farahidi

Arabic morphological analyzer for Python — a pure-Python re-implementation of AlKhalil Morpho Sys 2 (Oujda NLP Team).

Given an Arabic word, farahidi returns every valid morphological analysis — root, lemma, stem, pattern (wazn), part of speech with full features, case/mood, and segmented proclitics/enclitics — ranked by corpus frequency.

• Pure Python, zero dependencies. Works on CPython 3.11 – 3.14.
• Offline. The full lexicon plus the in-context language model ship compressed inside the wheel (~11 MB); nothing is downloaded at runtime.
• Faithful. Output is validated against the original Java AlKhalil2Analyzer (single-word) and ADATAnalyzer (in-context).

This is the Python implementation. A 1:1 JavaScript / TypeScript port lives at farahidi-js, with a live in-browser demo. Both re-implement the original Java AlKhalil Morpho Sys 2 by the Oujda NLP Team.

Named after al-Khalīl ibn Aḥmad al-Farāhīdī (الخليل بن أحمد الفراهيدي), the 8th-century founder of Arabic lexicography and prosody.

Install

pip install farahidi
# or
uv add farahidi

Usage

import farahidi

for a in farahidi.analyze("لِأَنَّهُمْ"):
    print(a.voweled_word, a.lemma, a.root, a.part_of_speech)

analyze() returns a list of Analysis objects, sorted by priority (most frequent analysis first). Each Analysis has these fields (Arabic script; "-" = not applicable, "#" = absent clitic):

field	meaning
voweled_word	fully diacritized surface form
proclitic / enclitic	segmented clitics with their descriptions
stem	the bare stem
lemma	dictionary form
root	the (3- or 4-letter) root
pattern_stem / pattern_lemma	canonical patterns (wazn)
diac_pattern_stem	diacritic pattern of the stem
part_of_speech	pipe-joined POS + morpho-syntactic features
case_or_mood	إعراب (case for nouns, mood for verbs)
priority	out-of-context ranking weight (higher = more frequent)

For repeated analysis, build one reusable analyzer (the lexicon loads lazily and is shared):

from farahidi import Analyzer

az = Analyzer()
results = az.analyze("مدرسة")

In-context disambiguation

analyze_text() picks the single best analysis per token across a sentence, returning one TokenResult per word with the chosen lemma, stem, and root:

import farahidi

for r in farahidi.analyze_text("ذهب الولد إلى المدرسة"):
    print(r.token, r.lemma, r.stem, r.root)
# ذهب ذَهَبَ ذَهَب ذهب
# الولد وَلَد وَلَد ولد
# إلى إِلَى إِلَى -
# المدرسة مَدْرَسَة مَدْرَسَة درس

A reusable Disambiguator is also exposed; disambiguate(tokens) takes a pre-tokenized list. TokenResult.analyzed is False for tokens the analyzer could not analyze (lemma/stem/root then fall back to the token).

Command line

Installing the package also provides a farahidi command with two subcommands mirroring the two entry points:

# Layer 1 — every analysis of each word
farahidi analyze الكتاب لأنهم

# Layer 2 — one disambiguated result per token, in context
farahidi text "ذهب الولد إلى المدرسة" --format table
# token    lemma  stem   root  analyzed
# -------  -----  -----  ----  --------
# ذهب      ذَهَبَ    ذَهَب    ذهب   true
# الولد    وَلَد    وَلَد    ولد   true
# إلى      إِلَى    إِلَى    -     true
# المدرسة  مَدْرَسَة  مَدْرَسَة  درس   true

Pick the output with -f/--format:

format	description
raw (default)	TAB-separated, one record per line, no header — pipe to cut/awk
table	aligned columns with a header (best-effort: Arabic combining marks are measured correctly, but RTL terminals may still reorder cells)
json	JSON Lines — analyze nests every analysis under its word; text emits one token result per line
csv	RFC-4180 with a header row — open in a spreadsheet or load with pandas

The flat formats (raw/table/csv) share one column schema; json keeps the nested structure. With no positional argument both subcommands read from stdin — analyze splits it on whitespace into words, text treats it as the text to tokenize:

echo "مدرسة كتاب" | farahidi analyze -f csv

Scope

• Layer 1 — out-of-context analysis of a single word (analyze), returning all candidates ranked by frequency.
• Layer 2 — in-context disambiguation (analyze_text / Disambiguator), a faithful port of AlKhalil's shipped ADATAnalyzer (lemmatizer + light/heavy stemmer). The chosen lemma is exact; the stem/root are then selected by corpus frequency among that lemma's analyses. On exact frequency ties the pick depends on analysis enumeration order, which can differ from the Java reference (its decoder draws stems/roots from a HashSet); the lemma decode is unaffected.

farahidi is a morphological analyzer — it ports AlKhalil Morpho Sys 2 in full (both shipped layers). It is not a POS tagger or a syntactic/dependency parser; those are separate systems that consume an analyzer's output and are out of scope.

Data & license

farahidi is licensed under the GPL-3.0-or-later, because it bundles and derives from AlKhalil Morpho Sys 2's GPL-3.0 linguistic data. Simply using the library (e.g. pip install and calling analyze) places no obligations on your own code or its outputs. See NOTICE for attribution to the Oujda NLP Team and LICENSE for the full terms.

Development

uv sync
uv run pytest
uv run ruff check

The bundled data is regenerated from the parent morph-analyzer export with python tools/build_data.py. Golden test fixtures are produced from the Java reference with tools/gen_golden.py: --mode words (Layer 1, via AlkhalilGolden.java) and --mode sentences (Layer 2, via AlkhalilSentenceGolden.java).