Supporting e-learning with automatic glossary extraction Experiments with Portuguese

1
Supporting e-learning with automatic
glossaryextraction Experiments with Portuguese

• Rosa Del Gaudio, António Branco
• RANLP, Borovets 2007

2
Presentation Plan

• LT4eL project
• ILIAS
• Corpus
• Tool
• Grammars
• Copula
• Other Verbs
• Punctuation
• Results
• Conclusion

3
LT4eL

• Improve retrieval and accessibility of LO in
  learning management systems
• Employ language technology resources and tools
  for the semi-automatic generation of descriptive
  metadata .
• Develop new functionalities such as a key word
  extractor and a glossary candidate detector,
  semantic search, tuned for the various languages
  addressed in the project (Bulgarian, Czech,
  Dutch, English, German, Maltese, Polish,
  Portuguese, Romanian).

4
ILIAS
5
Objective

• Build a Glossary in an automatic way to support
  e-learning process. In practice this means to
  extract a definition from unstructured text
  (scientific papers, enciclopedia, web pages)
• Better access to information for student
• Accelerate the work of the tutor

6
ILIAS Glossary Candidate Detector
7
The Corpus

• 274.000 tokens
• Tutorials
• PhD Thesis
• Scientific papers
• 3 Domains evenly represented
• e-learning
• Technology for non experts
• Calimera

8
XML format

• ltdefiningText continue"y" def"m147"
  def_type1"is_def" id"d5"gt
• ltmarkedTerm dt"y" id"m147" kw"y"gt
• lttok base"intranet" class"word" ctag"PNM"
  id"t9032" sp"y"gtIntranetlt/tokgt
• lt/markedTermgt
• lttok base"ser" class"word" ctag"V" id"t9033"
  msd"pi-3s" sp"y"gtélt/tokgt
• lttok base"uma" class"word" ctag"UM" id"t9034"
  msd"fs" sp"y"gtumalt/tokgt
• lttok base"rede" class"word" ctag"CN"
  id"t9035" msd"fs" sp"y"gtredelt/tokgt
• lttok base"desenvolver,desenvolvido" class"word"
  ctag"PPA" id"t9036" msd"fs" sp"y"gtdesenvolvida
  lt/tokgt
• lttok base"para" class"word" ctag"PREP"
  id"t9037" sp"y"gtparalt/tokgt
• lttok base"processamento" class"word" ctag"CN"
  id"t9038" msd"ms" sp"y"gtprocessamentolt/tokgt
• lttok base"de" class"word" ctag"PREP"
  id"t9039" sp"y"gtdelt/tokgt
• lttok base"informação" class"word" ctag"CN"
  id"t9040" msd"fp" sp"y"gtinformaçõeslt/tokgt
• lttok base"em" class"word" ctag"PREP"
  id"t9041" sp"y"gtemlt/tokgt
• lttok base"uma" class"word" ctag"UM" id"t9042"
  msd"fs" sp"y"gtumalt/tokgt
• lttok base"empresa" class"word" ctag"CN"
  id"t9043" msd"fs" sp"y"gtempresalt/tokgt
• lttok base"ou" class"word" ctag"CJ" id"t9044"
  sp"y"gtoult/tokgt
• lttok base"organização" class"word" ctag"CN"
  id"t9045" msd"fs"gtorganizaçãolt/tokgt
• lttok class"punctuation" ctag"PNT" id"t9046"
  sp"y"gt.lt/tokgt
• lt/definingTextgt

9
LxTransduce

• Match tree using elements
• Quick
• Unicode friendly
• freeware
• Easy to integrate in other tools (java)

• Input simple text or xml
• Regular expressions
• Substitution and markup
• Output the same file with changes

10
Rules in lxtransduce

• ltrule name"Conj"gt
• ltquery match"tok_at_ctag 'CJ'"/gt
• lt/rulegt
• ltrule name"Coor"gt lt!--Conjunctions or comma --gt
• ltfirstgt
• ltquery match"tok. ','"/gt
• ltref name"Conj" mult""/gt
• lt/firstgt
• lt/rulegt

• ltrule name"PARopen"gt
• ltquery match"tok.'\('"/gt
• lt/rulegt
• ltrule name"PARcl"gt
• ltquery match"tok.'\('"/gt
• lt/rulegt
• ltrule name"parenthetic"gt
• ltseqgt
• ltref name"PARopen"/gt
• ltrepeat-until name"tok"gt
• ltref name"PARcl"/gt
• lt/repeat-untilgt
• ltref name"PARcl"/gt
• lt/seqgt
• lt/rulegt

11
First developmentphase

• Less than 50 of the corpus
• Focus on the verb
• Precision manually marked/all automatic
• Recall correct automatic/manually marked
• F2 3(precisionrecall)/2precisionrecall

12
Second developing phase

• 75 of the corpus for developing
• 25 of the corpus for testing
• Specific grammar/rules for each type

13
Copula baseline grammar

• Verb to be third person singular or plural
  present indicative
• ltrule name"SERdef"gt
• ltbestgt
• ltref name"Ser3"/gt
• ltref name"PoderSer"/gt
• lt/bestgt
• lt/rulegt

• ltrule name"euristic"gt
• ltseqgt
• ltrepeat-until name"tok"gt
• ltref name"SERdef" mult""/gt
• lt/repeat-untilgt
• ltref name"SERdef" mult""/gt
• ltnotgt
• ltref name"PPA"/gt
• lt/notgt
• ltref name"tok" mult""/gt
• ltend/gt
• lt/seqgt
• lt/rulegt

14
Copula base result

• Sentence level results
• Problem with precision

15
Copula Grammar
16
Rules for is_type

• lt!-- To Be 3rd person pl and s --gt
• ltrule name"Serdef"gt
• ltquery
• match"tok_at_ctag V and _at_baseser and
• (_at_msdstarts-with(.,fi-3 )
• or _at_msdstarts-with(.,pi-3 ))
• lt/rulegt
• ....

• ltrule name"copula1"gt
• ltseqgt
• ltref name"SERdef"/gt
• ltbestgt
• ltseqgt
• ltref name"Art"/gt
• ltref name"adjadvprep" mult""/gt
• ltref name"Noun" mult""/gt
• lt/seqgt
• ....
• lt/bestgt
• ltref name"tok" mult""/gt
• ltend/gt
• lt/seqgt
• lt/rulegt

17
Confronting Results
Include that patterns that were excluded Try to
gather the syntactic pattern of non definition
and confront with the syntactic pattern of
definition.
18
Other_Verbs grammar

• ltrule name"Vpas"gt
• ltseqgt
• ltref name"tok"/gt
• ltnotgt
• ltref name"not"/gt
• lt/notgt
• ltref name"tok" mult"?"/gt
• ltquery match"tokmylex(_at_base) and
  (_at_ctag'PPA')" constraint"mylex(_at_base)/cat'pas'
  "/gt
• lt/seqgt
• lt/rulegt

• Collect verbs in a lexicon
• Three different category reflexive, active,
  passive.
• 22 different verbs
• ltlex word"chamar"gt
• ltcatgtreflt/catgt
• lt/lexgt
• ltlex word"chamar,chamado"gt
• ltcatgtpaslt/catgt
• lt/lexgt

19
Results for verb_type

• Analyze each verbs separately as with is_type
• Richer syntactic patterns

20
Punctuation Grammar

• ltrule name"punct_def"gt
• ltseqgt
• ltstart/gt
• ltref name"CompmylexSN" mult""/gt
• ltquery match"tok.\"/gt
• ltref name"tok" mult""/gt
• ltend/gt
• lt/seqgt
• lt/rulegt

• Preliminary work
• Definition introduced by colon mark (most
  frequent)

21
All-in-one

• Combination of the previous grammars
• The type is not take into account to calculate
  precision and recall

22
Conclusions and Future Work

• Overall results Recall 86, Precision 14
• Difference among domains the style of a document
  influence the result.
• Improve the rules for verb_type and punc_type
• Combining with other techniques such as ML