Word alignment¶

The corpus store and MCP server can hold word alignments between two sources, for example a translation and its original. This is what llmaligner builds on.

Alignment sets¶

An alignment set pairs the sentences with the same sent_id in two loaded sources (A and B) and stores word links between them, Ugarit style:

1-2S 2-3S 3-5P      word in A - word in B; S = sure, P = probable

Many-to-many alignments are several links (1-2S 1-3S). Each link can carry a confidence. Sets are JSON Lines files, one sentence pair per line:

{"sent_id": "B10K1V1", "links": [{"source": "1", "target": "2", "sure": true, "confidence": 0.95, "agent": "aligner"}]}

agent records who last set or changed the link (set_links’ agent argument): a link sent again unchanged (same S/P and confidence) keeps its agent, so a program can tell rule links, model links and links a scorer has already rated apart. get_parallel shows them as agents.

Like sources, a set is read-only (trusted, e.g. gold alignments, default confidence 1.0) or writable (default 0.0); a writable file is created when first saved, edits are saved immediately, and every change is appended to <file>.history.jsonl.

On the command line¶

uv run mcparser mcp serve \
    -s arabic=arabic.conllu -s greek=greek.conllu \
    -a verses=arabic,greek,alignments/verses.jsonl \
    -a gold=arabic,greek,alignments/gold.jsonl:ro

Here verses is writable (the default) and gold is read-only.

--alignment NAME=SOURCE_A,SOURCE_B,PATH[:ro|rw[:CONFIDENCE]] is accepted by mcp serve and mcp config (for Claude Desktop).

Tools¶

Tool	Purpose
`list_alignments`	alignment sets with their sources, mode and size
`get_parallel`	a sentence pair: both CoNLL-U sentences, the links with confidences, unlinked words on each side
`translation_candidates`	what a lemma (of A, or of B with `side="b"`) is linked to, with counts, unaligned occurrences and examples
`low_confidence_links`	links below a threshold, least confident first
`set_links`	set (or with `replace=false` add) the links of a pair; word ids are checked against both sentences; `dry_run=true` only checks
`remove_links`	remove links from a pair

From Python¶

from mcparser.corpus import CorpusStore

store = CorpusStore()
store.add_source("arabic", "arabic.conllu")
store.add_source("greek", "greek.conllu")
store.add_alignment("verses", "arabic", "greek", "verses.jsonl")
store.set_links("verses", "B10K1V1", "1-1S 2-2S 3-3P", confidence=0.9, agent="aligner")
store.translation_candidates("رَسُول", side="a").candidates     # → [ValueCount(value='ἀπόστολος', count=...)]