Difference between revisions of "Document-level Open IE"

Latest revision as of 22:37, 22 November 2013

Goals

• Extend sentence-based Open IE extractors to incorporate document-level reasoning, such as:
  • Coreference
  • Entity Linking
  • NER
  • Rules implemented for TAC 2013 Entity Linking
• Define necessary data structures and interfaces by Oct-9 (done)
• Preliminary End-to-end system evaluation by Nov-11 (done)
• Quantitatively determine how much this adds to present Open IE

Work Log

11-22

Trained and evaluated a linear classifier for best-mentions (instances of a rule-application or best-mention-resolution), which provides 95% precision at 90% yield over news data. Features include rule type (person, organization, location), whether coreference info was used, and the ambiguity of a given mention. Todo from here:

• Polish features:
  • Include coreference info when deciding candidates for a rule (currently coreference is only considered after applying rules)
  • Improve ambiguity measures: Return a value that indicates prominence of the chosen mention out of other ambiguous mentions
  • Improve location ambiguity measure - fix a technical issue reading tipster gazetteer. Consider city, stateOrProvince, and Country ambiguity separately.
  • Debug an issue where names are resolved to something only matching a prefix (e.g. Steven Miller -> Steven Tyler)
• Produce a formal evaluation
  • How much does this help "Open IE?"
    • How many extractions get annotated with additional, useful information?
    • How many more links get found as a result of best mentions? Coref? Are they higher confidence links?
    • How much does it increase running time to do Document-level processing with/without Coref?
• Code cleanup, packaging, and release.

11-12

Implemented serialization to allow extractor input to be saved to disk after pre-processing, saving us from redoing with each run:

• Parsing
• Chunking
• Stemming
• Coref
• Sentence-level Open IE
• NER tagging

This saves roughly 3 minutes per run (over 20 docs) and will greatly speed development time.

Started refactoring and cleaning up rules. Next step: get all substitution rules "on equal footing" programatically so that a classifier can be built to rank them.

11-8

Finished annotating data and discussed results. String substitution rules need to be tightened up, and a confidence measure over them would help greatly.

Extraction-level stats

From 20 documents, there were 528 total extractions in all runs.

Rules-diff: -- 206 extractions in diff -- 75 baseline better -- 99 rule-based system better -- 33 bad extractions (neither better)

Coref-diff: -- 280 extractions in diff -- 105 baseline better -- 115 coref+rule-based system better -- 59 bad extractions (neither better)

I took a closer look at the "baseline better" cases to see where we were getting it wrong:

Rule-based system -- 49 strange string errors e.g. "CDC" -> "CIENCE SLIGHTED IN" -- 16 location errors (e.g. "Washington" [DC] -> "Washington, Georgia") -- 8 entity disambiguation errors, e.g. ("he" [Scott Peterson] => "Laci Peterson") -- 1 incorrect link (e.g. "the theory" linked to "Theory" in FreeBase) -- 75 total

Coref+rule-based system -- 49 strange string errors -- 11 location errors -- 13 entity disambiguation errors -- 17 incorrect links -- 6 coref errors (e.g. "make it clear that" -> "make the CDC clear that") -- 105 total

Approximate running times over 20 documents: Baseline: 45 sec Rules: 45 Sec Rules+Coref: 230 sec

11-4

• Released system output for evaluation:
  • "Rules" configuration, using rule-based best-mention disambiguation, NO Coref.
  • "Coref" configuration, using coref-assisted rule-based best-mention disambiguation. Entity Linking context also extended via coreference.
  • Entity Linking output, showing differences in Entitylinks between each system configuration (and baseline).

Next: Stephen, John and I will annotate the output and analyze performance.

10-25

Met with Stephen, John, and Michael. Items:

• Create a (very simple) webapp for doc extractor
• Cleanup arguments before submitting them to the linker.
• Replace best-mention substrings rather than substituting best mentions for the entire argument.
• Reformat evaluation output to show only extractions that have been annotated with additional info (diff)
• Evaluate difference in linker performance with/without document-level info.

10-18

Met with Stephen and John. Discussed:

• Evaluation systems:
  • Baseline sentence extractor with entity linker, no coreference
  • Full system with best-mention finding rules
  • Full system without coreference.
• Evaluation data:
  • Sample of 20-30 documents from TAC 2013.
  • Moving away from QA/Query based approach, since the queries/questions will bias evaluation of the document extractor.
  • Instead, we will evaluate all (or a uniform sample) of extractions.
• Evaluation criteria:
  • Extractions "correct" if their arguments are as unambiguous as possible given the document text.
  • Measure prec/yield using this metric and compare systems.

10-17

Completed: Integrated sentence-level Open IE and Freebase Linker, test run OK.

Next Goals:

• Integrate best-mention finding rules.
  • First: Drop in code "as-is"
  • After: Factor out NER tagging, coref components
• Fix issues with tracking character offsets
  • Offsets are not properly computed for Open IE extractions
  • Find a good way for retrieving document metadata by character offset.

10-9

Short term goal - define necessary interfaces and data structures by 10-11

• Implemented interfaces for:
  • Document
  • Sentence
  • Extraction
  • Argument/Relation
  • Coreference Mention
  • Coreference Cluster
  • Entity Link
• Discussed interfaces at length with John and Michael
  • Interfaces to be incorporated into generic NLP tool library (nlptools):
    • Document
    • Sentence
    • CorefResolver