/analyze — Romanian morphological analysis

Given Romanian text, returns for every token in context: its lemma, part of speech, and morphological features.

The distinction that matters is in context. A dictionary can tell you that sare is either the noun sare ("salt") or the third-person present of sări ("to jump"). It cannot tell you which one you are looking at. /analyze can, because a language model resolves the reading from the sentence and the dictionary then supplies the lemma:

• Pune sare în mâncare. → sare, NOUN
• Pisica sare pe masă. → sări, VERB

Quick start

curl -X POST https://api.lexicro.com/analyze \
  -H 'Content-Type: application/json' \
  -H 'X-API-Key: lxr_your_key_here' \
  -d '{"text": "Elevii care s-au înscris nu vor fi afectați."}'

{
  "model_version": "phase2-baseline-0.1",
  "truncated": false,
  "sentences": [
    {
      "tokens": [
        {
          "form": "Elevii",
          "lemma": "elev",
          "upos": "NOUN",
          "feats": {
            "Case": "Acc,Nom",
            "Definite": "Def",
            "Gender": "Masc",
            "Number": "Plur"
          },
          "source": "lexicon"
        }
      ]
    }
  ]
}

Authentication

Pass your key in the X-API-Key header. Without one you get an anonymous allowance of 10 requests/day per IP address — enough to try the API, not enough to build on.

Keys look like lxr_ followed by 43 URL-safe characters. They are shown once, at issue time, and are not recoverable — store yours in a password manager. If you lose it, we revoke and reissue.

To get a key, fill in this short form — it takes about a minute. You'll get a confirmation email; click the link in it and your key is generated instantly and shown once on screen.

Didn't get the email? You can re-send it.

Rate limits

Exceeding the limit returns 429. For keyed requests the message names your key's prefix (the first 12 characters), so you can quote it in support requests without sending the key itself; anonymous requests are identified by IP.

Request

POST /analyze
Content-Type: application/json
X-API-Key: lxr_...

Send UTF-8. Romanian diacritics (ă â î ș ț) must survive the round trip — if they arrive mangled, so will the analysis.

Response

Token fields

The source field

This is provenance, not confidence: it tells you which subsystem produced the lemma, nothing about whether that particular lemma is right.

About 73% of tokens in ordinary text come back as lexicon.

Read the accuracy column as a description of each class, not of any one token. It is the accuracy measured across every token that landed in that class on the RRT test split — a population figure, not a per-token score, and /analyze has no way to tell you where in that population a given token sits. This matters most for model: a full stop and a genuine neologism both come back as model, and the class average blends a token that is trivially always right (punctuation) with tokens that are genuinely hard (unseen words). The 93.3% says nothing about which kind this token was.

So source is not a trust signal, and the table above is not one either — do not threshold on it. If you want to know whether a specific word was ambiguous, read candidates below: it is the one field that speaks about an individual token rather than a class of them. No general per-token confidence figure is published yet.

The candidates field

Present on a token only when the lexicon lists more than one reading for that form. Each entry is a {lemma, upos, feats} the form could have been:

{
  "form": "era",
  "lemma": "fi",
  "upos": "AUX",
  "feats": {"Mood": "Ind", "Tense": "Imp", "Number": "Sing", "Person": "3"},
  "source": "lexicon",
  "candidates": [
    {
      "lemma": "eră",
      "upos": "NOUN",
      "feats": {"Case": "Acc,Nom", "Definite": "Def", "Gender": "Fem", "Number": "Sing"}
    },
    {
      "lemma": "fi",
      "upos": "VERB",
      "feats": {"Mood": "Ind", "Number": "Sing", "Person": "3", "Tense": "Imp", "VerbForm": "Fin"}
    }
  ]
}

The reading chosen in context is the one in the token's own lemma / upos / feats. candidates is the dictionary's inventory beside it — which is exactly why this endpoint exists rather than a dictionary lookup.

Look closely and the two don't match: the token's own upos is AUX, but the lexicon's entry for the copula, sitting right there in candidates, is tagged VERB. That is not a bug in this example — MULTEXT-East doesn't distinguish auxiliary from main verb the way UD does, so the lexicon has no AUX entry for fi to offer. source is still "lexicon": the lemma fi came from that VERB entry, matched to the model's AUX prediction by tag class rather than by an exact tag match. The lemma is a lexicon fact either way; the tag shown in candidates is just the lexicon's own, in the lexicon's own convention. This is the concrete case the caveat below is about, not a hypothetical one.

Two things to know before you rely on it:

• The chosen reading is not guaranteed to appear in candidates. The lexicon and the treebank disagree on annotation conventions for about 6% of tokens — auxiliary versus main verb, participles as adjectives, determiner versus pronoun — so the model legitimately returns readings the lexicon never offered. Treat candidates as "what the dictionary knows", not as a closed set the answer was drawn from.
• Absence means one of two things, and source tells you which. source: "lexicon" with no candidates means the form is in the lexicon and unambiguous. source: "suffix" or "model" means the form is not in the lexicon at all, so there is no inventory to report.

About 35.68% of tokens in running text carry a candidates list. That is much higher than the 4.49% of lexicon forms that are lemma-ambiguous, mentioned below, and the two are not in conflict: ambiguity concentrates in common words, and candidates reports any alternative reading — including a different part of speech or different features for the same lemma — not only a different lemma.

Worked example: why context matters

{"text": "Pune sare în mâncare."}

sare → lemma sare, NOUN — the substance.

{"text": "Pisica sare pe masă."}

sare → lemma sări, VERB — the action.

Same four letters, two different lemmas. No lexicon can choose between them; the choice depends on the sentence. This is the whole reason the endpoint exists rather than a dictionary lookup.

Forms like this are not rare: 15,822 of the lexicon's 352,004 forms (4.49%) are lemma-ambiguous — the correct lemma genuinely depends on the reading, which is exactly what the model resolves before the dictionary is consulted.

Romanian clitics are handled too: s-au is split into s- and au, and each is analysed separately — matching the convention used by Romanian treebanks.

Accuracy

Measured on the UD Romanian RRT test split (16,311 tokens, gold tokenisation):

Sliced by dictionary coverage:

Two honest caveats:

These are gold-token evaluations — the standard convention for Universal Dependencies results, which is what makes them comparable to published work. The API tokenises raw text itself, so end-to-end accuracy on arbitrary input is a different and unmeasured quantity.

RRT's lemmas and XPOS are automatically produced, per the treebank's own metadata; UPOS and features are converted with corrections. Lemma accuracy is therefore agreement with an automatic annotation. This is the standard Romanian benchmark, but it is not a hand-verified ceiling.

Limitations

Worth knowing before you build on it:

• This is a tagger, not a parser. No dependency relations, no syntax tree.
• Sentence splitting is rule-based and will occasionally mis-split on unusual abbreviations. If you already have sentence boundaries, send one sentence per request for exact control.
• The model never abstains. Given a typo or an invented word it returns a plausible analysis rather than an error. source tells you when a lemma was predicted rather than looked up, and candidates shows what it was choosing between when the word was ambiguous.
• Very long sentences are truncated. If a single sentence exceeds the model's internal limit, the response carries "truncated": true. Tokens past the cut are still returned — so your token count matches your input — but with upos: "X" and no source key, because nothing analysed them. Send shorter sentences if you hit this.
• Standard Romanian only. Dialectal and heavily informal text is outside what the training data covers.
• 20,000 character limit per request; longer text returns 413.

Errors

Attribution

Built on openly licensed resources — the MULTEXT-East Romanian lexicon and the UD Romanian RRT treebank, both CC BY-SA 4.0, and Romanian BERT. Full citations in ATTRIBUTION.md.