Prague Markup Language (PML)

• Based on XML
• Designed to capture various types of linguistic annotation in a standardized form
• Morphology, dependency or phrase structures syntax, coreference, valency, discourse relations…
• Allows for
  • stand-off annotation principles
  • multi-layered
  • structured attributes
  • cross-referencing and linking
  • alternatives
  • extensibility
• PML-Schema

  Data types

  atomic, AVS, list, sequence, alternative, reference

  Data roles

  root/node of a tree, unique identifier, indexable…

PML—an Example

<sentence id="s5_5">
  <chunk id="hic43">
   <feats>
     <lemma>हो</lemma>
     <wxlemma>ho</wxlemma>
     <pos>v</pos>
     <g>m</g>
     <n>s</n>
     <p>m</p>
     <t>future</t>
   </feats>
   <ord>21</ord>
   <phrase>VG</phrase>
   <children>
     <chunk id="hic37">
      <drel>k7</drel>
      <feats>
        <lemma>अभाव</lemma>
        <wxlemma>aBAva</wxlemma>
        <pos>n</pos>
        <g>m</g>
        <n>s</n>
        <c>0</c>
      </feats>
      <ord>4</ord>
      <phrase>NP</phrase>

      <children>
       <chunk id="hic36">
         <drel>r6</drel>
         <feats>
           <lemma>पाठक</lemma>
           <wxlemma>pATaka</wxlemma>
           <pos>n</pos>
           <g>m</g>
           <n>s</n>
           <c>0</c>
         </feats>
         <ord>1</ord>
         <phrase>NP</phrase>
         <children>
           <word id="hiw74">
            <feats>
              <lemma>पाठक</lemma>
              <wxlemma>pATaka</wxlemma>
              <pos>n</pos>
              <g>m</g>
              <n>s</n>
              <c>0</c>
            </feats>
            <form>पाठक</form>
            <ord>2</ord>
            <phrase>NN</phrase>
            <wxform>pATaka</wxform>
           </word>

PML Schema—an Example

<?xml version="1.0" encoding="utf-8"?>
<pml_schema xmlns="http://ufal.mff.cuni.cz/pdt/pml/schema/" version="1.1">
  <revision>1.0.0</revision>
  <description>Converted Hyderabad Treebank morph data</description>


  <root name="hydtmorph">
    <structure>
      <member name="meta" required="0" type="meta.type"/>
      <member name="document" required="1">
        <sequence role="#TREES">
          <element name="sentence" type="sentence.type"/>
        </sequence>
      </member>
    </structure>
  </root>

  <type name="meta.type">
    <structure>
      <member name="annotation_info">
        <structure>
          <member name="version_info"><cdata format="any"/></member>
          <member name="desc"><cdata format="any"/></member>
        </structure>
      </member>
    </structure>
  </type>

  <type name="sentence.type">
    <container role="#NODE">
      <attribute name="id" role="#ID" required="1"><cdata format="ID"/></attribute>
      <sequence role="#CHILDNODES">
        <element name="chunk" type="chunk.type"/>
        <element name="word" type="word.type"/>
      </sequence>
    </container>
  </type>

  <type name="children.type">
    <sequence>
      <element name="chunk" type="chunk.type"/>
      <element name="word" type="word.type"/>
    </sequence>
  </type>

  <type name="chunk.type">
    <structure role="#NODE">
      <member name="id" role="#ID" as_attribute="1" required="1"><cdata format="ID"/></member>
      <member name="error" type="error.type"/>
      <member name="feats" type="feats.type"/>
      <member name="drel"><cdata format="any"/></member>
      <member name="ord" role="#ORDER"><cdata format="integer"/></member>
      <member name="phrase"><cdata format="any"/></member>
      <member name="children" type="children.type" role="#CHILDNODES"/>
    </structure>
  </type>

  <type name="word.type">
    <structure role="#NODE">
      <member name="id" role="#ID" as_attribute="1" required="1"><cdata format="ID"/></member>
      <member name="head"><choice><value>0</value><value>1</value></choice></member>
      <member name="error" type="error.type"/>
      <member name="feats" type="feats.type"/>
      <member name="form"><cdata format="any"/></member>
      <member name="wxform"><cdata format="any"/></member>
      <member name="ord" role="#ORDER"><cdata format="integer"/></member>
      <member name="phrase"><cdata format="any"/></member>
    </structure>
  </type>

  <type name="feats.type">
    <structure>
      <member name="wxlemma"><cdata format="any"/></member>
      <member name="lemma"><cdata format="any"/></member>
      <member name="pos"><cdata format="any"/></member>
      <member name="g"><cdata format="any"/></member>
      <member name="n"><cdata format="any"/></member>
      <member name="p"><cdata format="any"/></member>
      <member name="c"><cdata format="any"/></member>
      <member name="v"><cdata format="any"/></member>
      <member name="t"><cdata format="any"/></member>
    </structure>
  </type>

  <type name="error.type">
    <list ordered="0">
      <choice>
        <value>not-connected</value>
        <value>inside-chunk</value>
        <value>drel-like-name</value>
        <value>missing-parent</value>
        <value>duplicate-name</value>
      </choice>
    </list>
  </type>

</pml_schema>

Why Should I Use PML?

Tools to

• Browse the data
• Edit the data
• Query the data
• Validate the data (many errors reported during the conversion already)

Tree Editor TrEd

http://ufal.mff.cuni.cz./~pajas/tred Written in Perl/Tk: OS independent (MS Windows, Linux, Mac OS X) Drag-and-drop modification of trees User defined macros and style-sheets Minor modes The word “Tree” is important.

TrEd Extensions

Archive of modules, key-bindings, resources, style-sheets, etc. PDT 2.0 related (valency lexicon, multi-word expressions, discourse, coreference, etc.) a parser parallel treebank alignment support for other formats: Prague Arabic Dependency Treebank Slovene Dependency Treebank Prague Czech-English Dependency Treebank CoNLL 2009 ST (2006, 2007) Penn Treebank (+ Chinese, Arabic) Tiger Hyderabad Treebank Sinica

TrEd Extensions—Examples

• Bridging Anaphora Annotation
• Discourse Annotation
• Right-to-left: Penn Arabic Treebank 2
• Round-tripping: CoNLL (CLASSiC) or Tiger (Berlin)
  • No need to abandon your format if you want PML
• Visualization of errors (on-the-fly checking)

TrEd without GUI

btred: Batch TrEd

• Object-oriented API

  $this->parent
  $this->root
  $this->rbrother
  $this->firstson
  $this->attr("m/tag")

• Can search and/or modify the data directly

Example: Build a frequency table of POS tags in the WSJ part of the Penn Treebank

btred -q -T -N -e '
    writeln $this->{pos} unless $this->children
    ' ??/wsj*.pml \
    | sort | uniq -c | sort -n

Speeding btred Up

Parallelization

ntred

• works over ssh in a shared file system
• works in two steps:
  1. load the data into the memory of the servers
  2. submit requests serially

jtred

• works on SGE
• loads the data for each request, but several requests can be submitted at the same time

PML Tree Query

• Query and report language
• Two engines
  • SQL
  • Perl ([bnj]tred)
• Several clients
  • Browser
  • TrEd
  • CLI

The SQL engine requires the data to be converted to SQL and loaded into a database → stable datasets.

The Perl engine is slow, but works directly with the data files → data in progress.

PML-TQ: The Query Language

• Selecting all occurrences of nodes with given properties and in given relations (tree topology, cross-referencing, surface ordering, etc.)
• (Un)bounded iteration of relations (e.g. descendant{1,3}, coref_gram.rf{1,}, sibling{-1,1}, etc.)
• Multi-layered or aligned treebanks with structured attribute values
• Quantified or negated subqueries (e.g. “find all clauses with exactly three objects but no subject”)
• References (e.g. “find a parent and child that have the same case but different number”)
• Natural textual and graphical representation of the query
• Sublanguage for postprocessing and generating reports (filtering, grouping, aggregating, and sorting)
• Regular expressions, basic arithmetic and string operations

PML-TQ: Basic Syntax

Atomic attribute value equality, child relation:

nonterminal [
  cat = 'VP',
  child terminal [
    pos = 'VB'
] ]

More relations and operators:

nonterminal $n := [
  cat in { 'S','VP' },
  descendant terminal [
    pos ~ '^V',
    parent nonterminal [
      cat = $n.cat
] ] ]

• Example-based building of queries

PML-TQ: Basic Syntax (2)

Relations

• child, parent
• descendant, ancestor
• same-tree-as
• sibling
• depth-first-precedes, depth-first-follows
• order-precedes, order-follows
• References
• Implementation specific

Predicates

• (!) =, ~, ~*, in
• <, >, <=, >=

PML-TQ: Basic Syntax (3)

Operators

+, -, *, div, mod, &

Functions

• descendants(), sons()…
• length(), substr(), match()…
• round(), exp(), sqrt()…
• count(), sum(), max(), ration()…

PML-TQ: Verb Clause without a Subject (I)

Tiger Treebank nonterminal [ cat = 'S', 0x * [ label = 'SB' ], nonterminal [ cat = 'VP' ] ]
Penn Chinese Treebank nonterminal [ cat = 'IP', 0x nonterminal [ functions = 'SBJ' ], nonterminal [ cat = 'VP' ] ]

PML-TQ: Verb Clause without a Subject (II)

Penn Treebank (WSJ) nonterminal [ cat = 'S', 0x nonterminal [ functions = 'SBJ' ], nonterminal [ cat = 'VP', coindex.rf nonterminal [ cat = 'VP', sibling nonterminal [ functions = 'SBJ' ] ] ] ]

Penn Arabic Treebank nonterminal [ cat = 'VP', 0x nonterminal [ cat = 'VP' or functions = 'SBJ' ] ]

PML-TQ: Verb Clause without a Subject (III)

Prague Dependency Treebank a-node [ $$ = $aux and m/tag ~ '^V[^fs]' or $$ != $aux and m/tag ~ '^V[sf]', 0x echild a-node [ afun = 'Sb' ], ? echild a-node $aux := [ afun = 'AuxV', m/tag ~ '^V[^f]' ] ]

PML-TQ: Negation

Atomic

cat = 'NP'

∃ x ∈ cat ( x = 'NP' )

cat != 'NP'

∃ x ∈ cat ( x ≠ 'NP' )

! cat = 'NP'

∀ x ∈ cat ( x ≠ 'NP' )

* cat = 'NP'

∀ x ∈ cat ( x = 'NP' )

Same for in, ~, etc.

Existencial

nonterminal [ 0x descendant terminal [pos = "NN"]]

Layers in PML-TQ

t-node [
  a/lex.rf $n4, 
  t-node [ a/lex.rf $n3] ];
a-node $n3 := [
  a-node $n4 := [  ] ];

PML-TQ: Grammar Extraction

For Penn Treebank (WSJ), extract the underlying grammar. For each rule, show the number of applications.

nonterminal $p := [ * $ch := [  ] ]
>> give $p, $p.cat,
        first_defined($ch.cat,$ch.pos),
        lbrothers($ch)
>> give $2 & " -> " 
        & concat($3," " over $1 sort by $4)
>> for $1 give count(),$1 sort by $1 desc

PML-TQ: Non-Projectivity

CoNLL 2009 treebanks

node $upper := [ 
  same-tree-as $between,
  node $lower := [ ] ];
node $between := [ ! ancestor $upper,
  ( (order-precedes $upper 
     and order-follows $lower)
    or (order-follows $upper
        and order-precedes $lower) ) ]

$upper is not an ancestor of $between

Word Order Typology

For German

Based on CoNLL ST 2009 data

node $p := [ substr(pos,0,1) = 'V', 
  ? node $ch := [
    deprel in {'SB','OA','OC','OA2','OP'}
] ];
>> give $p.xml:id,
     if($p=$ch,
       if($p.deprel = 'ROOT','V','v'),
       substr($ch.deprel,0,1)),
     $ch.order
>> give distinct $1,
     concat($2,'' over $1 sort by $3)
>> give 
     substitute($2,'([OS])\\1+','\\1','g')
>> filter ($1 ~ 'O' and $1 ~ 'S')
>> for $1 give $1,count() sort by $2 desc

Conclusion

PML Framework

• Multi-layered typed XML data format
• OS independent tree editor TrEd
• PML-TQ Tree Query Language and Engine(s)
• Used in development of Prague Dependency Treebank, but tested on more than ten different treebanks