Update week 8 to 14 April 2024:

• Went over airflow and research dataset repos
• Sketched an overview of our current code base workflow and a few research airflow repos.

The exploratory part of link-recommendation for add-a-link is done.

Update 1/4/2024 - 7/4/2024:

• Update MR according to comments
• Run an experiment with 100k cap of samples on ALL wikis.
• Start running another experiment with 1M cap of samples on ALL wikis.
• Discuss Airflow with Fabian.

Update 25/3/2024 - 31/3/2024:

• Modify and push code. Collect all wikis data.
• Train a model on all wikis
• Discuss possible pipeline solutions

Update 18/3/2024 - 20/3/2024:

• Clean and upload main and secondary wiki results. Add single language result for comparison.
• Prepare and run script to gather data for all languages (anchor dicts etc)
• Clean code to push as an intermediate stage of language agnostic modeling

Update 11/3/2024 - 17/3/2024:

• Scale the model to 50 languages: Run pipeline for 50 languages and train a model with max 100k samples per language. Used fall back chains to select languages at the center.
• Test on a different set of 50 languages (randomly chosen) in a 0-shot manner and compare performance.

Update 4/3/2024 - 10/3/2024:

• Finished grid search and stored best fit model
• Add mwtokenizer in one more place, fix code to accommodate wiki_db feature, fix label encoder, push draft MR

Update 26/2/2024 - 3/3/2024:

• Trained a combined model with all data and with stratified split
• Hyperparameter tuning in progress

Update 19/2/2024 - 25/2/2024:

• Replace w2v with outlink embedding, created baseline and ran 11 test wikis, MR sent
• Training and evaluating a combined model with all test language wikis
  • with and without wiki_db feature

Update 29/01/2024 - 04/02/2024:

• Understanding the feature generation and model training component of link-recommendation model
• Tested language performance of several models by changing the model's language. Tested multilingual models 2 languages at a time.

Update week 22/1/2024 - 28/1/2024::

• Fixed regex that was causing a lot of the models to have low-recall
• MR sent

Results of evaluations after solving this ticket can be found here: https://meta.wikimedia.org/wiki/Research:Improving_multilingual_support_for_link_recommendation_model_for_add-a-link_task#Results.

Update week 15/1/2024 - 21/1/2024::

• MR sent to fix unicode errors. Multiple languages tested.
• Tested all previously failed languages. wikipedia2vec==2.0.0 introduces a new IndexError that occurs in several languages.
  • Reverted to 2 venvs. This time conda has w2v==2.0.0 for jawiki and fywiki. venv has w2v==1.0.5 for rest of the languages.
  • Sent MR4

Update week 8/1/2024 - 14/1/2024:

• Test and fix jawiki error by adding required dependencies.
• Attempt to fix Unicode errors in zhwiki and fywiki (using different version of Wikipedia2Vec)

Update week 1/1/2024 - 7/1/2024:

• mwtokenizer MR merged, new version released
• link-recommendation MR updated and refactored to integrate new mwtokenizer
• Ran non WS languages ad some previously Failed languages. There were some improvements. More debugging required.
• MR merged

Update week 19/12/2023 - 24/12/2023:

• Make changes in mwtokenizer
  • replace ▁ with " " in the tokenizer
  • separate punctuation from tokens.
  • Sent MR
• In progress:
  • Use the updated mwtokenizer to improve link-recommendation.
  • Refactor and consolidate ngram functions in link-recommendation code.

Update week 11/12/2023 - 17/12/2023:

• Fix sentence tokenization errors in link-recommendation. Send MR. Improves bowiki, but no improvement in mywiki. WS languages remain same.
• Some analysis into the cause of the issues above.

Update week 27/11/2023 - 3/12/2023:

• mwtokenizer issues resolved. MR merged.
• Read through link recommendation docs
• Pull code, set up dev env, run code for test wikis.
• Some errors reported and fixed (T352525)

Update week 20/11/2023 - 26/11/2023:

• Revised MR for issue 38. Merged after few iterations.
• Working on issue 32. Pushed.

Update week 13/11/2023 - 19/11/2023:

• Discussed Issue 32 and 38.
• Pushed code for Issue 38
• Make diff for 38
• Create new issues for edge cases found while solving 38

Update week 6/11/2023 - 12/11/2023:

• Issue 37 solved and merged.
• Start looking at Issue 38 and 32

Update week 30/10/2023 - 5/11/2023:

• Working on Issue 37. Dissecting regex.

Update week 23/10/2023 - 29/10/2023:

• Done finding sentence terminator symbols using the sentence per paragraph method. Also listed more symbols from Terminal_punctuation. See DOC for details. MR 21 sent, merged.
• Discussed Issue 37 and 9. MR merged for Issue 9.
• Started Issue 37

Update week 16/10/2023 - 22/10/2023:

• Done Issue 41 to move Cree off of non-WS language list
• Worked on Issue 40: Find list of languages where most paragraphs have 1 sentence, analyze few random wiki pages, detect missing sentence ending punctuation, if any.

Update week 9/10/2023 - 15/10/2023:

• Working on updating sentence tokenizer for non white space languages

Update week 2/10/2023 - 8/10/2023:

• 1st MR merged
• Discussed how evaluation was set up and next steps.
• Researched and took notes for non-whitespace separated languages. Doc.
• Wikimedia Connect!!

Update week 25/09/2023 - 1/10/2023:

Thanks @colewhite. I'm all set!

Update week 18/09/2023 - 24/09/2023:

• Read onboarding docs to familiarize with mwtokenizer and add-a-link tasks: link-recommendation-2023_work-plan
• Going over mwtokenizer code base
• Issues setting up:
  • Don’t have kinit access yet
  • Cannot access Jupyter as well
  • Issues setting up venv in personal pc (couldn’t solve)
  • Created venv and error installing in stat1008. [resolved. needed proxy]

I am getting this error when I kinit
kinit: Client 'akhatun@WIKIMEDIA' not found in Kerberos database while getting initial credentials
Am I supposed to get a temporary password though email?

Week 26/6/23 - 2/7/23 Update:

• Listed and analyzed redirects in Wiktionary.

Week 19/6/23 - 25/6/23 Update:

Week 12/6/23 - 18/6/23 Update:

• Finished report (Images not added yet)
• MR11 sent for readme and figures

Week 5/6/23 - 11/6/23 Update:

• Fixed template parsing error, MR10 got merged.
• Started working on report

Week 29/5/23 - 4/6/23 Update:

• Fixed template parsing to accommodate the use of lang param in template
• Parsed and saved all language wiktionary misspelling.
• Did some analysis, de-duplication, and saved all_wiki combined wiktionary misspellings.
• More errors found in template parsing (named params occur before un-named params causing incorrect parsing)

Week 22/5/23 - 28/5/23 Update:

• Updated MR9 with summary
• Created Issue 14. Changed wiktionary parser script to make it work with all languages. Need to figure out some changes in template params.

Week 15/5/23 - 21/5/23 Update:

• Checked example use of misspelling of templates in all the collected 16 languages. All languages look similar to enwiktionary except trwiktionary (small change) and viwiktionary

Week 8/5/23 - 14/5/23 Update:

• Created Issues 12 and 13. Started working on them: identify misspelling of templates in other languages and find usage of these templates. The templates would be collected from Q50368067, misspelling of named templates in other languages, and their redirects.

Week 1/5/23 - 7/5/23 Update:

• Incorporated feedback and had MR7 merged (refactor repo)
• Analysis done on extracted misspellings, sent MR8. Based on feedback, some more analysis done.

Week 24/4/23 - 30/4/23 Update:

• Incorporate Isaacs feedback for MR5 and 6. All MRs merged after some editing and discussion.
• Created MR7 to Refactored repo
• extracted misspelling from all language wikipedias.
  • Todo: analysis on extracted data

Week 10/4/23 - 16/4/23 Update:

• Add info on language detection (language, confidence, text sent to model). Analyze examples.
• Add proxy tables: tables that were not detected by mwparserfromhell.
• Separate cell data of tables: each cell in table is now a node. Stuck with cell data/paragraph text to send to model.

Week 3/4/23 - 9/4/23 Update:

• Pushed revised code that includes all additional formatting as a list (as discussed).
• Fixed quotations detected. Added fasttext language detection.
• Analysed collected misspellings from context. Some work need to be done to increase precision of detected language.

Week 27/3/23 - 2/4/23 Update:

• Apply additional filter information to extracted misspellings: Capitalization, word length, part of a list item, inside of quotations (in any language)
• Still need to figure out the data's structure and add fasttext detected language information

Week 20/3/23 - 26/3/23 Update:

Week 13/3/23 - 19/3/23 Update:

Week 6/3/23 - 12/3/23 Update:

• Compared collected en and fr misspellings with AutowikiBrowser Typo list. Merge requested. Summary here
• Started working on extracting wikipedia text to find the ratio of misspellings

Week 27/2/23 - 5/3/23 Update:

• Address comments for Issue #5
  • Parse sections line by line, consider templates in # items (numbered list)
  • Count the number of definitions by # count, excluding ## #: #; and #*
  • Also change the data format a bit to make it more readable
• To address Issue 6: get list of misspellings from another Language and compare the collected lists to existing approaches
  • collected bnwiktionary templates. It does not have much Bangla words. Its the same as present in enwiktionary. Will work with existing collected Spanish misspellings instead.
  • for English, compared collected list with enwiki Lists_of_common_misspellings

Week 20/2/23 - 26/2/23 Update:

Week 13/2/23 - 19/2/23 Update:

Week 6/2/23 - 12/2/23 Update:

• Set up jupyter notebook (fix issues with getting spark3)
• Get list of enwiktionary pages that use missplelling_of template using the following tables:
  • mediawiki_templatelinks, mediawiki_linktarget, mediawiki_wikitext_current
• Parsed enwiktionary pages to get heading name (typically POS: Noun, Adj, etc), language of misspelling, and the correct spelling from the template
• Some analysis on parsed wikis to get languauge and heading distribution

Thank you, accessed!

Week 1/2/23 - 5/2/23 Update:

• Caught up on previous work on copy editing both in research team and growth team
• Learned about templates in Wiktionary in different langauges and the possible categories they may be in

In T303831#8063021, @EBernhardson wrote:

In terms of the exact code causing this, spark is terrible at telling us exactly where but trying to infer from the SparkUI output i think it's this join:

def getTopSubgraphItems(topSubgraphs: DataFrame): DataFrame = {
  wikidataTriples
    .filter(s"predicate='<$p31>'")
    .selectExpr("object as subgraph", "subject as item")
    .join(topSubgraphs.select("subgraph"), Seq("subgraph"), "right")

Update:
I tested a few options in the statbox, I am not sure how much this will represent the prod env, but here goes:

In T303831#8058159, @EBernhardson wrote:

the airflow patch is deployed but i only turned on *_init dags and subgraph_mapping_weekly today (ran out of time, will do rest tomorrow).

subgraph_mapping_weekly failed the first time through. I updated executor memory from 8g to 12g but the second execution is still failing. something is quite unbalanced about the topSubgraphItems, of the 8 shards they have inputs varying from 100MB to 450MB giving executions times of ~30s on the small ones and ~8m before the final one fails.

Not specifically related to this patch, but i wonder if we could change up the SparkUtils.saveTables method to somehow take parameters in the path to specify coalesce vs repartition and the number of partitions to save by, so we only have to update the airflow invocation and not the jar as well to test variations there.

The analysis is done here (for Q-ids): Wikidata_Item_ORES_Score_Analysis

In T288262#7629267, @Lydia_Pintscher wrote:

@AKhatun_WMF: You mention on the wiki that some Items don't have an ORES score. All Items should have one 😬 Do you have an example of one that does not?

In T288262#7628599, @MPhamWMF wrote:

@AKhatun_WMF , sorry, it's been a while since I wrote this, but I think what I meant when I wrote the question about "optimal separation" is given some distribution of ORES scores (e.g. a normal distribution), is it clear what the threshold is for what qualifies as a "high" vs "low" score: e.g. anything over .75 is a high score. But that's assuming the scores are continuous. I guess it's moot if they're binary (I don't actually know).

If this isn't a sensible way of thinking about the issue, let me know if there's a better way.

@MPhamWMF Hi, could you please clarify the question Is there an optimal separation between high/low ORES scores?. Separation in what respect? To my mind comes the separation of items with respect to the subgraph it is part of.

@ACraze Indeed! I was confusing the models for revision (item quality) with edits (damaging/good faith). The latest revision is all I will need. Thank you!

Counts of queries and triples for astronomical objects were done here: Wikidata_Subgraph_Query_Analysis, along with the top ~300 large subgraphs.
For the specific case of Astronomical objects (and only astronomical objects), a list of all its subclasses was obtained and manually inspected for relevance to astronomical objects. The subclass list also consists of subclasses of subclasses and so on.

Details can be found here: Wikidata_Subgraph_Query_Analysis