[vthommeret]

I saw a few posts discuss using the Wiktionary dump directly vs. the freeDictionary API, which is difficult to do because the raw wiki text isn't immediately usable. I actually created and open sourced a project several years ago that I never publicized that lexes and parses the Wiktionary dump:

https://github.com/vthommeret/glossterm

Specifically it can understand and execute 21 different wiki text templates (e.g. "cog", "borrow", "gloss", "prefix", "qualifier”), e.g. {{inh|es|la|gelātus}}:

https://github.com/vthommeret/glossterm/tree/master/lib/tpl

And eventually parse it into this structure, which has a list of all definitions (distinguished into nouns, adjectives, verbs, adverbs, etc...), etymology, links, and descendants for a given word:

https://github.com/vthommeret/glossterm/blob/master/lib/gt/p...

Further parts of the pipeline turned different relationships into edges that I could stick into a graph database and do certain graph queries. This allowed me to do certain queries like find French, Spanish, and English words that share a Latin root.

I ended up parallelizing this specific query using Apache Beam and then dumping the results into Firestore so they could be queried via a web app. Here's an example for the Spanish word: helado

https://cognate.app/words/es/helado

Under the "Cognates" section, it knows that it comes from the Latin root "gelatus" from which English has borrowed the word "gelato".

I originally started this project when I was learning Spanish. If you just look up the definition of helado (ice cream) it doesn't necessarily help you learn it. But I found that if I could relate it to languages I already knew (e.g. English and French), it was easier to remember. In this case helado is related to gelato, but you won't find that in e.g. Google Translate or SpanishDict.

Ultimately, I found that while the Wiktionary data is amazing, it’s also a bit of a quagmire for finding cognates. I would miss certain etymologies where you had to follow a descendant tree 2 or 3 levels deep. Or a definition would just mention a word it was related to. But if I expanded the query to include these instances, then it significantly increased the amount of non-cognates that showed up in the results.

So I created a useful set of tools (which I never wrote about until now), but I realized the end result of a web UI that showed the relationships between words would require a significant investment in data quality that likely wasn’t possible without changing Wiktionary itself / community investment.

[achempion]

Great work

I'm working on similar dictionary app and found wiktionary insanely usable as dictionary source.

Here is one more project aiming to make wiktionary data usable as json data structure: https://github.com/tatuylonen/wiktextract.

It has a link to a site https://kaikki.org/ which hosts dictionary data dumps.

[vthommeret]

Thanks! Yeah I've seen a few similar projects (particularly written Java which I wasn't excited about). That looks like a nice project. It's written in Python and says it can take from an hour to several days depending on the computer and they don't recommend running it on a Mac.

I don't have up-to-date benchmarks but my project is written in Go and everything was designed to be as highly parallel as possible, broken up into multiple pipeline steps (splitting the Wiktionary dump, lexing, parsing, resolving, etc...) with a high emphasis on performance so I would assume it's faster but would need to do a head-to-head test.