Elex2015Summary

FrontPage

SkE workshop †

• Herstmonceux Castle, East Essex
  • 10 August, 2015

Sketch engine data †

• August 2015
  • 400 corpora for 82 languages
  • 100+ corpora (more than 100 million tokens)
  • 30+ corpora (more than 1 billion tokens)
  • 60+ languages with POS tagged corpus
  • 42 languages with word sketches
  • 26 languages with integrated tagger
  • parallel corpora

Sketch engine marketing †

• ambassadors and advocates in the SkE community
• workshop

SkE future †

• President: Adam's wife
• MJ buys the company: Czech company will own SkE
• no immediate changes to any customers
  • at least a couple of years
    • in 5 years, no employees in UK now; all the members in Brno

Research agenda †

• parallel and distributed processing of very large text corpora
• buidling very large text corpora from the web
• corpus heterogeneity and homogeneity
  • what's in the corpus?
• corpus evaluation
  • lexicographic tasks to compare different corpora
• terminology extraction
  • more and more users; cheap opportunities; supporting tools for translators;
• corpora and language teaching
  • a book by James Thomas
• language change over time

Bilingual terminology extraction by Vit Baisa †

• monolingual terminology extraction is supported for 14 languages
• what is a term?

unithood: grammatically defined (e.g. noun phrases) †

• formalism
  • 2:NN/JJ/VVG + 1: NN(head) -- English
    • Depending on languages, more complex rules need to be made, taking into case, gender, number agreements

• Statistics
  • simple math parameter N: f(focus) + N / f(ref) + N f: relative frequency

• Output:
  • Download TBX (exchange format of terminology)/SCV

• Fine tuning
  • stoplists
  • minimum freq
  • minimum score
  • minimum character length
  • alphanumeric character only

• Multilingual terminology extraction

• Future work:
  • API: sketch engine available
  • plugins for SDL, Kilgray products

Parallel corpora by Jan Michelfeit (Lexical Computing Ltd) †

• Features
  • sentence alignment only
  • mapping file showing the alignment of sentences

• Preloaded corpora
  • Europal7
  • DGT (4M English sentences, translation memory)
  • OPUS2 (130M English sentences)
    • various docuents (EU, UN, movie, Tatoeba, etc.)

• TMX
  • a good input format for 1:1 bilingual data

Underused functionality †

Global attributes in CQL †

• 1.number=2.number tokens #1 and token #2 have the same number (plural or singular)
• 1.id = 2.ref ??

• m4 macros
  • A noun --> define('noun', '[tag="N.*"])
  • An adjective --> define ('adj', '[tag="JJ.?"]') etc.
    • make the agreement rules simpler once definition is made

• wordlist functions
  • using word attributes

• WebBootCaT
  • get seed words from Wikipedia

New functions in CQL †

• adopted by Manatee
• Query language
  • CQL: Oliver Christ
• Useful options:
  • general NOT perator; NOT within;
    • !within
    • <doc year="2010"> ! containing [word="castle"]
    • general ! <query>

• Global conditions
  • 1:[lemma="alt"] 2:[lemma="Scholoss"] & 1.case = 2.case # agreement in case
  • 1.[lemma="alt"] 2:[tag="NOUN"] & 1.head = 2.id #dependency relation

• New operators
  • [ws("test-n", "modifier", ".*")]
    • modifiers of test(noun)

• [swap(1, ws("test-n", "modifier", ".*")) & tag="*.DAT.*"]
  • with collocates being in dative
  • node の位置を変えるためのコマンド

• [lempos~10"test-n"]
  • search for the top 10 synonyms of "test-n" (synonym search)

• REGEX optimization
  • simple OR optimization
  • prefix optimization
  • n-gram prefetching
    • pre index all character in one- bi- trigrams

Ted Briscoe Keynote †

Appplications †

CLC †

• 40 M words with 80 error types coded
• 24 different exam types
• CEFR levels benchmarked to exam grades

Error-coded script †

• Transcribe
• Error coding
  • <M|our having> M:missing

Parse it with RASP †

• Parsing errors: use this to machine learning
• dependency grammar: representation of texts

Overall assessment of proficiency †

• textual features which are proxies for writing competence
• predict a grade

Machine learning †

• Training : CLC
  • Classification: number of labels (pass vs. fail)
  • Regression: try to predict real number
  • Ranking: A>B>C (e.g. rank the web pages)

• Features
  • secript length
  • errors
  • n-grams
  • parse trees

• Rank Preference --> predictiona

Pairwise Ranking SVM: †

• Joachims 2002
• Learn an optimal ranking funciton that explicitly models the grade relatihoships

Features †

• word sequences
• POS sequences
• Grammatical constructions
• Other features
  • error rate estimate (against n-gram corpora)
  • readability score

• Correlation between human grades and the system-predicted grades
  • Upper bound = 0.796 (word+POS+readability+constructions+error rate)

Highly ranked features †

• 100,000 features
  • see images

Error detection and correction suggestions †

• ensure high accuracy and reasonable coverage

• corpus-derived rules
  • error rules from CLC
  • detect incorrect word sequences
  • At least 90% incorrect occurrences

• online dictionary-derived rules

• Precision:90% Recall:10% (effect on learning rate)

• Sentence evaluation
  • Limited linguistic evidence that can be extracted automatically
  • difficulty in acquiring

Incremental Semantic Analysis (ISA) †

• Distributional semantics
  • fully incremental variation of Random Indexing
  • Similarity between words measured by context vectors

Future work †

• prompt relevance
• task achievement
• discourse organization feedback
• L1-specific feedback
• content word error detection and correction suggestions
• link to online dictionary/thesaurus
• link to courseware multiple choice exercises