User:Khannatanmai

From Apertium
Jump to navigation Jump to search
Google Summer of Code 2019: Proposal [First Draft] [INCOMPLETE]

Anaphora Resolution

Who am I?

What open source software do you use?

Have used apertium in the past, ubuntu, firefox, vlc.

What are your professional interests?

I’m currently studying NLP and I have a particular interest in Linguistics

What are your hobbies?

I love singing, reading, debating.

What is your skill set?

Creating NLP tools, Thorough Linguistic Analysis, Writing clean and understandable code

What do you want to get out of GSoC?

I’ve studied about apertium and it’s amazing to me that I get an opportunity to work with them. NLP is what I want to do in life and working with a team to develop tools that actual people use will be invaluable experience that classes simply cannot match.

Personal Details

Name: Tanmai Khanna

E-mail address: khanna.tanmai@gmail.com

Other information that may be useful to contact you (e.g. IRC):

IRC: khannatanmai

GitHub: khannatanmai

Why is it that you are interested in Apertium? / Why am I interested in Machine Translation?

Apertium is an Open Source Rule-based MT system. Each part of their mission statement interests me and excites me to be working with them. I have been part of the Machine Translation lab in my college and it interests me because it’s a huge problem and is often called NLP-Complete by my professors, i.e. it uses most of the tools NLP has to offer and hence if one learns to do good MT they learn most of Natural Language Processing. While Neural Networks, Deep Learning is the fad these days, they only work for resource rich languages and that’s why I feel a project which is rule based and open source really helps the community with language pairs that our resource poor and gives them free translations for their needs.

Which of the published tasks are you interested in? What do you plan to do?

Anaphora Resolution - Currently uses default male. I don’t plan to make a perfect anaphora resolution in 3 months, but I’m confident that I can make one which works significantly better than the default male and it can increase the fluency and intelligibility of output significantly.

I feel that this project affects almost all language pairs in apertium and hence it affects almost everyone using the tool. Pronouns are present in a lot of languages and with gendered pronouns, singular, plural, etc., we need to find out what they refer to. Why is this important?

A sentence like “The group agreed to release his mission statement” is grammatically incoherent and an incorrect pronoun will more often than not confuse people in more complex sentences. What puts people off Machine Translation is the lack of fluency, and I feel this is an important contribution to fluency and will definitely generate more fluent sentences leading to more trust in this tool.

IDEAS

Use elimination to figure out which noun. Like heads of chunks can be referred to. Can we use semantics or not? Should it be language independent? Very less or no dependence on external tools, like wordnet, framenet, etc. For basic sentences, it will definitely help (with female subjects) When translating from languages with gender in pronouns, retain that info. Can be used for anaphora resolution in target language. If it knows about animacy, in the coding challenge I can give accurate result.

QUESTIONS

Is this tool supposed to be language independent? For eg., Anaphora Resolution of English can use certain tools which capture semantics to perform better. If it is language independent then we can’t depend on external tools, which would need solutions which use only the information available out of biltrans. Suddenly stopping the default male system and putting another one could give worse results. Going step by step makes more sense. For eg., sentences with just one noun and if that noun is female, the later pronoun has to be female. There we should use female anaphor. Eg. “La chica comió su manzana” translates to “The girl ate his apple”.

Even without touching the default male system, if the only candidate antecedent is female, the anaphor should be female. Apart from this, I feel a good method might be to use elimination to figure out the best antecedent for an anaphor. Biltrans does seem to have some element of animacy. We can use that to eliminate. Also, if a chunk exists, such as “Groups of the Parliament”, the head of the chunk is “groups” and it is more likely that an anaphor refers to the head of a chunk.

STUFF TO DO

  • Understand the system, Get familiar with the files that I need to modify
  • Formalise the problem, limit the scope of anaphora resolution (To Anaphora needed for MT)
  • May include general anaphora (Need to decide on scope [for gisting])
  • Annotation of anaphora for evaluation
  • Flowchart of proposed system and Pseudocode
  • Implement a scoring system for antecedent indicators [work for Target Language:English for now]
  • Decide on a definite context window
  • Implement basic anaphora outside the pipeline (python)
  • Implement basic transfer rules to see if final system will work
  • A basic prototype of final system ready
  • Port the basic prototype to C++ (All further coding to be done in C++)
  • TEST the system
  • Document the outline
  • Implement system to work out all possible antecedents
  • Add ability to give antecedents a score
  • TEST basic sentences with single antecedents, Test the pipeline
  • Deliverable #1: Anaphora Resolution for single antecedents, with transfer rules [The full pipeline]
  • Implement Antecedent Indicators:
  • Implement Boosting Indicators
  • Implement Impeding Indicators
  • Implement tie breaking systems
  • Implement fallback for anaphora (in case of too many antecedents or not past certainty threshold)
  • Code to remember antecedents for a certain window
  • TEST Scoring System
  • Implement transfer rules to deal with new additions
  • Evaluate current system and produce precision and recall
  • Implement Expectation-Maximization Algorithm
  • Use Monolingual corpus to get probabilities of anaphora
  • Implement choosing anaphora with max probability:
  • If Scoring System has a tie
  • As an independent system
  • Compare the effectiveness of the above two possibilities
  • Test EM Algorithm and implemented system
  • Evaluate if addition of EM gives us significant benefits
  • Deliverable #2: Anaphora Resolution with antecedent scores, fallback mechanism, EM algorithm
  • Document Antecedent Indicators, Scoring System, EM Algorithm, Fallback
  • Insert into Apertium pipeline
  • Implement code to accept input in chunks and process it
  • Output with anaphora attached
  • EXTENSIVELY TEST final system with multiple pairs, see what needs to be changed for pairs
  • TEST for backwards compatibility and ensure it
  • Project Completed

Additional Information

  • Agreement: Different for different languages?
  • Agreement rules in Arabic, however, are different. For instance, a set of non- human items (animals, plants, objects) is referred to by a singular feminine pronoun.
  • Since Arabic is an agglutinative language, the pronouns may appear as suffixes of verbs, nouns (e.g., in the case of possessive pronouns) and preposi- tions.
  • Antecedent Indicators:
  • Boosting Indicators[Scoring different for different languages]
  • First NPs
  • Indicating Verbs
  • Lexical Reiteration
  • Section Heading Preference
  • Collocation Pattern Preference
  • Immediate reference (if it is pronoun then its reference)
  • Sequential Instructions
  • Impeding Indicators
  • Indefiniteness
  • Prepositional NPs

Reasons why Google and Apertium should sponsor it:

I feel that this project affects almost all language pairs in apertium and hence it affects almost everyone using Apertium. A decent anaphora resolution will give the output an important boost in it’s fluency and intelligibility, not just for one language, but all of them. It’s a project which has promising future prospects - apart from the fact that language specific features can be added to improve it, we’ll be doing anaphora resolution even for languages which don’t need it to pick the correct pronoun. Doing this will enable Apertium to do gisting translation, which is an important tool and anaphora resolution is an essential cog in that wheel.

A description of how and who it will benefit in society:

It will definitely benefit most users of Apertium and hopefully will attract more people to the tool. I’m from India and for a lot of our languages we don’t have the data to create reliable Neural MT systems. Similarly, for all resource poor languages, Apertium provides an easy and reliable MT system for their needs. That’s how Apertium benefits society already.

I feel that currently what repels people from Machine Translation is unintelligible outputs and too much post editing which makes it useless for them. While Apertium aims to make minimal errors, as of now it selects a default male pronoun and that leads to several unintelligible outputs. Fixing that and making the system more fluent and intelligible overall should definitely attract people to using Machine Translation and help them to reduce costs of time and money.

Work plan

  • Week 1:
  • Week 2:
  • Week 3:
  • Week 4:
  • Deliverable #1
  • Week 5:
  • Week 6:
  • Week 7:
  • Week 8:
  • Deliverable #2
  • Week 9:
  • Week 10:
  • Week 11:
  • Week 12:
  • Project completed

Include time needed to think, to program, to document and to disseminate.

If you are intending to disseminate to a conference, which conference are you intending to submit to. Make sure to factor in time taken to run any experiments/evaluations and write them up in your work plan.

List your skills and give evidence of your qualifications. Tell us what is your current field of study, major, etc. Convince us that you can do the work.

List any non-Summer-of-Code plans you have for the Summer, especially employment, if you are applying for internships, and class-taking. Be specific about schedules and time commitments. we would like to be sure you have at least 30 free hours a week to develop for our project.

Non-Summer-Of-Code Plans

I will have a 3 month vacation from May to July so will heave no other commitments in that period and will be dedicated to GSoC full time (40 hours?) I am going on a short trip to London from 15 May to 22 May but I will have internet there and will be working a little less than normal but will catch up.

Coding Challenge

https://github.com/khannatanmai/apertium

Successfully completed the coding challenge:

  • Wrote code to do basic anaphora resolution (last seen noun), input taken as a stream (byte by byte)
  • Wrote code to do basic anaphora resolution (last seen noun), input taken as a stream (as a chunk)
  • Compared speed-ups and wrote a report