Supporting Universal Dependencies in Tree Editor TrEd

Outline

1. NLP
2. What are Universal Dependencies?
3. What is TrEd?
4. How do they play together?

Natural Language Processing

• Interdisciplinary: Computer science / Linguistics
• Three stages according to Wikipedia:
  1. Symbolic (1950s-1990s)
  2. Statistical (1990s-2010s)
  3. Neural (present)

Natural Language Processing

• Interdisciplinary: Computer science / Linguistics
• Three stages according to Wikipedia:
  1. Symbolic (1950s-1990s)
  2. Statistical (1990s-2010s) ← We are somewhere here.
  3. Neural (present)

The Difference between a Corpus and a Treebank

Phrase versus Dependency Trees

Phrase		Dependency
		Used in UD.

Universal Dependencies

• Framework for grammar annotation
• PoS tags, morphological features, syntactic dependencies
• Goals
  • Development of multilingual parsers
  • Cross-lingual training
  • Research of language typology and parsing
  • Linguistic research across different languages

UD—The Community

• Open community effort
• More than 600 contributors
• Over 250 treebanks in more than 150 languages

UD—Morphology

• Tokenisation:
  • syntactic words (e.g. don't expanded to 2 separate words, do and not)

UD—Morphology

• Tokenisation
• Lemma:
  • base form of the word
  • Highly language-dependant

UD—Morphology

• Tokenisation
• Lemma
• Part-of-speech tag:
  • 17 universal POS tags (e.g. NOUN, VERB, DET, …)
  • A language can use a subset, but anything beyond the listed tags should go to the Features or XPOS.

UD—Morphology

• Tokenisation
• Lemma
• Part-of-speech tag
• Extra POS:
  • Optional language/treebank specific part of speech tag

UD—Morphology

• Tokenisation
• Lemma
• Part-of-speech tag
• Extra POS
• Features:
  • Inflection (e.g. Gender=Masc|Number=Sing>)
  • Lexical (e.g. PronType: categorises pronouns or pronominal adjectives, numerals, and adverbs as personal, reciprocal, interrogative, relative, demonstrative, …)
  • A list given, but can be extended per language/treebank

UD—Syntax

• Relations between words
• Exactly 1 word is the head of the sentence, each other word depends on another word
• Fixed set of basic relation types (e.g. nsubj, obj, iobj, …)

UD—Syntax (Example)

The cat could have chased all the dogs down the street.

UD—Enhanced Dependencies

• Implicit relations between words
  • additional subject relation in control and raising constructions
  • coreference in relative clauses
• Relations to elided words
• Propagation of dependencies into or from conjuncts
• Further specification of the basic relation

UD—Enhanced Dependencies (2)

• Enhanced dependencies don’t have to form a tree
• The enhanced graph can exclude some basid dependencies
• Language and treebank dependant

UD—Enhanced Dependencies (Example 1)

Is that Microwave that you gave Dan really expensive?

UD—Enhanced Dependencies (Example 2)

At least you get to go to Florida in JANUARY.

UD—Data Format

Tab separated values

Tree Editor TrEd

• Written in Perl
  
• Using Tk (yay, Tk::Canvas!)
• Open source (moving to GitHub)
• About 100K LOC
• Editing trees by dragging nodes with a mouse
• Keyboard shortcuts
• Additional links
• Stylesheets

TrEd—Screenshot

TrEd—Batch Version

• Command line version without the GUI
• Searching or editing data programmatically
• Can be parallelised on a multi-CPU machine or in a computer cluster (SGE, Slurm)
• Similarly to perl -n or -p, one can run

  btred -N -T -e '
      FPosition()
          if $this->{deprel} eq "nsubj"
          && $this->parent->{upostag} ne "VERB"
      ' data/*.conllu | tred -l-
  

TrEd—Data Format

Prague Markup Language

• Implemented in Treex::PML based on XML::LibXML
• PML schema to organise PML data structures in the data
• Perl generates and precompiles data loading code from the schema to speed up processing

TrEd—Data Format

• PML Data Types
  • list
  • alternative
  • sequence
  • structure
  • container
  • atomic value
• Roles
  • #NODE
  • #CHILDNODES
  • #ORDER
  • …

Searching

PML Tree Query

• A query language
• 3 client interfaces
• 2 engines

Searching (2)

PML-TQ: The Query Language

• Relation between nodes
• Cross-layer relations
• Arbitrary boolean combinations of statements
• Sub-language for quantitative analysis of the results

Searching (3)

PML-TQ: The Client Interfaces

1. GUI client with an interactive query builder in TrEd
2. CLI
3. Web interface with some assistance for query building

Searching (3)

PML-TQ: The Engines

1. Perl
2. SQL

Searching (4)

PML-TQ: An example

ud.node [ deprel = "nsubj", parent ud.node $p := [ upostag != "VERB" ] ]; >> give $p.upostag >> for $1 give $1, count($1) sort by $2

PUNCT 2 X 9 PART 11 INTJ 26 DET 54 ADP 130 SYM 131 NUM 299 PROPN 605 ADV 735 AUX 929 PRON 1264 NOUN 5622 ADJ 7169

TrEd Extensions

TrEd plugins

• Perl modules
• Resources (e.g. schemata)
• Style sheets
• Annotation modes (keyboard & mouse bindings) or Minor modes
• Extension repositories (55)

UD in TrEd

• Instead of conversion scripts, works with the UD files directly (“Backend”)
  • read
  • write
  • test

    package Treex::PML::Backend::UD;
    use warnings;
    use strict;
    
    use Treex::PML::IO qw{ open_backend close_backend };
    
    sub test {
        my ($filename, $encoding) = @_;
        my $in = open_backend($filename, 'r', $encoding)
            or die "$filename: $!";
        while (<$in>) {
            return unless /^#/;
            return 1
                if /^\#\ (?:new(?:doc|par)\s+id
                           |sent_id
                           |text
                         )
                         (?:\ =)? \ ./x;
        }
    }
    
    

UD in TrEd: Stylesheet Example

Вперше за останні півроку працюю з 11-ї і встаю, коли на вулиці вже сонце.

Conclusion

UD in TrEd

• TrEd and its related tools can be used to browse UD files
• TrEd can now be used to annotate UD data from scratch

Questions

• Non-linguistic trees?
• Yours...