User:Krvoje/Application2012

From Apertium
Jump to navigation Jump to search

(This is a draft)

GSoC application: Rule-based finite-state disambiguation

Hrvoje Peradin

hperadin@gmail.com,

krvoje on IRC: #apertium

Why is it you are interested in machine translation?

It's a perfect combination of Computer Science and Linguistics. I am very fascinated with languages, both natural or artificial. With MT it fascinates me to see a natural message transfered across a language barrier via only a process of computation. While the results are rarely perfect, it takes down communication barriers, and opens up new opportunities for learning and communication.

Why is it that you are interested in the Apertium project?

I have worked on a language pair in last year's GSoC, and it gave me great insights on rule based NLP. It gave me an invaluable chance to do real-life work on an immensely interesting topic, and to create an open-source resource. It was a great experience that taught me a lot about software development and NLP - and it also gave me the theme for my master thesis. So, I could say Apertium has a special place in my heart, and I would love to continue working on it.

Which of the published tasks are you interested in?

Writing the module for rule-based finite-state disambiguation.

Why should Google and Apertium sponsor it?

The module is intended to supplement the current bigram tagger, and Constraint Grammar, by implementing constraint based-disambiguation in a finite-state manner. Since the most common disambiguation rules can be expressed in a finite-state way, this will greatly improve speed of disambiguation, and will be beneficial for working with large texts.

How and whom it will benefit in society?

It will provide a fast tool for rule-based disambiguation, which will enable faster processing of larger corpora, and potentialy help improve translation quality in any language pair in Apertium.

What do you plan to do?

I will design an XML formalism for writing disambiguation rules, a validator for it, a compiler for representing the rules as a finite-state transducer, and a processor which applies the rules to an Apertium input stream.

The compiler and the processor will be written in C++, based on the designs of Apertium's transfer module, and the lexical selection module. My initial intention was to use the lttoolbox API for the design and compilation of rules, and while it provides a great formalism for morphological transducers, it proved to be less flexible and expressive than I needed for writing the example rules. Since expressivity is quite important I will implement transducers in foma's formalism, or in one with equal expressivity.

Another important issue is performance and complexity of the resulting transducers, for which I will do a literature review before the beginning of the coding period, and design the implementation so that it does not blow out of proportion. Regarding minimization, Foma's api uses either Brzozowski's minimization algorithm (the same is implemented in lttoolbox), or a variant of the Hopcroft's algorithm so those will be my starting points.

The disambiguation module will be documented with use-case examples in various languages, based on my current experience with developing Constraint Grammar for Croatian.

Examples of rule formalism

The following are few examples of the proposed XML formalism, together with a rudimentary transducer implementation in the finite-state calculus of foma[1]. The transducers operate on simple patterns, and although they perform validation of the Apertium's stream format, no account was given to the robustness rule of Constraint Grammar (keeping the last remaining reading). The source code of the script can be seen here, or checked out from my folder on SVN.

A removal rule

<rule n="preposition_not_accusative">
	<lu>
		<tag n="pr"/>
		<tag n="acc"/>
	     	<remove> 
			<tag n="acc"/> <!--Removes readings where the tag is Accusative!-->
		</remove>
	</lu>
	<lu>
		<tag n="n"/>
		<tag n="loc"/>
	</lu>
</rule>

This rule matches the pattern ["LU containing the tag <pr> and the tag <acc>"] ["LU containing the tags <n> and <loc>"] and removes all readings containing the tag <acc>.

Code snippet from the foma script:

define RuleRemoveAccusative  [ [Preposition & Accusative & Locative] .o. MarkSuspicious(Contains({<acc>}))] 
       			     [ Noun & Locative ] ;

The transducer matches two LU's, with tags as described in the XML formalism, and the first match is given by composition to the MarkSuspicious transducer, which marks readings containing <acc> as pending removal.

The entire transducer applying this rule to the stream is made as the following composition.

ValidInput .o. PrepareInput .o. RuleRemoveAccusative .o. Remove .o. CleanMarks ;

Where ValidInput validates the input stream according to the Apertium stream format, PrepareInput assigns temporary tags, and the Remove and CleanMarks transducers perform the removal of readings marked for removal, and the cleanup of the temporary tags. TODO: možda slike ovih nekompliciranih

A select rule example

<rule n="noun_not_dative">
	<lu>
		<tag n="pr"/>
		<tag n="loc"/>
	</lu>
	<lu>
		<tag n="n"/>
		<tag n="loc"/>
		<tag n="dat"/>
		<select>
			<tag n="loc"/> <!--Select the readings with the locative tag!-->
		</selec>
	</lu>
</rule>

The rule matches the pattern ["LU containing <pr> and <loc>"] ["LU containing <n>, <loc> and <dat>"], and selects all readings with <loc> in the second LU.

The foma snippet for this rule is:

define RuleSelectLocative [ Preposition & Locative ] 
       			  [ [Noun & Locative & Dative] .o. MarkSelection(Contains({<loc>})) ] ;

Here the second match is given to the transducer MarkSelection, which marks all the readings containing <loc> for keeping, and all other readings for removal.

The entire rule is similarly made as the composition:

 ValidInput .o. PrepareInput .o. RuleSelectLocative .o. Remove .o. CleanMarks ;

Combination of both rules

These two rules can easily be combined in one that is more compact:

<rule n="combined_rule">
	<lu>
		<tag n="pr"/>
		<tag n="loc"/>
		<tag n="acc"/>
		<remove>
			<tag n="acc"/> <!-- Remove readings containing accusative -->
		</remove>
	</lu>
	<lu>
		<tag n="n"/>
		<tag n="loc"/>
		<tag n="dat"/>
		<select>
			<tag n="loc"/> <!-- Select readings containing locative -->
		</selec>
	</lu>
</rule>

This rule performs a similar match, and the transducers for <select> and <remove> are composed to both of the cohorts, yielding a more compact representation than the former two rule variants:

define RuleCombine  [ [Preposition & Locative & Accusative] .o. MarkSuspicious(Contains({<acc>}))]
		    [ [Noun & Locative & Dative] .o.  MarkSelection(Contains({<loc>})) ] ;

The entire rule is again a composition with auxilliary transducers

 ValidInput .o. PrepareInput .o. RuleCombine .o. Remove .o. CleanMarks ;

Comments on the rule formalism

The rule formalism will be taylored according to my current practical experience with Constraint Grammar (my master thesis is a CG for Croatian), thus containing functionality for e.g. for defining analogous structures to SETs and LISTs (similar to categories in apertium-transfer), allowing groups of tags to be iterated over and used on a single rule, allowing arbitrarily long matches etc. My main role models are LanguageTool[2], and Apertium's lexical selection tools[3], as well as Apertium's transfer module which use a similar pattern/action paradigm. The aim will not be to mimic Constraint Grammar, but to make the formalism as expressive as possible in given frames.

Work already done

Community bonding period

- written the program for the coding challenge

- started familiarising myself with lttoolbox, written a small program that composes strings and regexes into an FST

- done an implementation of example disambiguation rules in the calculus of foma

Work To do

Before the coding period:

- explore the API

- write a simple prototype, that implements a simple hardcoded rule (e.g. preposition-based case disambiguation for Serbo-Croatian)

- do a review of the literature on minimization algorithms for FST's

The coding period:

TODO

- Week 1: A thorough design of the XML formalism along with writing a validator.

- Week 2: Expressing the XML rules in finite-state calculus.

- Week 3: Expressing the XML rules in finite-state calculus.

- Deliverable #1 : A complete XML formalism for expressing finite-state disambiguation rules, along with expressing them in finite-state calculus.

- Week 4-8: Writing a compiler and a stream processor.

The syntax of the disambiguation will be based on the syntax of the lexical selection module. The funcionality will be expanded by adding a negation element (to implement funcionality analogous to barriers and negation in CG), and adding elements to match one or more LU's (analogous to scanning in CG, or the +,? and * quantifiers in regular expressions syntax).

- Deliverable #2 : The first version of the compiler and the processor.

- Week 9-12: Testing and polishing the system, writing the documentation, along with use-case examples on various languages.

For testing of the rules I will use examples of rules of varying complexity based on my current experience with Constraint Grammar. The documentation will contain use-case examples in form of tutorials, and will be based on my current work with Constraint Grammar for Croatian. The examples will be based mostly on the rules I am currently developing for Croatian, as well as on Constraint Grammar rules available for other languages.

- Deliverable #3 : The complete disambiguation system, with a compiler, a processor, and the documentation.

Non-GSoC activities

TODO:

Bio

I am an Graduate student of Computer Science and Mathematics at the Faculty of Science, University of Zagreb.

During my courses I have worked with C/C++, Python, C\#, Java, JavaScript, PHP + CSS + HTML, XML, SQL, Coq... Besides Coq, I also have a basic knowledge of functional programming through Haskell and the GF formalism. Currently I am writing my master thesis on disambiguation for the Croatian language with Constraint Grammar.

I have worked on the language pair apertium-sh-mk for the GSoC of 2011., and have been a mentor for Google Code-In 2011 for several tasks involving that and similar language pairs.

Regarding the technologies used in machine translation we I've been enrolled in courses with finite state machines, and context free grammars (implementation of a parser using yacc+flex), and machine learning.

References