GATE track 1 session

A full week of learning GATE text mining/information extraction language processing and talks. Session wiki

GATE is written in Java and very Java centric. This makes it portable, fast, and heavyweight. A programming library is available. It's 14 years old and has many users and contributors.

Using GATE developer

• GATE developer is used to process sets of Language Resources in Corpus using Processing Resources. They are typically saved to a serialized Datastore.
• ANNIE, VG (verb group) processors.
• Preserve formatting embeds tags in HTML or XML.
  • Different strengths using GATE's graph (node/offset) based XML vs. preserved formatting (original xml/html)

Information Extraction

• IR - retrieve docs
• IE - retreive structured data

• Knowledge Engineering - rule based
• Learning Systems - statistical

Old Bailey IE project - 17th century english (Online)

• POS - assigned in Token (noun, verb, etc)

• Gazateer - gotcha, have to set initialization parameter listsURL before it's loaded. Must also "save and reinitialize."
• Gazeteer creates Lookups, then transducer creaties named entities
• Then orthomatcher (spelling features in common) coreference associates those

• Annotation Key sets and annotation comparing
  • Need setToKeep key in Document Reset for any pre annotated texts

Evaluation / Metrics

• Evaluation metric - mathematically against human annotated
• Scoring - performance measures for annotation types

• Precision = correct / correct + spurious
• Recall = correct / correct + missing
• F-measure is precision and recall (harmonic mean)
• F=2⋅(precision⋅recall / precision+recall)
• GATE supports average, strict, lenient

• Result types - Correct, missing, spurious, partially correct (overlapped)

• Tools > Annotations Diff - comparing human vs machine annotation

• Corpus > Corpus quality assurance - compare by type
• (B has to be the generated set)

• Annotation set transfer (in tools) - transfer between docs in pipeline
  • useful for eg HTML that has boilerplate

To investigate

• markupAware for HTML/XML (keeps tags in editor)
• AnnotationStack
• Advanced Options

JAPE

• Rules based on tokens and lookups

Phase: MatchingStyles
Input: Lookup
Options: control = appelt
Rule: Test1
(
({Lookup.majorType == location})?
{Lookup.majorType == loc_key}
):match
-->
:match.Location = {rule=Test1}

Copying features: :match.Location = { type = :match.Lookup.minorType}

To review, gotchas

• Rule types : first takes only first match, excludes compound
  • a? b for "a b" will match "a b"
• multiplexor tranducers
• multi-constraint statements
• macros
• To reuse created annotations has to be a separate rule

Matching types

To follow up

• WebSphinx crawler CREOLE plugin
• GATE NLP web services

Other notes

Lucene data store and ANNIC

• Use <null> for default set
• Go to Datastore for queries
  • eg {Person}({Token})+{Money}
• Useful for debugging JAPE and results

Demos

• Mímir for querying large volumes of data (uses MG4J)
• Translating parts of speech between languages using Compound editor and Alignment editor
• Predicate extractor (MultiPaX)
  • Mixed results at best
• OwlExporter
  • NLP ontology

Conclusions

While it can do a lot out of the box and benefits from development time and breadth of connectivity, to be useful to more than patient specialists, it needs usability testing. A lot of things are inobvious and too domain specific that with a bit of work could be more broadly useful. Interaction could include a lot more immediate, useful and interesting looking displays. A web based version could have these features. However the team seems somewhat ambivalent about development. :)

Looking forward to learning about programming using GATE libraries.

RSS

Blikied on Aug 30, 2010