User:Khannatanmai/GSoC2020 Final Report

From Apertium
Jump to navigation Jump to search

This is the final report for the GSoC 2020 project initially titled "Modifying the apertium stream format and eliminating dictionary trimming", which ultimately became "Modifying the apertium stream format to introduce wordbound blanks and solve the markup reordering problem".

The problem

Format/Markup handling has been a problem in Apertium and is described in detail here. The way format handling was done is using superblanks that protect markup information inside superblanks during the translation process. While this works well to protect the formatting of the document, during translation, words/phrases move around, get deleted, split, merge, etc., and the markup information on the words needs to go through these with the words, otherwise we will end up with erroneous markup, which is what happened in Apertium.

$ echo '<i>Perro</i> <b>blanco</b>' | apertium spa-eng -f html
<i>White</i> <b>dog</b>

The solution

A new kind of blank was proposed - wordbound blank. This blank contains any information that needs to stay attached to a word/phrase during the translation process. At this point, this is markup information, but introducing wordbound blanks basically introduces alignments in Apertium and there are several use cases for information inside wordbound blanks.

Wordbound blanks will be denoted by double square brackets and will always appear right before the Lexical Unit to which they are attached.

[[wordboundblank]]^LU<tags>$

If there is no Lexical Unit in the stream (before the morph analyser and after the generator), then we have an end wblank as well.

[[wordboundblank]]word[[/]] word2 word3 [[wordboundblank]]word4[[/]]


Method

A new deformatter and reformatter was written, that converts markup into either superblanks or wordbound blanks, and each module in the entire pipeline was modified to work with wordbound blanks.

When wordbound blanks are on text that isn't tokenised using LUs, there's a closing wordbound blank

[[/]]

to limit it's span. After the input is tokenised by the analyser, the program apertium-wblank-attach attaches wordbound blanks to all the LUs inside the span of that wordbound blank and removes the closing wordbound blanks from the stream. Right before the generator, the opposite process happens with the apertium-wblank-detach - closing wordbound blanks are put in the stream again in preparation for the reformatter.

Here's a snippet of the process:

Input:
<i>Perro</i> <b>blanco</b>

tf-extract:
[[t:i:4_tPUA]]Perro[[/]] [[t:b:ls9z2Q]]blanco[[/]]

Analyser (lt-proc -z):
[[t:i:4_tPUA]]^Perro/Perro<n><m><sg>$[[/]] [[t:b:ls9z2Q]]^blanco/blanco<n><m><sg>/blanco<adj><m><sg>$[[/]]

apertium-wblank-attach:
[[t:i:4_tPUA]]^Perro/Perro<n><m><sg>$ [[t:b:ls9z2Q]]^blanco/blanco<n><m><sg>/blanco<adj><m><sg>$

.
.
.

Transfer Output:
[[t:b:ls9z2Q]]^White<adj><sint>$ [[t:i:4_tPUA]]^dog<n><sg>$

apertium-wblank-detach:
[[t:b:ls9z2Q]]^White<adj><sint>$[[/]] [[t:i:4_tPUA]]^dog<n><sg>$[[/]]

Generator (lt-proc -g):
[[t:b:ls9z2Q]]White[[/]] [[t:i:4_tPUA]]dog[[/]]

tf-inject:
<b>White</b> <i>dog</i>

Here is the full pipeline:

$ echo "<i>Perro</i> <b>blanco</b>"      |
 tf-extract                              |
 lt-proc -z 'spa-eng.automorf.bin'.      |
 apertium-wblank-attach                  |
 apertium-tagger -z -g $2 'spa-eng.prob' |
 apertium-pretransfer -z                 |
 lt-proc -z -b 'spa-eng.autobil.bin'     |
 lrx-proc -z -m 'spa-eng.autolex.bin'    |
 apertium-transfer -z -b 'apertium-eng-spa.spa-eng.t1x'  'spa-eng.t1x.bin' |
 apertium-interchunk -z 'apertium-eng-spa.spa-eng.t2x'  'spa-eng.t2x.bin'  |
 apertium-postchunk -z 'apertium-eng-spa.spa-eng.t3x'  'spa-eng.t3x.bin'.  |
 apertium-wblank-detach                  |
 lt-proc -z -g 'spa-eng.autogen.bin'     |
 lt-proc -z -p 'spa-eng.autopgen.bin'    |
 tf-inject

<b>White</b> <i>dog</i>

Development

Here is a list of changes made to the Apertium codebase during GSoC 2020.

Transfer (Pull Request 1, Pull Request 2, Commit 1, Commit 2)

Chunker/Single-stage transfer

  • Wordbound blanks are a part of transfer word as a new side: blank.
  • Are ignored in pattern matching
  • Wordbound blanks are added just before the output LU from the LU that the lem/lemh is clipped from.
  • If the lem/lemh comes from a variable in the output then the balnk come from the LU which the lemma comes from, by tracing its variable assignment in <let>.
  • No regression. Stream without wordbound blanks work as-is.
  • Normal blanks don't move around while wordbound blanks move around.
  • When MLUs are formed the blanks are merged.
  • Tests added
  • If rule pattern has only one LU, the wordbound blank gets output with all output LUs of the rule
  • When using apertium-transfer -n, the wblanks print as they're supposed to.

Interchunk

  • No change needed as inter chunk doesn't access LUs inside the chunk.

Postchunk

  • Wordbound blanks are ignored in pattern matching
  • Wordbound blanks are added just before the output LU from the LU that the lem/lemh/whole is clipped from.
  • If the lem/lemh comes from a variable in the output then the blank comes from the LU which the lemma comes from, by tracing its variable assignment in .
  • No regression. Stream without wordbound blanks work as-is.
  • Normal blanks don't move around while wordbound blanks move around.
  • When MLUs are formed the blanks are merged.
  • Tests added
  • If rule pattern chunk has only one LU, the wordbound blank gets output with all output LUs of the rule

Recursive Transfer (Pull Request)

  • Wordbound blanks are read as part of LUs as a new side->wblank.
  • Wblanks reorder with the LUs in transfer based on where the lemma is clipped from.
  • Works even if lemma is clipped into a variable and the variable is later added in the output.
  • No regression. Stream without wordbound blanks work as-is.
  • Normal blanks don't move around while wordbound blanks move around.
  • When MLUs are formed the blanks are merged.
  • Tests added

Pretransfer (Pull Request)

  • Wordbound blanks distribute across parts when compounds are split into individual LUs

Separable (Pull Request)

  • Merge wordbound blanks and add to all LUs in rule output.
  • Works for both autoseq and revautoseq.

Analysis, Biltrans, Generation (Pull Request)

  • Parsing wordbound blanks as normal blanks for analysis, generation, biltrans.
  • Added a test for wordbound blank analysis.

HFST Analysis, Generation (Pull Request)

  • Parsing wordbound blanks as normal blanks for analysis and generation in hfst-proc.

Streamparser (Pull Request)

  • Wordbound blanks parsed as part of a lexical unit in the stream parser.
  • Can be accessed by class member: LexicalUnit.wordbound_blank.

Postgeneration (Pull Request)

  • Wordbound blanks merge when words merge.
  • Wordbound blanks apply to all output words when output of postgen rule are more than input words.
  • No regression for postgeneration without wordbound blanks.
  • Lots of tests added.

Tagger (Pull Request)

  • Parse wblanks as normal blanks

References