Chunking: A full example

From Apertium
Jump to navigation Jump to search

This will be a full example of chunking, which we build from the ground up.

We will look at Esperanto <-> English and try to translate the sentence "La libro estas blua" to "The book is blue".

Overview

First a little overview of how 3-stage transfer normally works:

  • Transfer stage: Words are translated using the bidix and categorized and put into chunks (in the .t1x file). Here the tags in the words can also be added, removed or made into 'pointers' that points to the tags in the enclosing chunk.
  • Interchunk stage: Chunks are reordered, combined and split and chunk tags changed (in the .t2x file)
  • Postchunk stage: The words in the chunks are restored (in the .t3x file)


If we look at how "The blue book is good" goes throgh the system, we have just before transfer:

^The<det><def><sp>$ ^blue<adj>$ ^book<n><sg>$ ^be<vbser><pres><p3><sg>$ ^good<adj><sint>$

which is tranfered to Esperanto and chunked into

^det_adj_nom<SN><sg><nom>{^La<det><def><2><3>$ ^blua<adj><2><3>$ ^libro<n><2><3>$}$ 
^ser<SV><pres><p3><sg>{^esti<vbser><pres>$}$ 
^adj<SN><sg>{^bona<adj><sg><nom>$}$

Here 'det_adj_nom' is the name of the chunk and <SN><sg><nom> the chunk's tags. The content of the chunk is {^La<det><def><2><3>$ ^blua<adj><2><3>$ ^libro<n><2><3>$} where the <2> and <3> are pointers to the chunk's tag (<sg> and <nom> respectively). This allows us to change the values at chunk level later on, if necessary.

In this simple case nothing happens at the interchunk stage. After the postchunk stage it looks like:

^La<det><def><sg><nom>$ ^blua<adj><sg><nom>$ ^libro<n><sg><nom>$ ^esti<vbser><pres>$ ^bona<adj><sg><nom>$

which becomes "La blua libro estas bona".

Starting from the ground

Now we will try the same sentence Esperanto -> English, but with more or less empty t1x, t2x and t3x files. After tagger disambiguation "La blua libro estas bona" is:

^La<det><def><sp>$ ^blua<adj><sg><nom>$ ^libro<n><sg><nom>$ 
^esti<vbser><pres>$ 
^bona<adj><sg><nom>$

Without any rules the result will just be that each work get a 'default' chunk:

^default{^The<det><def><sp>$}$ ^default{^blue<adj><sg><nom>$}$ ^default{^book<n><sg><nom>$}$  
^default{^be<vbser><pres>$}$  
^default{^good<adj><sint><sg><nom>$}$

Now let's define some rules. To begin with we will define some basic one-word rules à la Apertium New Language Pair HOWTO#Transfer rules.

The determiner

First for the determiner. First we define the category 'c_det' which contains all words marked as <det> and <det><other tags>. Then we define attributes for number and kind of determiner.

  <def-cat n="c_det">
     <cat-item tags="det"/>
     <cat-item tags="det.*"/>
  </def-cat>

...
  <def-attr n="a_nbr">
     <attr-item tags="sg"/>
     <attr-item tags="sp"/>
     <attr-item tags="pl"/>
  </def-attr>

  <def-attr n="a_det">
     <attr-item tags="det.def"/>
     <attr-item tags="det.ind"/>
     <attr-item tags="det.pos"/>
     <attr-item tags="det.qnt"/>
  </def-attr>
...
  <rule>
     <pattern>
       <pattern-item n="c_det"/>
     </pattern>
     <action>
       <out>
         <chunk name="det" case="caseFirstWord">
           <tags>
             <tag><lit-tag v="SN"/></tag>
             <tag><clip pos="1" side="tl" part="a_nbr"/></tag>
           </tags>
           <lu>
             <clip pos="1" side="tl" part="lem"/>
             <clip pos="1" side="tl" part="a_det"/>
             <clip pos="1" side="sl" part="a_nbr" link-to="2"/>
           </lu>
         </chunk>
       </out>
     </action>
  </rule>

The rule looks for category c_det (anything starting with <det>) and therefore recognizes the ^La<det><def><sp>$ and outputs

^det<SN><sp>{^The<det><def><2>$}$ 

"outside", on the chunk is put first attribute <SN> and then the number attribute. Inside the chunk (in a 'lexical unit') is put the translated lemma (the bidix has la -> the), the determiner attributes (<det><def>) and a pointer to tag 2 (the part="a_nbr" must be there but is ignored in chunking mode).


Adjectives

Here we have ^blua<adj><sg><nom>$ and ^bona<adj><sg><nom>$ which we - if they stand alone (that is, we are still at the one word = one chunk level) - would like to get chunked as

^adj<SN><sg>{^blue<adj>$}$
^adj<SN><sg>{^good<adj><sint>$}$

First we define the category c_adj and then the attributes that can appear on an adjective:

  <def-cat n="c_adj">
     <cat-item tags="adj.*"/>
  </def-cat>
...
  <def-attr n="a_adj">
     <attr-item tags="adj"/>
     <attr-item tags="adj.comp"/>
     <attr-item tags="adj.sup"/>
     <attr-item tags="adj.sint"/>
     <attr-item tags="adj.sint.comp"/>
     <attr-item tags="adj.sint.sup"/>
  </def-attr>

Here it becomes a rather long list as adjectives can be comparative and superlative (blue, more blue, most blue) and synthetic (see List of symbols#Adjectives). This is taken care of by the bidix so we really don't have to worry about the syntheticness of ^good<adj><sint>$ :

<e><p><l>blua<s n="adj"/></l><r>blue<s n="adj"/></r></p></e>
<e><p><l>bona<s n="adj"/></l><r>good<s n="adj"/><s n="sint"/></r></p></e>

The transfer rule that recognizes adjectives and output chunks called adj with attributes <SN>+number of the adjective (<sg>) and inside the chunk is the lemma (good) with the adjectives attributes (<adj><sint>):

  <rule>
     <pattern>
       <pattern-item n="c_adj"/>
     </pattern>
     <action>
       <out>
         <chunk name="adj" case="caseFirstWord">
           <tags>
             <tag><lit-tag v="SN"/></tag>
             <tag><clip pos="1" side="tl" part="a_nbr"/></tag>
           </tags>
           <lu>
             <clip pos="1" side="tl" part="lem"/>
             <clip pos="1" side="tl" part="a_adj"/>
           </lu>
         </chunk>
       </out>
     </action>
  </rule>

So before, during and after transfer the adjectives looks like

^blua<adj><sg><nom>$
^adj<SN><sg>{^blue<adj>$}$
^blue<adj>$

Where did the <nom> tag go?

You may have noticed that blue in Esperanto actually was ^blua<adj><sg><nom>$. What happened to that? Well, as it wasnt mentioned in the <out> part of the rule it just disappeared. In Esperanto nouns and adjectives in accusative get an extra -n in the end, so:

  • "Mi havas la bluan libron (^blua<adj><sg><acc>$ ^libro<n><sg><acc>$) is "I have the blue book" whereas
  • "Min havas la blua libro (^blua<adj><sg><nom>$ ^libro<n><sg><nom>$) is "The blue book have me".

For now we will just ignore the <nom> and <acc> tags.


Nouns

Nouns are more or less the same game as adjectives, apart from that nouns in English are flexed according to numbers, so its important to keep the <sg> or <pl> on the noun:

  <rule>
     <pattern>
       <pattern-item n="c_nom"/>
     </pattern>
     <action>
       <out>
         <chunk name="nom" case="caseFirstWord">
           <tags>
             <tag><lit-tag v="SN"/></tag>
             <tag><clip pos="1" side="tl" part="a_nbr"/></tag>
           </tags>
           <lu>
             <clip pos="1" side="tl" part="lemh"/>
             <clip pos="1" side="tl" part="a_nom"/>
             <clip pos="1" side="tl" part="a_nbr" link-to="2"/>
             <clip pos="1" side="tl" part="lemq"/>
           </lu>
         </chunk>
       </out>
     </action>
  </rule>

Before, during and after transfer the nouns looks like

^libro<n><sg><nom>$
^nom<SN><sg>{^book<n><2>$}$
^book<n><sg>$

Multi-word chunking

Word/chunk reordering

Now that "La libro estas bona" -> "The book is good" works, lets look at how chunk reordering works. In Esperanto you make a sentence into a question by putting "Ĉu" in the start of the sentence: "Ĉu la libro estas bona?" In English the verb needs to come first: "Is the book good?".