Difference between revisions of "User:Srbhr/GSOC 2020 Proposal: Automatic PostEditing"

From Apertium
Jump to navigation Jump to search
Line 170: Line 170:
--[[User:Mlforcada|Mlforcada]] ([[User talk:Mlforcada|talk]]) 16:10, 25 March 2020 (CET) I miss an indication of the range of applicability: for instance, it may only work when PE(MT(S)) and MT(S) are similar.
===Language Pairs & Data===
===Language Pairs & Data===

Revision as of 15:10, 25 March 2020

Title: Automatic Post-Editing/Improving Language Pairs by Mining Post-Edits

Contact Information

Name: Saurabh Rai

IRC Nick: srbhr

Location: New Delhi, India

Time Zone: UTC+5:30 (IST)

Email: srbh077@gmail.com

Github: https://github.com/srbhr[1]

LinkedIn: https://www.linkedin.com/in/saurabh-rai-9370a0194/ [2]

Who am I?: I'm a Undergraduate Computer Science Student, in 3rd year of college from GGS Indraprastha University, New Delhi. I'm interested in Machine Learning and Natural Language Processing, and always seek to find ways to improve stuff based on them. I love talking about technology, AI, and Cyberpunk 2077.

FOSS Software I have used: My Work always involve FOSS Software and Frameworks, from Python to Tensorflow, from Ubuntu to Arch Linux, I've used many and tried to tweak the software that I use. I have tried to contribute to some of the FOSS Frameworks as well. And I have taken Part in Making some open source projects as well of my own.

Languages I know: Hindi(Native), English

Skills and Knowledge

I'm a Computer Science Student in my 3rd year of College

Languages that I know:

   1. Python
   2. C/C++
   3. Java
   4. JavaScript

Languages that I'm familiar with:

   1. Julia
   2. Bash

Machine Learning, Deep Learning, and Natural Language Processing.

   ML Frameworks that I've worked with: Tensorflow, PyTorch, Spacy, Gensim, nltk, Flair, AllenNlp, openCV etc.
   Data Visualisation and stats libraries: pandas, numpy, seaborn, plotly, matplotlib, scipy etc.
   I've taken Online as well as offline courses to hone up my skills in the field.
   Not only this I've work with projects regaridng the same and I'm currently writing a research paper on Information Retrieval and Extraction
  (Stalled due to sudden college closure coz of CORONA Virus Pandemic).

Cources that I've taken during my college time (Only Relevant to project):

   1. Mathematics (I-IV, it includes statistics, calculus(both),probability and linear algebra)
   2. Computer Programming
   3. AI
   4. Data Structures
   5. Algorithm Design and Analysis

Why is it that you are interested in Machine Translation?

Machine Translation is one such field that tells us about how tasks like translating languages and serving humanity can be done easily, and not only this, it tells about the scope of how things can be in future, I'm interested in Mahcine Translation because I'm interested in discovering ways to improve it further and with more feautures, and the project I'm interested in working with does the same. To improve machine-translation by mining post-edits, and learn based on it.

Why is it that you are interested in Apertium?

Apertium was one of the first Machine Translating Tool invented in the early 2000's in Spain. Till then and since now, it has helped a lot of people translating languages and it's the only Machine Translating Tool/Engine that works offline as well as works with Low Resource lanugages. Not only this, it's an Open Source and free software for all, and it allows learning students to reasearch and work with them. Working with the Apertium Team(Mentors, and people) is a great opportunity to learn and improve one such great tool to newer accuracy measures and feautures.Not only this, but it also provides resources to learn and research with tools like the Morphological Dictionaries, Transfer Rules, Stream-Parser which can be used to create other tools as well for eg. Automatic Post-Editing tool to create dictionary entries after learning from it. etc.

My Proposal

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

I'm interested in the Project: Automatic Post-Editing, Mining Post-Editing Dumps(Parallel Corpora) to improve Translation under the Mentor: Mikel L. Forcada (More Mentors to be added). The Project Aims to find the difference in translation by the Apertium Machine Translation, and the Human Post-Edited and takes appropriate measures and learning algorithms to define the problems/mis-translations for the same and creates the required into information that can be inserted in that Apertium language pair. These information can be either:

  • Dictionaries
  • Constraint Grammar Rules
  • Lexical Selection Rules

--Mlforcada (talk) 16:06, 25 March 2020 (CET)The proposal is not too specific when it comes to decide how you plan to use your operators to generate valid(able) linguistic data for teh language pair in question. A sketch of the algorithm would be necessary. Please study the work done by Deltamachine in 2018, and see if there is anything there you can salvage. Refer to what was done there.

Task Description

The main goal of the Project Automatic Post-Editing, Mining Post-Editing Dumps(Parallel Corpora) to improve Translation is to create automatically Dictionary Entries (Monodix, Bidix) as automatically and complete by miniming Post-Editing Dumps(Human Verified Parallel Corpora) to improve translation, and performance of an Apertium Language Pair, a long goal of this will be to automate the process of creating/enriching the Built Language Pairs that are under-incubation.

The Project Consists of two phases :

First Phase

In this phases the Data needs to gathered and converted into specific format, or Post-Editing Operators:

  • S : Source Text
  • MT(S): Machine Translation of S
  • PE(S) or PE(MT(S)): The Post-Edited Sentence (Considered to be accurate)

Now then from this a structruted data needs to be created, possibly a Pandas DataFrame (CSV, JSON) for Different Languages.

--Mlforcada (talk) 15:51, 25 March 2020 (CET) The operators would be made of substrings of S, MT(S) and PE(MT(S)). I would strongly favour (in the spirit of most of Apertium work) text-based format to store any intermediate results of your processing; this allows debugging and easy scripting around it.

Second Phase

For this phase the morphological data generated by Apertium's Stream Parser will be used to get the tags/operations of each text. And comparision of the PE and MT of S, will yield some information that can be used to improve the Apertium Language Pair by creating Dictionary, Grammar, Lexical Selection Rules.

--Mlforcada (talk) 15:53, 25 March 2020 (CET) Apertium streamparser is just a way to parse the text-file format of Apertium. Again, you should consider applying it to smaller operators.


The Rationale for this Process is described at [Rationale[3]]

  • The First Step is to get the Language Pair data and make it available in the required format(S, MT(S), PE(MT(S))).
  • Use the Data with the Appropriate Edit-Distance Algorithms(Explained Later) to find where we need to improve, and get the set of triplets where there is need

to improve.

  • Then using the streamlined data, our first approach will be to expand the sentences/word's morphological tags using Apertium's Streamparser and convert data

into Apertium's Stream format. This will reveal the operators that we are looking for to improve upon.For further easy of use we can store the whole stream or break down the words and store it in the DataFrame.

  For Example Consider the Case for English-Galician Pair:
  S: Never engage in action for the sake of reward.
  MT(S): Nunca comprometer en acción para o *sake de recompensa.
  Here sake is not translated.
  We can adapt an approach here to find the missing translation for *sake from the Parallel Corpus of Post-Edits and try to create an entry for it
  Consider another Scenario:
  S: Never engage in action for the purpose of reward.
  MT(S): Nunca comprometer en acción para o propósito de recompensa.
  Here for the same sentence we have word "sake"  and it's synonym "purpose".As "sake" doesn't have a translation pair, whereas "purpose -> propósito" does.
  We can use this data to create the entry for "sake" to Galician "ben". Take the example "For the sake of" to "Por ben de." in English Galician. 
  But this is highly recommmend that a Human Post-Edited Parallel Corpus should be available to verify the data.

--Mlforcada (talk) 15:55, 25 March 2020 (CET) Having a post-edited corpus is indeed the starting point of the approach. --Mlforcada (talk) 15:57, 25 March 2020 (CET) How would you generate the synonym-paraphrased source text?

  This example was explained by Mikel:
  S: Los marineros oteaban el horizonte.
  MT(S): The sailors *oteaban the horizon.
  PE(S): The sailors scanned the horizon.
  Here we can do scan.v -> otear.v [or scanned.v -> oteaban.v]
  A morphological guess will make it like Spanish verbs with lemmas ending in "ar" have "aban" as their imperfect 3rd person plural.
  Allowing us to build an entry from here itself.

--Mlforcada (talk) 15:55, 25 March 2020 (CET) You'd have to build a guesser. Check first on IRC as there was someone else considering doing that.


In computational linguistics and computer science, edit distance is a way of quantifying how dissimilar two strings (e.g., words) are to one another by counting the minimum number of operations required to transform one string into the other. Edit distances find applications in natural language processing, where automatic spelling correction can determine candidate corrections for a misspelled word by selecting words from a dictionary that have a low distance to the word in question.

Edit-Distance Algorithms that can be used for this task:

   Levenshtein Distance
   Damerau-Levenshtein Distance
   Jaro Distance
   Jaro-Winkler Distance
   Match Rating Approach Comparison
   Hamming Distance

--Mlforcada (talk) 16:03, 25 March 2020 (CET) Before considering all these distances, please give a reason why Levenshtein alone is not enough. Also, you could consider Snover et al.'s TER [see: http://mt-archive.info/AMTA-2006-Snover.pdf] which also allows shifts, which is kind of a generalization of Damerau-Levenshtein Giving a list of distances in a proposal without any indication of their features is superfluous. Some of them would not produce different alignments. Hamming's does not apply if strings have diferent length. Please improve.

For Token Comparision:

   Jaccard Index
   Tversky Index
   Overlap Coefficient
   Cosine Similarity, etc.

--Mlforcada (talk) 16:03, 25 March 2020 (CET)Please explain why you need to compare tokens. How would you apply cosine similarity if you've got no embeddings?

All these algorithms are available through two Python Libraries (and are optimised).

  • Jellyfish[[4]]
  • Textdistance[[5]]

--Mlforcada (talk) 16:10, 25 March 2020 (CET) I miss an indication of the range of applicability: for instance, it may only work when PE(MT(S)) and MT(S) are similar.

Language Pairs & Data

For testing of the project, I'm working on Available(X) to English or English to Available(X). Our aim is to test this on X to English and English to X as I'm fluent in English and can work with it well. But the longer goal for this is to create this process of discovering pairs Language Free. (The Algorithm for Discovering the Text).

For the Dataset: I'll be using the Available Post-Edits from OPUS Project [6]. (Note: WikiMedia Dumps is a Part of OPUS).(It's FOS Data)'

--Mlforcada (talk) 16:04, 25 March 2020 (CET) Do you have (S,MT(S),PE(MT(S))) triplets in OPUS? Give one example.

Work Plan

Community Bonding Period (4th May to 30th May)

Make myself familiar with the Tools, Language Models (Dicts, LUs, Tags etc.), The Howto of Creating Language Pairs, and Anything necessary to the Project.

Create the GitHub Wiki Roadmap and Git Kraken Globoards for the Project and Map it's timeline and share with the team.

Talk with the Mentors to get more familiar with them, and their views and improvements about the Project.

If the time Allows, start the Project Early.

Week 1-4 (1st June to 29th June)

The Aim for The First Phase

  • Get the Data for the some Linguistic Pairs.(4-5 Testing Pairs)
  • Convert the Data into the required Format and in a Data Frame. (S,MT,PE)

--Mlforcada (talk) 16:07, 25 March 2020 (CET)Please stick to text-based formats where possible.

  • Use the Edit-Distance Algorithm to find the mis-matches, missing words in Translation.
  • Using the Necessary Apertium Libraries, expand the Morphological Dictionary (and Possibly store into it's respective dataframe for ease of use later).

--Mlforcada (talk) 16:07, 25 March 2020 (CET)What do you mean by dataframe here?

  • Try to get the synonyms of words from nltk.synset to check if we have something similar available that can be used as well.
  • Document the Whole Code.

Deliverable (29th June to 3rd July): A DataFrame Consisting of the Morphological Dictionaries and LUs (Which are later to be used by Algorithm) Note: We're saving time on Edit-Distance Algorithm by using External Libraries.

Week 5-8 (3rd July to 27th July)

The Aim for The Second Phase

  • This Phase Focuses Solely on creating the Algorithm that is able to learn post-edits, get the difference between two structure and tags.
  • And work upon them to get the data from the Operators and be able to create Dictionary Pairs or Apertium Data and Document the Whole Code.
  • This is the most crucial step for the Project, hence a whole work phase of 4 weeks is dedicated to this.
  • This whole process is Explained in Task Description (Above [6.2]).

Deliverable (27th July to 31st July): Data Generated from the Streamlined Algorithm in the form of monodix, bidix, Grammars, etc.

Week 9-12 (31st July to 24th August)

The Aim for The Third Phase

  • Testing of the Data Produced by the Algorithm in Phase Two, and check the accuracy of it.

--Mlforcada (talk) 16:08, 25 March 2020 (CET)Say how

  • Improve the Algorithm based on this process.

--Mlforcada (talk) 16:08, 25 March 2020 (CET) No algorithm sketched

  • Documenting the Whole Process.
  • Testing the Algorithm onto newer Data and New Language Models
  • Trying to shift the Project's Testing of English to other Language Pairs. (This Phase is Experimental)
  • If the Time Allows, think of newer ways/ideas to improve/integrate to the Project.(This involves Discussion with the mentors)

Final Submission (24th August to 31st August): The Whole Code along with the Documentation and the Data upon which it was tested and trained.

A Description of Who and How it will Benefit the Society

This Project's sole aim is to Improve the Language Pairs, by creating Data Elements for them, from Parallel Corpus. These Language Pairs are used by the Apertium Developers to create new translational models and my Automatic Post-Editing task will provide them with the resource in for of Dictionary Elements, Grammers, Lexical Rules,which can aid the process of creating/generating Data for an existing Model Pair, henceforth enriching the dictionary and help it achieve state of the art.

For the Apertium User, this will end up as better translation tool. Making the End User more satisfied, and that's a software's sole purpose to help the end-user.

Why Google and Apertium Should Sponsor it?

The Project I'm working on is an Apertium Integration Idea, Automatic Post-Editing, that is with the help of tools, to make the Process of working with Language Pairs more easier by Automatically mining dictionary elements, or Apertium Data. This will in a long run will help enrich the existing pair of Language Tools. Apertium's Developers are sure to get the benefit of such autmation task, while working on a language pair given an Parallel Corpus.

By Funding this Project Apertium will get work done on one of it's integration idea, and explore newers automation methods, for creating data. For Google to fund this project, it'll help an Open Source Software that is serving a wide community, help Open Source Development for it's ideas and promote FOSS Development and Software in general.

And for me, I wish to stay with this community and work on other ideas as well, work on new things and to Finally be a part of PMC.

Coding Challenge

The Coding Challenge for this task is completed and was discussed with the Mentor Mikel L. Forcada. He considered the challenge a success/pass.

Code is available here Along with the Necessary Details: https://github.com/srbhr/Test-Edits

--Mlforcada (talk) 16:08, 25 March 2020 (CET) Yes.

Other Plans Besides GSOC

No Other Plans

I have no other Plans besides GSOC, and I can make sure that I will be able to give in full time (40 hrs. a week) for this. However, one situation that can arise from this can be my exams (maybe I don't know yet) being in the first-phase of the Coding Challenge, due to the COVID-19 Pandemic, and lock-downs. This is a situation that can occur but I'm not sure of this. Else, everything goes just as stated in the Work-Plan. And even in exam time I will be able to give in time for coding (20 hrs a week).