Annotate postposed adjectives that do not agree in case; reynir 3.7 compatibility

.github/workflows/python-package.yml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest]
-        python-version: ['3.9', '3.10', '3.11', '3.12', '3.13', '3.14', 'pypy-3.11']
+        python-version: ['3.11', '3.12', '3.13', '3.14', 'pypy-3.11']
 
     steps:
     - uses: actions/checkout@v4
@@ -28,19 +28,19 @@ jobs:
         python -m pip install uv
         uv pip install --system --upgrade wheel setuptools pytest tokenizer reynir
         # No need to test the sentence classifier in every build (also doesn't work with PyPy)
-        if [ "${{ matrix.python-version }}" == "3.9" ]; then
+        if [ "${{ matrix.python-version }}" == "3.11" ]; then
           uv pip install --system -e ".[sentence_classifier]"
         else
           uv pip install --system -e ".[dev]"
         fi
     - name: Lint with ruff
       run: |
-        if [ "${{ matrix.python-version }}" == "3.9" ]; then uv pip install --system ruff; fi
-        if [ "${{ matrix.python-version }}" == "3.9" ]; then ruff check src/reynir_correct; fi
+        if [ "${{ matrix.python-version }}" == "3.11" ]; then uv pip install --system ruff; fi
+        if [ "${{ matrix.python-version }}" == "3.11" ]; then ruff check src/reynir_correct; fi
     - name: Typecheck with mypy
       run: |
-        if [ "${{ matrix.python-version }}" == "3.9" ]; then uv pip install --system mypy; fi
-        if [ "${{ matrix.python-version }}" == "3.9" ]; then mypy --ignore-missing-imports --python-version=3.9 src/reynir_correct; fi
+        if [ "${{ matrix.python-version }}" == "3.11" ]; then uv pip install --system mypy; fi
+        if [ "${{ matrix.python-version }}" == "3.11" ]; then mypy --ignore-missing-imports --python-version=3.11 src/reynir_correct; fi
     - name: Test with pytest
       run: |
         python -m pytest

pyproject.toml

@@ -17,8 +17,6 @@ classifiers = [
     "Natural Language :: Icelandic",
     "Programming Language :: Python",
     "Programming Language :: Python :: 3",
-    "Programming Language :: Python :: 3.9",
-    "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
@@ -28,7 +26,7 @@ classifiers = [
     "Topic :: Utilities",
     "Topic :: Text Processing :: Linguistic",
 ]
-requires-python = ">=3.9"
+requires-python = ">=3.11"
 dependencies = ["reynir>=3.5.7", "icegrams>=1.1.2", "typing_extensions"]
 
 [project.urls]
@@ -62,7 +60,10 @@ filterwarnings = [
 line-length = 120
 
 [tool.ruff.lint]
-#select = ["ALL"] # We use default rules for now
+# Pin the rule selection (this was ruff's default selection before
+# version 0.15; later versions enable many more rules by default)
+select = ["E4", "E7", "E9", "F"]
+#select = ["ALL"]
 # extend-select = ["E501"] # Complain about line length
 # Ignore specific rules
 # (we should aim to have these as few as possible)

src/reynir_correct/checker.py

@@ -176,6 +176,11 @@ def __init__(
         pipeline: CorrectionPipeline,
         **options: Any,
     ) -> None:
+        # Greynir (reynir >= 3.7) rejects unknown keyword arguments;
+        # only pass through the options that it recognizes
+        known = getattr(self, "_KNOWN_OPTIONS", None)
+        if known is not None:
+            options = {k: v for k, v in options.items() if k in known}
         super().__init__(**options)
         self.settings = settings
         self.pipeline = pipeline

src/reynir_correct/classifier.py

@@ -66,7 +66,7 @@ def __init__(self) -> None:
             )
             warnings.warn(warningtext)
             raise ImportError(warningtext)
-        self.pipe: Any = pipeline(
+        self.pipe: Any = pipeline(  # type: ignore[call-overload]
             "text2text-generation",
             model=self._model_name,
             tokenizer="google/byt5-base",

src/reynir_correct/errfinder.py

@@ -84,6 +84,14 @@ def fget(self, tree: SimpleTree) -> str:
 # Case name prefixes
 CASE_NAMES = {"nf": "nefni", "þf": "þol", "þgf": "þágu", "ef": "eignar"}
 
+# Adjective inflection variants other than case, kept unchanged when
+# suggesting a corrected case form. Note that strong/weak declension
+# variants ('sb'/'vb'/'esb'/'evb') are deliberately not included, as
+# BinPackage's lookup_variants() does not map them onto the FSB/FVB/
+# ESB/EVB inflection marks; BÍN returns strong (FSB) forms first,
+# which is the standard declension in this construction anyway.
+ADJECTIVE_NON_CASE_VARIANTS = frozenset(("et", "ft", "kk", "kvk", "hk", "mst"))
+
 # Replacements for numeric ordinals, in the various genders and cases
 ORDINALS = {
     1: {
@@ -552,6 +560,38 @@ def VillaEinkunn(self, txt: str, variants: str, node: Node) -> AnnotationDict:
             suggest=correct_pronoun,
         )
 
+    def VillaLoEftirNlMeðAndlagi(self, txt: str, variants: str, node: Node) -> Optional[AnnotationDict]:
+        # Lýsingarorð með fallstýrðu andlagi, á eftir nafnlið sem það
+        # sambeygist ekki í falli, t.d.
+        # 'móðurfélag fleiri félaga tengdum rekstri' -> 'tengdra'
+        correct_case = variants.split("_")[0]
+        # The offending adjective is the first terminal within the error phrase
+        tnode = self._terminal_nodes[node.start]
+        wrong_adj = tnode.text
+        # Preserve the adjective's other inflection features,
+        # replacing only the case
+        keep = [v for v in tnode.all_variants if v in ADJECTIVE_NON_CASE_VARIANTS]
+        suggestion = PatternMatcher.get_wordform(
+            wrong_adj.lower(), tnode.lemma, tnode.cat, keep + [correct_case]
+        )
+        if not suggestion or suggestion == wrong_adj.lower():
+            # Can't find a correct form, or it's identical to the original:
+            # don't annotate
+            return None
+        suggestion = emulate_case(suggestion, template=wrong_adj)
+        start, end = tnode.span
+        return AnnotationDict(
+            text="'{0}' á sennilega að vera '{1}'".format(wrong_adj, suggestion),
+            detail=(
+                "Lýsingarorðið '{0}' á að vera í {1}falli, í samræmi við "
+                "nafnliðinn á undan".format(wrong_adj, CASE_NAMES[correct_case])
+            ),
+            start=start,
+            end=end,
+            original=wrong_adj,
+            suggest=suggestion,
+        )
+
     def AðvörunSemOg(self, txt: str, variants: str, node: Node) -> AnnotationDict:
         # 'sem' er sennilega ofaukið
         ch1, ch2 = self._terminal_nodes[node.start : node.end]

test/test_annotator.py

@@ -252,6 +252,30 @@ def test_corrected_meanings(api) -> None:
     )
 
 
+def test_adjective_agreement(api) -> None:
+    """Check annotation of a postposed adjective with a case-governed
+    complement that does not agree with its head noun phrase in case:
+    'forstöðumenn fyrirtækja tengdum sjávarútvegi' -> 'tengdra'"""
+    nts = api.gc.parser.grammar.nonterminals
+    if not any(name.startswith("VillaLoEftirNlMeðAndlagi") for name in nts):
+        # The installed reynir version does not include the
+        # VillaLoEftirNlMeðAndlagi error grammar rule
+        pytest.skip("reynir version lacks the VillaLoEftirNlMeðAndlagi grammar rule")
+    s = "Ég ræddi við forstöðumenn fyrirtækja tengdum sjávarútvegi."
+    check_sentence(api, s, [(5, 5, "P_NT_LoEftirNlMeðAndlagi")])
+    # The corresponding correct sentence should not be annotated
+    s = "Ég ræddi við forstöðumenn fyrirtækja tengdra sjávarútvegi."
+    check_sentence(api, s, [])
+    # Check the suggested correction
+    sent = reynir_correct.check_single(
+        "Ég ræddi við forstöðumenn fyrirtækja tengdum sjávarútvegi."
+    )
+    anns = [a for a in sent.annotations if a.code == "P_NT_LoEftirNlMeðAndlagi"]
+    assert len(anns) == 1
+    assert anns[0].suggest == "tengdra"
+    assert anns[0].original == "tengdum"
+
+
 def test_lhþt_variant(api) -> None:
     """Check for a regression in the handling of LHÞT variants in BinPackage"""
     s = (