Automatic postediting at GSoC 2018

Related links

Idea description

Proposal for GSoC 2018

https://github.com/deltamachine/naive-automatic-postediting

Progress notes

Data preparation

Russian - Belarusian

• Mediawiki: 2059 sentences - source, Apertium translated and postedited by humans (only bel -> rus)
• Tatoeba: 1762 sentences: source, target and both ways Apertium translated (bel -> rus, rus -> bel)

Total amount of sentences: 3821.

Russian - Ukranian

• Tatoeba: 6463 sentences - source, target and both ways Apertium translated (ukr -> rus, rus -> ukr)
• OpenSubtitles: 2000 manually filtered and corrected source - target pairs from OpenSubtitles2018 corpora preprocessed with bicleaner + both ways Apertium translations (ukr -> rus, rus -> ukr).

Total amount of sentences: 8463.

Code refactoring

Two old scripts, learn_postedits.py and apply_postedits.py were refactored. Also now both scripts work approximately 10 times faster: now scripts collect all subsegments in one large file and translate/analyze the whole file. Instead of calling Apertium few times for every subsegment, now it is called only two times (for translating and for analyzing) for all subsegments of a sentence.

Operations extraction

There were three attempts to extract postediting operations for each language pair: with threshold = 0.8 and -m, -M = (1, 3). In fact, results are not very meaningful: the reason might lie in problems in learn_postedits.py and in the method itself (but it should be checked carefully).

New algorithm for operations extraction

Because of meaningless results of using the old algorithm, the new algorithm was created. It is based on the custom alignment. It seems that the new code will work okay on close-related languages, but I'm not sure about others. The code can be found here: https://github.com/deltamachine/naive-automatic-postediting/blob/master/new_alg/new_learn_postedits_algorithm.py and the rationale can be found here: https://github.com/deltamachine/naive-automatic-postediting/blob/master/new_alg/rationale.md.

Classifying operations

A script for classifying extracted postedits (https://github.com/deltamachine/naive-automatic-postediting/blob/master/new_alg/extract_types.py) indentifies three types of operations: potential monodix/bidix entries (when a pair doesn't have a translation for a given word), grammar mistakes (when Apertium chooses incorrect form of translated word) and other mistakes (it can be, for example, a potential lexical selection rule).

How it works:

1) It takes file with postedit triplets (s, mt, pe).

2) If here is '*' in mt, algorithm adds triplet to "potential bidix entries" list.

3) If not, the script calculates the following metric:

x = ((l - d) / l) * 100

where l = number of letters in pe and d = Levenshtein distance betweeen mt and pe.

If 50 <= x < 100, the algorithm adds triplet to "grammar mistakes" list.

4) Otherwise the algorithm checks, if mt != pe, and if not, adds triplet to "other mistakes" list.

Cleaning

The extracting postedits algorithm is not perfect and extracts a lot of garbage along with potentially good triplets. For cleaning files with postedits a following script was written: https://github.com/deltamachine/naive-automatic-postediting/blob/master/new_alg/clean_postedits.py. On the first step it tags every part of every triplet using apertium-tagger and then drops triplets which contain punctuation. It helps filter out wrong triplets like, for example, (',', '*видець', ',').

Then it calculates the same metric as in classifying step between s and mt, mt and pe, s and pe. If every result >= 30 and triplet is not from "other mistakes" list, the algorithm saves this triplet, if not - drops it. It helps filter out wrong alignment cases.