in which the machine tries to design a solution for synonyms

This is a genuinely^°hard word interesting linguistic^°hard word modeling problem. After reviewing^°hard word the current data model, here are my thoughts.

The Core^°hard word Problem

Each Lemma^°hard word currently represents^°hard word one concept^°hard word anchored^°hard word to an English^°hard word label^°hard word. "hat" and "cap^°hard word" are two separate lemmas^°hard word, each with one translation^°hard word per^°hard word language. But languages carve up semantic^°hard word space differently:

English^°hard word: "hat" (broad brim^°hard word) vs^°hard word "cap^°hard word" (fitted^°hard word, no brim^°hard word) — distinct^°hard word

Lithuanian: "skrybel^°hard wordė" (wide brim^°hard word) vs^°hard word "kepur^°hard wordė" (general headwear^°hard word/cap^°hard word) — different boundary^°hard word

Chinese^°hard word: "帽子" covers nearly all headwear^°hard word generically^°hard word

French^°hard word: "chapeau^°hard word" (hat, broad) vs^°hard word "casquette^°hard word" (baseball cap^°hard word) — yet^°hard word another split^°hard word

The current model has no way to say "these two lemmas^°hard word occupy^°hard word overlapping^°hard word semantic^°hard word territory^°hard word, and in language X they might collapse into one word."

1. Synonym^°hard word Clusters^°hard word via^°hard word Existing Relation Groups (Minimal^°hard word Change)

Add a new relation_type = "synonym^°hard word_cluster^°hard word" to the existing LemmaRelationGroup^°hard word / LemmaRelationMember^°hard word system. Group "hat" and "cap^°hard word" together under concept^°hard word_label^°hard word "headwear^°hard word".

Pros^°hard word: Almost no schema^°hard word change. Reuses^°hard word existing infrastructure^°hard word and release^°hard word format^°hard word (lemma^°hard word_relations/synonym^°hard word_cluster^°hard word/headwear^°hard word.jsonl^°hard word).

Cons^°hard word: Too blunt^°hard word. It says "these are related^°hard word" but doesn't capture^°hard word how — it doesn't tell the app what to do in Lithuanian vs^°hard word Chinese^°hard word. A cluster^°hard word of 5 near-synonyms^°hard word doesn't tell you which ones merge^°hard word in which languages.

2. Per-Language^°hard word Acceptable^°hard word Alternatives^°hard word on Translations^°hard word

Add an alternatives^°hard word field (JSON^°hard word list) to LemmaTranslation^°hard word. For "hat" in Chinese^°hard word, the primary^°hard word translation^°hard word is "帽子", and for "cap^°hard word" in Chinese^°hard word, the primary^°hard word translation^°hard word is also "帽子" (or maybe "鸭舌帽"). The alternatives^°hard word field would note that "帽子" is acceptable^°hard word for both.

More concretely, a new table like:

TranslationAlternative^°hard word:

lemma^°hard word_id^°hard word → FK^°hard word to Lemma^°hard word

language_code^°hard word → str^°hard word

alternative^°hard word_text^°hard word → str^°hard word (the acceptable^°hard word alternative^°hard word translation^°hard word)

source_lemma^°hard word_id^°hard word → FK^°hard word to Lemma^°hard word (which lemma^°hard word "owns" this word natively^°hard word)

notes → str^°hard word (why it's acceptable^°hard word)

So for "hat" (N09_X) in Chinese^°hard word, you'd have: primary^°hard word = "帽子", no alternatives^°hard word needed (it's the generic^°hard word). For "cap^°hard word" (N09_Y) in Chinese^°hard word, primary^°hard word = "鸭舌帽", alternative^°hard word = "帽子" (source: hat's^°hard word lemma^°hard word).

Pros^°hard word: Directly solves^°hard word the Trakaido^°hard word use case — the app knows which answers to accept. Per-language^°hard word granularity^°hard word. Works well for quiz^°hard word/testing scenarios^°hard word.

Cons^°hard word: Doesn't capture^°hard word the why — it's a flat list of "also accept this." Maintenance^°hard word burden^°hard word grows combinatorially^°hard word. Doesn't help with teaching (explaining the distinction^°hard word to learners^°hard word).

3. Semantic^°hard word Overlap^°hard word Groups with Language-Specific^°hard word Merge^°hard word Flags

A hybrid^°hard word: create synonym^°hard word clusters^°hard word (idea 1), but annotate^°hard word per^°hard word language whether the cluster's^°hard word members^°hard word are distinct^°hard word or merged^°hard word in that language.

SynonymCluster^°hard word:

id^°hard word, concept^°hard word_label^°hard word ("headwear^°hard word")

SynonymClusterMember^°hard word:

cluster^°hard word_id^°hard word, lemma^°hard word_id^°hard word

SynonymClusterLanguageInfo^°hard word:

cluster^°hard word_id^°hard word

language_code^°hard word

merge^°hard word_type → "distinct^°hard word" | "merged^°hard word" | "partial^°hard word"

dominant^°hard word_lemma^°hard word_id^°hard word → if merged^°hard word, which lemma's^°hard word translation^°hard word covers the group

notes → free text^°hard word ("kepur^°hard wordė covers both unless^°hard word brimmed^°hard word")

For the headwear^°hard word cluster^°hard word in Chinese^°hard word: merge^°hard word_type = "merged^°hard word", dominant^°hard word_lemma^°hard word_id^°hard word = hat (because 帽子 covers everything). In English^°hard word: merge^°hard word_type = "distinct^°hard word". In Lithuanian: merge^°hard word_type = "partial^°hard word" with a note.

Pros^°hard word: Captures^°hard word the actual^°hard word linguistic^°hard word reality^°hard word. The app can use merge^°hard word_type to decide behavior (skip testing distinctions^°hard word in languages where they're merged^°hard word, accept either answer for "partial^°hard word"). Useful^°hard word for pedagogy^°hard word — you can teach learners^°hard word "in Chinese^°hard word, there's one word for all of these."

Cons^°hard word: More complex schema^°hard word. "partial^°hard word" is vague^°hard word and needs the notes field to be useful^°hard word. Requires^°hard word human curation^°hard word per^°hard word language per^°hard word cluster^°hard word.

4. Concept^°hard word Hierarchy^°hard word with Language-Specific^°hard word Lexicalization^°hard word Levels

Model a hypernym^°hard word/hyponym^°hard word tree:

headwear^°hard word

├── hat (brimmed^°hard word headwear^°hard word)

│ ├── sun hat

│ └── top hat

├── cap^°hard word (fitted^°hard word, no brim^°hard word)

│ ├── baseball cap^°hard word

│ └── beanie^°hard word

└── helmet

Then^°hard word annotate^°hard word per^°hard word language which level has a common word. Chinese^°hard word lexicalizes^°hard word at "headwear^°hard word" level (帽子). English^°hard word lexicalizes^°hard word at "hat"/"cap^°hard word" level. Lithuanian has its own split^°hard word point.

Pros^°hard word: Linguistically^°hard word the most accurate^°hard word model. Powerful for curriculum^°hard word design — teach the generic^°hard word word first, then^°hard word the specific^°hard word ones at higher difficulty^°hard word. Naturally handles the "how many words does this language need for this concept^°hard word space" question.

Cons^°hard word: Most complex to build and maintain^°hard word. Over-engineered^°hard word for many cases (do you really need a tree for "big" vs^°hard word "large"?). The hierarchy^°hard word itself may be English-biased^°hard word — who says "headwear^°hard word" is the natural top node^°hard word?

5. Pragmatic^°hard word Recommendation^°hard word: Cluster^°hard word + Alternatives^°hard word (Ideas 1+2 Combined^°hard word)

My actual^°hard word recommendation^°hard word is to combine^°hard word the simplest^°hard word version^°hard word of two ideas:

A) Synonym^°hard word clusters^°hard word using the existing relation group system (relation_type = "synonym^°hard word_cluster^°hard word"). This groups related^°hard word lemmas^°hard word and is cheap to implement^°hard word. It lives in lemma^°hard word_relations/synonym^°hard word_cluster^°hard word/ in the release^°hard word format^°hard word.

B) Acceptable^°hard word alternatives^°hard word as a lightweight^°hard word annotation^°hard word on LemmaTranslation^°hard word (a JSON^°hard word field or a small sibling^°hard word table). This gives the app the per-language^°hard word "accept this answer too" data it needs.

The cluster^°hard word provides^°hard word the human/editorial^°hard word grouping for curation^°hard word. The alternatives^°hard word provide^°hard word the app-facing^°hard word behavior. They're populated^°hard word semi-independently^°hard word — you might have a cluster^°hard word of 4 words but only 2 of them share a translation^°hard word in Lithuanian.