# Running the extractor

`gemination.py` has no dependencies of its own. To turn new words into
gemination patterns it needs HornMorpho 5.3.6 (Michael Gasser,
https://github.com/hltdi/HornMorpho) with two small patches.

## One command

```bash
./install_hornmorpho.sh .venv-hm          # Python 3.9 or newer; set PYTHON=... to choose one
LANG=C.UTF-8 .venv-hm/bin/python gemination.py ይገኛል ክልል መለወጥ
LANG=C.UTF-8 .venv-hm/bin/python test_gemination.py
```

Expected output of the second line (tab-separated):

```
ይገኛል  ይገ፟ኛ፟ል  positions=[1, 2]  sources=['t-prefix', 'root']  pos=V  root_freq=3688  seg=----ይ</ገ/ኝ>---ኣ/ል---
ክልል   ክል፟ል   positions=[1]     sources=['lexicon']            pos=N  root_freq=1     seg=-<ክ/ልል>------
መለወጥ  መለ፟ወጥ  positions=[1]     sources=['t-prefix']           pos=N  root_freq=84    seg=-<መ/-ለወጥ>------
መለወጥ  መለወ፟ጥ  positions=[2]     sources=['root']               pos=N  root_freq=84    seg=-<መ-ለ/ወጥ>------
```

መለወጥ has two readings (passive mälläwäṭ, active mäläwwäṭ), so it prints two
patterns. Only sentence context can choose between them.

## What the script does, step by step

1. **Install the wheel.** HornMorpho is not on PyPI. The script downloads
   `dist/hornmorpho-5.3.6-py3-none-any.whl` from the HornMorpho repository at
   commit `7e3d93af760e27ea6dbc3b7a078d2d9c3335f618` and checks its sha256
   (`e1b0e4ed...fa7`). Dependencies `conllu` and `requests` come with it.
2. **Patch 1, Tk GUI stub** (`patches/gui.py`, replaces `hm/morpho/gui.py`).
   `hm/morpho/corpus.py` does `from .gui import *` at import time, so
   `import hm` fails on any Python without `_tkinter`. The stub defines the
   one name `corpus.py` refers to. Applied only when `import tkinter` fails.
3. **Patch 2, lexicon loader** (`patches/morphology.patch`, applied with
   `patch -p1` inside site-packages). `words1.srf` lines are
   `ortho phon POS` with `#` comment lines, but the loader built a dict from
   the raw split, which raises on the first real entry.
4. **Amharic data.** `src/hm/languages/a.tgz` from the same commit
   (sha256 `3274f71d...95c`), unpacked into `hm/languages/`.

Always run with a UTF-8 locale (`LANG=C.UTF-8`).

## The call that matters

```python
import hm
readings = list(hm.anal("a", "ይገኛል", degem=False))   # degem=False keeps the '/' marks
```

With the default `degem=True` the `/` marks are stripped from `seg`, and the
only gemination information left is the `+gemN` root feature, which misses
every geminate created by a prefix or a suffix. That was the bug in v1 of this
lexicon.

```python
import gemination as G
G.analyze("ይገኛል")                 # runs HornMorpho, returns patterns, most frequent root first
G.word_gemination(word, readings)  # same, from readings you already have
G.reading_gemination(word, {"seg": "----ይ</ገ/ኝ>---ኣ/ል---"})   # one seg, no HornMorpho needed
G.mark("ይገኛል", [1, 2])            # 'ይገ፟ኛ፟ል'
```

## Rebuilding the release files

`build_lexicon.py` converts the internal lexicon (one entry per word with the
stored HornMorpho segs) into `data/`. It re-derives every pattern from its segs
and stops if any disagrees, so the data cannot drift from this extractor.