Cpascoe pimms or2012_


Published on

The PIMMS project and Natural Language Processing for Climate Science

Published in: Career, Technology, Education
1 Like
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • skip
  • Cpascoe pimms or2012_

    1. 1. The PIMMS project and Natural Language Processing for Climate ScienceExtending the Chemical Tagger natural language processing tool with climate science controlled vocabularies Charlotte Pascoe, Hannah Barjat Peter Murray-Rust and Gerry Devine June 9th 2012, Open Repositories 2012
    2. 2. Portable Infrastructure for the Metafor Metadata System
    3. 3. Common Information Model Data Software We can talk about DataObjects collected together in any number of ways, stored in a particular medium Shared ISO We reuse various ISO classes Quality We can talk aboutSome concepts hierarchicalare shared ModelComponents with We can record the ModelProperties, som quality of things A particular Activity uses e of which can be a particular coupled together Grids Activity SoftwareComponent We can talk about Simulations run in support of Experiments. Experiments consist of Requirements; We can define a GridSpec Simulations conform to or some other geometry Requirements
    4. 4. Common Information Model
    5. 5. Mind MapsMind maps are used to captureinformation requirements from domainexperts and build a controlled vocabulary.
    6. 6. Python ParserThe python parser processes the XML files generated by the mind maps<component name="Radiation"> <definition status="missing">Definition of component type Radiation required</definition> <parameter name="RadiativeTimeStep" choice="keyboard"> <definition status="missing">Definition of property name RadiativeTimeStep required</definition> <value format="numerical" name="time step" units="time units"/> </parameter> <parametergroup name="Longwave"> <parameter name="SchemeType" choice="XOR"> <definition status="missing">Definition of property name SchemeType required</definition> <value name="Wide-band model"/> <value name="Wide-band (Morcrette)"/> <value name="K-correlated"/> <value name="K-correlated (RRTM)"/> <value name="other"/> </parameter> <parameter name="Method" choice="XOR"> <definition status="missing">Definition of property name Method required</definition> <value name="Two stream"/> <value name="Layer interaction"/> <value name="other"/> </parameter> <parameter name="NumberOfSpectralIntervals" choice="keyboard"> <definition status="missing">Definition of property name NumberOfSpectralIntervals required</definition> <value format="numerical" name=""/> </parameter> </parametergroup>
    7. 7. Web FormsWeb forms generate content in CIM xml format
    8. 8. CIM Viewer
    9. 9. Chemical Tagger is an open-source tool that uses OSCAR4 and NLP techniques for tagging andparsing experimental sections in the chemistry literature.
    10. 10. Chemical Tagger &• Java project Developed by the Peter Murray-Rust group, Cambridge. Online demo:• Adapted for use with ACP Abstracts (Lezan Hawizy and Hannah Barjat). – Modification by use of dictionaries and changes to grammar. – First use case outside of laboratory chemistry. – Still with a significant chemistry component. – Wider physical science. • Open Source NLP tool for processing• Open Source NLP tool for processing chemical text chemical text• Combines Chemical Entity Recognitions (OSCAR) with NLP • techniquesChemical Entity Recognitions Combines• Extendible and Reconfigurable Taggers and Parsers (OSCAR) with NLP techniques • Extendible and Reconfigurable Taggers and Parsers generated using ANTLR (ANother Tool for Language Recognition)
    11. 11. Chemical Tagger & PIMMS• To extend chemical tagger to be more suited to climate modelling. – Specifically: • Palaeoclimate modelling and how process of text mining might differ from development of a controlled vocabulary. • High-lighting of text for comparison with CIM documents. • Initially only using XML Abstracts e.g. from EGU’s Geoscientific Model Development and Climate of the Past. – Brief look at PDF to Text.11
    12. 12. Paleoclimate Language• Time periods and climatic events – Includes named Ages, Epochs, Eras etc. [Including all those in a mind map produced for the PIMMS project at Bristol]. – context of proper nouns e.g. with words such as ‘period’, ‘era’, ‘epoch’ – Numbers with appropriate units e.g. Mya, yr BP – Likely date numbers e.g. 1750 AD. – Acronyms – known’LGM’ e.g. [in context ACRONYMS have not been investigated] – Related adjectives e.g. seasonal, decadal, glacial, interglacial, stadial, interstadial, maximum, minimum where used as proper nouns.• Palaeoclimate Models – Can guess model names from context • e.g. proper noun or acronym followed by model • e.g. reconstruction / simulation with XXX – Can develop/use glossary of model names.• Palaeoclimate Acronyms – Time periods and models. – Theories, techniques, physical and chemical parameters? – Can develop/use glossary of acronyms – problem area: often not unique even within subject.
    13. 13. Natural Language vs CV• Quick compilation of proper nouns used for time periods (primarily from Wikipedia) contains 185 words. – Use of these words together with adjective/ dates / details of events would produce a very large number of phrases.• Controlled Vocabulary from Bristol contains around 24 of these. • Use of these words together with other proper nouns / adjectives / dates gives only 44 phrases within the Bristol CV.• Map natural language to CV? – Straightforward for most dates? – Understanding of context important • Does context refer to main emphasis of paper?13 • Is an event/time period described unambiguously? e.g. “Last Glacial
    14. 14. Preliminary ResultsPreliminary Results (from 68 files) Tag / Tags Example Comment <timePhrase> (i) Holocene, (ii) 8 kyr BP <PALAEOTIME> (iii) <referencePhrase> (i) (Otto et al. 2009b) Important to distinguish (ii) Giraudeau et al. 2000 year pattern from dates relevant to the study. <locationPhrase> (i) around Lake Kotokel, False positives: e.g. “from (ii) over Tibetan Plateau Sphagnum” <LOCATION> (i) 52°47´ N, 108°07´ E, Cannot currently do 458 m a.s.l (ii) London. degrees from pdf-text. <TempPhrase> „warm‟ and „cool‟: verbs in synthetic chem unlike env. chem.
    15. 15. Tag / Tags Example Numbers found<CAMPAIGN> (i) PMIP, (ii) PANASH Less relevant here than to ACP in general<MODEL> (i) REVEALS model, (ii) ECBILT-CLIO intermediate complexity climate model<acronymPhrase> (i) Modern Analogues May pick up campaigns / Technique ( MAT ) models where phrases (ii) REVEALS ( Regional above have failed. Estimates of VEgetation Abundance from Large Sites )<QUANTITY> (i) 10 ppm (ii) 0.53 mm/day units dictionary could be more extensive<MOLECULE> (i) CO2, (ii) calcium Many false positives as carbonate what chemical tagger was designed for.
    16. 16. Chemical Tagger Rendering of PALEOTIMEXML rendered with CSS 16
    17. 17. GMD Journal Article
    18. 18. CIM Document Viewer The acronym / nameMIROC4 is not explained – so reproduce sentence The description is just first few sentences after appearance of <MODEL>
    19. 19. CIM Document Viewer Makes use of existing chemical tagging.
    20. 20. CIM Document Viewer Number of spectral intervals were not found! No place for “not found”
    21. 21. Climate Models – General Constraints• Unless paper is specifically about the model we are unlikely to find much MEAFOR type CV in the abstract – Look at experimental / methods sections • model name • model resolution • model schemes – Problem with PDF -> text. – Only certain elements easy to extract (e.g. resolution)
    22. 22. Refine ACPgeo Output• Add a few more phrases e.g. specific phrases to look for model resolution, using expected vocabulary (e.g. grid, levels, resolution, directions etc).• Refine output of ACPgeo to look for specific CV terms.• Try to put CV terms in context: – Look for proximity of CV terms to other phrases: • Within phrase; within sentence or within a number of sentences22
    23. 23. <MOLECULE>– Chemical Tagger was designed to be used primarily with chemistry. • Unsurprising that there is a tendency to to assign acronyms; hyphenated words; and words with common chemical endings as molecules. – It is possible to filter some of these wrongly assigned words by probability.– There are still conflicts e.g. C3 and C4 could refer to hydrocarbons or plants. • Extensive testing and modifying / machine learning might reduce these.– Better to get right first time if important!
    24. 24. Harvested Metadata vs Documented Metadata was designed to be populated by modellers with the (probably over simplistic) assumptionthat if something isnt in the CIM document then it either isnt in the model or isnt relevant. ButCIM documents created by harvesting information from papers will naturally not covereverything about a model, so missing info doesnt mean that those things werentincluded/arent relevant.PIMMS will need to describe different protocols for interpreting CIM documents depending onhow they were created, but we will also want to ensure that that CIM accounts for missing datamore intelligently in future releases.In essence the difference between journal article descriptions and metadata documentation isNarrative. Journal articles need to tell a story so the information they include is only that whichis relevant to the narrative, whereas metadata documentation is an attempt to include as muchas possible across the board. The general nature of metadata documentation is probably why ithas historically been perceived as such a boring task to complete.PIMMS will make metadata documentation more fun by bringing back the Narrative, oncePIMMS is established at an institution users will be able to create generalised metadata havingonly described those things that are relevant to the story of their experiment.