process was 94
hours.
● We also indexed 150 million similar web pages, using two hundred
Amazon EC2 Large Instances running for a week to produce a
federated index
● Mimir runs on GateCloud.net, so easy to scale up

Search for string “Harriet Harman”

Search for string “Harriet Harman says”

Search with morphological variants:
Harriet Harman root:say

Replace strings with NEs: {Person} root:say

{Person} AND root:say – 11803 hits

{Person} [0..5] root: say – 5495 hits

Patent Annotation: Data Model
118

An Example Text
119

Hands-On with patent data
http://demos.gate.ac.uk/mimir/patents/search/index
Text. Matches plain text.
Example: nanomaterial
Linguistic variations of text
Example: (root:nanomaterial | root:nanoparticle )
Annotation. Matches semantic annotations.
Syntax: {Type feature1=value1 feature2=value2...}
Example: {Abstract lang="DE"}
Sequence Query. Sequence of other queries.
Syntax: Query1 [n..m] Query2...
Example: from {Measurement} [1..5] {Measurement}

Inclusion Queries
IN Query. Hits of one query only if inside another.
Syntax: Query1 IN Query2
Example: (root:nanomaterial | root:nanoparticle ) IN {Abstract}
Number of times these words are mentioned in patent abstracts (as
well as links to the actual documents)
OVER Query. Hits of a query, only if overlapping hits of another.
Syntax: Query1 OVER Query2
Example: {Abstract} OVER (root:nanomatrial | root:nanoparticle )
Finds all abstracts that contain nanomaterial(s) or nanoparticle(s)

Date restrictions
(
{Abstract lang="EN"} OVER
(root:nanomaterial | root:nanoparticle )
)
IN
{PatentDocument date > 20050000}
YYYYMMDD

Find references to literature or patents in
the prior art or background sections, which
contain nanomaterial/nanoparticle
({Reference type="Literature"}
|
{Reference type="Patent"}
) IN
({Section type="PriorArt"}
|
{Section type="BackgroundArt"}
)
OVER
(root:nanomaterial | root:nanoparticle)

Queries Using External Knowledge
{Measurement spec="1 to 100 volts“}
● Uses GNU Units (http://www.gnu.org/software/units/) to convert
measurements and normalise them to ISI units
{Measurement spec="1 to 100 kg m^2 / A s^3"}
● Example hits: 10 volts, 2V, +20 and -20 volts; ±10V; +/- 100V; +3.3
volts
{Measurement spec="1 to 100 m / s"}
● Example hits: 40 km/hr, 60m/min, 100cm/sec, 60 fps; 10 to 2000
cm/sec

Searching LOD with SPARQL
• SQL-like query language for RDF data
• Simple protocol for querying remote databases over HTTP
• Query types
• select: projections of variables and expressions
• construct: create triples (or graphs) based on query results
• ask: whether a query returns results (result is true/false)
• describe: describes resources in the graph

SPARQL Example
# Software companies founded in the US
PREFIX rdf:<http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX dbpedia: <http://dbpedia.org/resource/>
PREFIX dbp-ont: <http://dbpedia.org/ontology/>
PREFIX geo-ont: <http://www.geonames.org/ontology#>
PREFIX umbel-sc: <http://umbel.org/umbel/sc/>
SELECT DISTINCT ?Company ?Location
WHERE {
?Company rdf:type dbp-ont:Company ;
dbp-ont:industry dbpedia:Computer_software ;
dbp-ont:foundationPlace ?Location .
?Location geo-ont:parentFeature dbpedia:United_States .
}

SPARQL Results
● Try at: http://factforge.net/sparql

Documents mentioning Persons born in Sheffield
{Person sparql = "SELECT ?inst WHERE { ?inst :birthPlace
<http://dbpedia.org/resource/Sheffield>}"}
The BBC web document does not mention Sheffield at all:
http://www.bbc.co.uk/news/uk-politics-11494915
The relevant text snippet is:

Try these on the BBC News demo:
http://services.gate.ac.uk/mimir/gpd/search/index
● Gordon Brown [0..3] root:say
● {Person} [0..3] root:say
● {Person inst ="http://dbpedia.org/resource/Gordon_Brown"}
[0..3] root:say
● {Person sparql="SELECT ?inst WHERE { ?inst a :Politician }"}
[0..3] root:say
● {Person sparql = "SELECT ?inst WHERE { ?inst :party
<http://dbpedia.org/resource/Labour_Party_%28UK%29> }" }
[0..3] root:say
● {Person sparql = "SELECT ?inst WHERE {
?inst :party <http://dbpedia.org/resource/Labour_Party_
%28UK%29> .
?inst :almaMater
<http://dbpedia.org/resource/University_of_Edinburgh> }"
} [0..3] root:say

User Interfaces for SPARQL-based Semantic
Search
• SPARQL-based semantic searches, tapping into LOD resources are
extremely powerful
• However, impossible to write by the vast majority of users
• User interfaces for SPARQL-based semantic search:
• Faceted searches (see ExoPatent next)
• Form-based searches (see EnviLOD, PIN)
• Text-based searches (natural language interfaces for querying
ontologies), e.g. FREyA

Faceted Search: ExoPatent Example
● Use semantic information to expose linkages between
documents based on the intersecting relationships between
various sets of data from
● FDA Orange Book (23,000 patented drugs
● Unified Medical Language System (UMLS) – database of medical
terms (370,000)
● Patent bibliographic information
● Search for diseases, drug names, body parts, references to
literature and other patents, numeric values, ranges
● Demo uses a small set of patents (40,000)

ExoPatent: Faceted Search

Annotated Patent Document

Find all applicants who filed patents related
to mitochondria, as well as drug names and
active ingredients

Semantic Search over Content and Annotations
http://demos.gate.ac.uk/trendminer/envilod

EnviLOD Semantic Search UI

Example Results

…and the underlying SPARQL Query
{Sem_Location dbpediaSparql="select distinct ?inst
where {
{{ ?inst <http://dbpedia.org/property/north> ?loc} UNION
{ ?inst <http://dbpedia.org/property/east> ?loc } UNION
{ ?inst <http://dbpedia.org/property/west> ?loc } UNION
{ ?inst <http://dbpedia.org/property/south> ?loc } UNION
{ ?inst <http://dbpedia.org/property/northeast> ?loc } UNION
{ ?inst <http://dbpedia.org/property/northwest> ?loc } UNION
{ ?inst <http://dbpedia.org/property/southeast> ?loc } UNION
{ ?inst <http://dbpedia.org/property/southwest> ?loc }
FILTER(REGEX(STR(?loc), "Sheffield", "i"))} }"} AND (root:"flood")
Ongoing work: Use GeoSparql instead and be able to specify distances and
reason with the richer information in GeoNames

Flooding in Oxford

Some Caveats….
● EnviLOD demonstrator built in a 6 month project
● VERY limited environmental science content indexed
– Some Defra, Environment Agency, Scottish Government reports
● PDFs of the articles are not connected to

We don't just have to look at politicians saying and
measuring things
● If we first process the text with other NLP tools such as
sentiment analysis, we can also search for positive or negative
documents
● Or positive/negative comments about certain
people/topics/things/companies
● In the Decarbonet project, we're looking at people's opinions
about topics relating to climate change, e.g. fracking
● We could index on the sentiment annotations too
● Other people are using the combination of opinion mining and
MIMIR to look at e.g. customer feedback about products /
companies on a huge corpus

A positive tweet

A negative tweet

A Sarcastic Tweet

Explicitly Choosing The Search Classes

Choosing A Specific Instance

What diseases are in these documents?

What Pathogens?

Disease vs Disease Co-ocurrences

Diseases vs Pathogens

Summary
● Text mining is a very useful pre-requisite for doing all kinds of
more interesting things, like semantic search
● Semantic web search allows you to do much more interesting
kinds of search than a standard text-based search
● Text mining is hard, so it won't always be correct, however
● This is especially true on lower quality text such as social
media
● On the other hand, social media has some of the most
interesting material to look at
● Still plenty of work to be done, but there are lots of tools that
you can use now and get useful results
● We run annual GATE training courses where you can spend a
whole week learning all this and more!

Acknowledgements and further information
● Research partially supported by the European Union/EU under the Information
and Communication Technologies (ICT) theme of the 7th Framework Programme
for R&D (FP7) DecarboNet (610829) http://www.decarbonet.eu and TrendMiner
(287863)
● Semantic Search over Documents and Ontologies (2014) K Bontcheva, V Tablan,
H Cunningham. Bridging Between Information Retrieval and Databases, 31-53
● V. Tablan, K. Bontcheva, I. Roberts, and H. Cunningham. Mímir: An open-source
semantic search framework for interactive information seeking and discovery.
Journal of Web Semantics: Science, Services and Agents on the World Wide
Web, 2014
● These and other relevant papers on the GATE website:
http://gate.ac.uk/gate/doc/papers.html
● Annual GATE training course in June in
Sheffield:https://gate.ac.uk/family/training.html
● Download gate: http://gate.ac.uk/download

Text analysis and Semantic Search with GATE

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Text analysis and Semantic Search with GATE (20)

More from Diana Maynard (20)

Recently uploaded (19)

Text analysis and Semantic Search with GATE