From fa862e95f5f586837c5d69ca9fe108719d44f123 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 21 Aug 2026 19:17:31 +0000
Subject: [PATCH 01/10] Register reflexive verb frames instead of dropping them
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Verbs.conf frames with a reflexive pronoun as an object or inside a
prepositional phrase ('átta |sig /á þgf', 'tryggja sér |þf') were
classified as 'complex' and silently skipped when loading, so verbs
whose only frames are reflexive (átta, ímynda, furða, notfæra,
trúlofa, ...) were unknown to the parser and could not take objects
at all, and frames such as 'tryggja sér |þf' never licensed
'tryggja e-m e-ð'. 89 verbs and 227 verb+case keys were affected.

Such frames are now registered under the pronoun's case and marked
as weak. The reducer penalizes a verb terminal whose cases are only
licensed by weak frames, so that e.g. 'á' in 'á pí-deginum' remains
a preposition, and verb_score() ignores the $score() pragmas of weak
frames when strong frames exist for the same key. Frames with fixed
phrases are still skipped; registering them licensed far too much
('eiga |þgf /til góða' -> generic 'eiga' + dative).

The 'átta so < töl' preference is strengthened so that the numeral
reading still wins in 'þrjátíu og átta fermetra'.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/config/Prefs.conf |  2 +-
 src/reynir/reducer.py        | 11 +++++++++++
 src/reynir/verbframe.py      | 38 ++++++++++++++++++++++++++++++------
 test/test_parse.py           | 30 ++++++++++++++++++++++++++++
 4 files changed, 74 insertions(+), 7 deletions(-)

diff --git a/src/reynir/config/Prefs.conf b/src/reynir/config/Prefs.conf
index cce1ba34..70c72042 100644
--- a/src/reynir/config/Prefs.conf
+++ b/src/reynir/config/Prefs.conf
@@ -251,7 +251,7 @@ aukinn no < so
 ári so < no
 ára so < no
 át no << so
-átta so < töl
+átta so <<< töl
 áttu no < so
 banka so < no
 bankar so < no
diff --git a/src/reynir/reducer.py b/src/reynir/reducer.py
index 6704401c..0fee98af 100644
--- a/src/reynir/reducer.py
+++ b/src/reynir/reducer.py
@@ -138,6 +138,10 @@ class ResultDict(TypedDict, total=False):
 
 _VERB_PREP_BONUS = 7  # Give 7 extra points for a verb/preposition match
 _VERB_PREP_PENALTY = -2  # Subtract 2 points for a non-match
+# Penalty for a verb terminal whose argument cases are only licensed by
+# frames with a reflexive pronoun ('eiga sér |þf' -> eiga_þgf_þf), so that
+# a reading with arbitrary objects in those cases is discouraged
+_VERB_WEAK_ONLY_PENALTY = -6
 _LENGTH_BONUS_FACTOR = 10  # For length bonus, multiply number of tokens by this factor
 
 _CASES_SET = BIN_Token.CASES_SET
@@ -721,6 +725,8 @@ def _calc_terminal_scores(self, w: Node) -> ScoreDict:
                         # In the (rare) cases where there are conflicting scores,
                         # apply the most positive adjustment
                         adjmax: Optional[int] = None
+                        # Is the terminal only licensed by weak frames?
+                        weak_only: Optional[bool] = None
                         for m in token.meanings:
                             if m.ordfl == "so":
                                 key = m.stofn + t.verb_cases
@@ -730,7 +736,12 @@ def _calc_terminal_scores(self, w: Node) -> ScoreDict:
                                         adjmax = score
                                     else:
                                         adjmax = max(adjmax, score)
+                                if numcases > 0 and VerbFrame.matches_arguments(key):
+                                    wo = VerbFrame.weak_only(key)
+                                    weak_only = wo if weak_only is None else (weak_only and wo)
                         sc[t] += adj + (adjmax or 0)
+                        if weak_only:
+                            sc[t] += _VERB_WEAK_ONLY_PENALTY
                     if t.is_bh:
                         # Discourage 'boðháttur'
                         sc[t] -= 4
diff --git a/src/reynir/verbframe.py b/src/reynir/verbframe.py
index 98d93c2c..50f80605 100644
--- a/src/reynir/verbframe.py
+++ b/src/reynir/verbframe.py
@@ -235,8 +235,14 @@ def __init__(
         preps: Iterable[Tuple[str, str]],
         particle: Optional[str],
         score: Optional[int],
+        weak: bool = False,
     ) -> None:
         self.verb = verb
+        # True if the frame contains a reflexive pronoun (sig, sér, sín)
+        # as an object or within a prepositional phrase ('átta |sig /á þgf',
+        # 'tryggja sér |þf'); such a frame is registered under its object
+        # cases but is only a weak license for arbitrary objects in those cases
+        self.weak = weak
         # All arguments
         self.args = args
         self.obj = obj
@@ -277,6 +283,7 @@ def create_from_config(cls, s: str) -> None:
         # Format: verb [arg1] [arg2] [/preposition arg]... [*particle] [$pragma(txt)]
 
         complex = False
+        weak = False
 
         def get_score(s: str) -> Tuple[str, Optional[int]]:
             """Handle the $score() pragma, if present"""
@@ -364,7 +371,7 @@ def get_indirect_object(s: str) -> Tuple[str, str]:
                 raise ConfigError("Verb '{0}' is not a valid word".format(verb))
             case = ""
             if len(a) > 1:
-                case = get_case_and_kind((" ".join(a[1:])))
+                case = get_case_and_kind(" ".join(a[1:]))
             return verb, case
 
         def get_case_and_kind(w: str) -> str:
@@ -384,16 +391,17 @@ def get_case_and_kind(w: str) -> str:
             case = ""
             # kind: int = 1    # Default value
             refl = REFLPRN_CASE.get(w, "")
-            nonlocal complex
+            nonlocal complex, weak
             if w in ALL_CASES:
                 # Case 1
                 case = w
                 # kind = 1
             elif refl and refl in ALL_CASES:
-                # Case 2
+                # Case 2: a reflexive pronoun is registered under its case,
+                # so that e.g. 'átta |sig /á þgf' yields the frame 'átta_þf'
                 case = refl
                 # kind = 2
-                complex = True
+                weak = True
             elif w in SUBCLAUSES:
                 # Cases 5-8
                 case = w
@@ -432,7 +440,10 @@ def get_case_and_kind(w: str) -> str:
             args.append(iobj)
         if obj:
             args.append(obj)
-        # Add frame to verb database, if this construct is supported (not 'complex')
+        # Add frame to verb database, if this construct is supported.
+        # Frames with subclause arguments (nh, nhx, falls, spurns, ...)
+        # or fixed phrases are not yet supported; frames with reflexive
+        # pronouns are registered under the pronoun's case, marked as weak.
         if complex:
             return
         vf = cls(
@@ -443,6 +454,7 @@ def get_case_and_kind(w: str) -> str:
             score=score,
             obj=obj,
             iobj=iobj,
+            weak=weak,
         )
         case_key = vf.case_key
         if error:
@@ -501,6 +513,16 @@ def matches_particle(cls, verb_with_cases: str, particle: str) -> bool:
         # Check the frames one by one and return True if any matches are found
         return any(vf.matches_pcl(particle) for vf in verb_frames)
 
+    @classmethod
+    @lru_cache(maxsize=1024)
+    def weak_only(cls, verb_with_cases: str) -> bool:
+        """Return True if the given key, e.g. 'tryggja_þgf_þf', is only
+        licensed by frames containing a reflexive pronoun ('tryggja sér |þf')"""
+        verb_frames = cls.CASE_FRAMES.get(verb_with_cases)
+        if not verb_frames:
+            return False
+        return all(vf.weak for vf in verb_frames)
+
     @classmethod
     @lru_cache(maxsize=1024)
     def verb_score(cls, verb_with_cases: str) -> Optional[int]:
@@ -510,7 +532,11 @@ def verb_score(cls, verb_with_cases: str) -> Optional[int]:
             # No such verb frame, hence no score
             return None
         # Return the highest score associated with a verb frame with the
-        # given arguments, or None if no score was given
+        # given arguments, or None if no score was given. Weak (reflexive)
+        # frames only contribute if no strong frame exists for the key,
+        # so that an idiom's $score() does not apply to arbitrary objects
+        if any(not vf.weak for vf in verb_frames):
+            verb_frames = [vf for vf in verb_frames if not vf.weak]
         score = None
         for vf in verb_frames:
             if vf.score is not None:
diff --git a/test/test_parse.py b/test/test_parse.py
index 4d412846..ad0ac789 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2499,6 +2499,35 @@ def test_rel_clause_pp_attachment(r) -> None:
     )
 
 
+def test_reflexive_verb_frames(r) -> None:
+    """Verb frames in Verbs.conf whose object is a reflexive pronoun
+    ('átta |sig /á þgf', 'tryggja sér |þf') used to be dropped entirely,
+    leaving verbs such as 'átta' unknown to the parser. They are now
+    registered under the pronoun's case, so that 1st and 2nd person
+    objects ('áttaði mig') as well as non-reflexive objects in the same
+    frame ('tryggði gestunum sigur') parse."""
+    for sentence, verb in (
+        ("Ég áttaði mig á þessu.", "átta"),
+        ("Hann áttaði sig á villu síns vegar.", "átta"),
+        ("Við áttuðum okkur ekki á því hvað þetta var erfitt.", "átta"),
+        ("Liðið tryggði gestunum sigur.", "tryggja"),
+        ("Hún ímyndaði sér þetta.", "ímynda"),
+        ("Hún furðaði sig á þessu.", "furða"),
+        ("Hann notfærði sér aðstöðuna.", "notfæra"),
+        ("Hann faðmaði hana.", "faðma"),
+        ("Þau minnka losun.", "minnka"),
+        ("Þau trúlofuðu sig árið 2011 og gengu í hjónaband árið 2013.", "trúlofa"),
+    ):
+        s = r.parse_single(sentence)
+        assert s is not None and s.tree is not None, sentence
+        assert verb in s.tree.verbs, sentence
+    # A verb reading that is only licensed by a reflexive frame
+    # ('eiga sér |þf') must not displace a preposition reading
+    s = r.parse_single("Eðlisfræðingurinn lést í dag, á pí-deginum.")
+    assert s is not None and s.tree is not None
+    assert s.tree.verbs == ["láta"]
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2559,4 +2588,5 @@ def test_rel_clause_pp_attachment(r) -> None:
     test_designation_numeral(g)
     test_skommu_eftir_ad(g)
     test_rel_clause_pp_attachment(g)
+    test_reflexive_verb_frames(g)
     g.__class__.cleanup()

From f8102ca2de1022856e8f4077c297b4c05fc94083 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 21 Aug 2026 19:19:33 +0000
Subject: [PATCH 02/10] Allow a verb particle before the object
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The grammar had no construction for a particle preceding the object:
'bjó til hús', 'kastaði upp boltanum', 'skilaði inn skýrslunni' and
'sagði upp störfum' failed to parse, while 'tók upp bókina' only
parsed by misanalysing 'upp bókina' as a prepositional phrase.

A new nonterminal SagnÖgn (tagged verb_particle) may now appear
between the verb and its object in Sögn_1, Sögn_1_sagnb, Sögn_1_nh,
Sögn_1_bh and NhSögn, all tagged apply_particle_bonus. The reducer
notes the particle text at the SagnÖgn node, carries it up to the
enclosing verb phrase and adjusts that family's score: a bonus if
Verbs.conf lists the particle for the verb with these argument cases
(*til, *upp, ...), using VerbFrame.matches_particle(), which had no
caller before, and otherwise a penalty that outweighs the
preposition penalty, so that 'kastaði upp á þakið' keeps its PP
reading.

'búa |þf *til' is added to Verbs.conf; it was missing altogether.

The expected tree in test_rel_clause_pp_attachment now has
'í umdæmi' in the dative, which is the correct reading.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar   | 36 +++++++++++++++++++++------
 src/reynir/config/Verbs.conf |  1 +
 src/reynir/reducer.py        | 48 +++++++++++++++++++++++++++++++++++-
 test/test_parse.py           | 34 ++++++++++++++++++++++++-
 4 files changed, 110 insertions(+), 9 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index 31fe92bb..3596145b 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -829,6 +829,17 @@ $tag(begin_prep_scope) SagnRunaKnöpp/tala/pers/kyn
 # Sagnir innan undirtrés geta ekki orsakað bónus á forsetningar utan þess
 $tag(purge_verb) SagnRuna/tala/pers/kyn
 
+# Sagnögn sem stendur á milli sagnar og andlags:
+# 'bjó til hús', 'kastaði upp boltanum', 'sagði upp störfum'.
+# Reducer gefur plús ef sögnin hefur ögnina í Verbs.conf (*til, *upp...),
+# en annars mínus, þannig að forsetningarliður sé tekinn fram yfir ögn
+SagnÖgn →
+    "upp:ao" | "niður:ao" | "inn:ao" | "út:ao" | "fram:ao" | "aftur:ao"
+    | "til:ao" | "af:ao" | "frá:ao" | "undan:ao" | "yfir:ao" | "um:ao"
+    | "á:ao" | "við:ao" | "saman:ao" | "burt:ao"
+
+$tag(verb_particle) SagnÖgn
+
 SagnInnskot →
     FsAtv
     | ÁttFs # suður götuna
@@ -2565,19 +2576,23 @@ Sögn_1_sagnb →
     # Jón getur verið [af náttúrunnar hendi] hálfgerður asni
     So_1_sagnb_nf FsRunaEftirSögn? NlSagnfylling_nf
     # Þingmenn geta/hafa lagt [fyrir Alþingi] frumvörp (til laga)
-    | So_1_sagnb/aukafall FsRunaEftirSögn? NlBeintAndlag/aukafall
+    | So_1_sagnb/aukafall SagnÖgn? FsRunaEftirSögn? NlBeintAndlag/aukafall
     | NlEnginnAndlag/fall So_1_sagnb/fall # (Þó getur hann) enga samninga gert
 
+$tag(apply_particle_bonus) Sögn_1_sagnb
+
 NlEnginnAndlag/fall →
     'enginn:fn'/fall/tala/kyn Nl/fall/tala/kyn
 
 # Sögn í nafnhætti með einu viðfangi
 
 Sögn_1_nh →
-    So_1_nh/fall FsRunaInnskot? NlBeintAndlag/fall
+    So_1_nh/fall SagnÖgn? FsRunaInnskot? NlBeintAndlag/fall
     # sjá/láta/'horfa á' Pál breyta [í snarheitum] sér (í kött)
     | So_1_þf_nh Nl_þf So_1_nh/fall FsRunaInnskot? Nl/fall
 
+$tag(apply_particle_bonus) Sögn_1_nh
+
 # Gefa forsetningarliðum milli sagnar og andlags mínus
 # enda eru þær klasturslegt mál ('Hann mun sjá í snatri fílinn')
 FsRunaInnskot →
@@ -2591,7 +2606,9 @@ Sögn_1_bh →
     # Vertu [í alvörunni] þú sjálfur
     So_1_bh_nf FsRunaEftirSögn? NlSagnfylling_nf
     # Bakaðu [í snarheitum] kökuna
-    | So_1_bh/aukafall FsRunaEftirSögn? NlBeintAndlag/aukafall
+    | So_1_bh/aukafall SagnÖgn? FsRunaEftirSögn? NlBeintAndlag/aukafall
+
+$tag(apply_particle_bonus) Sögn_1_bh
 
 # Sagnbót með tveimur viðföngum
 
@@ -2680,8 +2697,9 @@ Sögn_1/tala/pers/kyn →
     # (Hún) sagði lafmóð/sjálf söguna
     so_1_nf/tala/pers OgSögn_1_nf/tala/pers*
         FsRunaEftirSögn? LoTengtSögn_nf/tala/kyn? NlSagnfylling_nf NhFylling?
+    # (Hann) bjó til hús / (Hann) kastaði upp boltanum
     | so_1/aukafall/tala/pers OgSögn_1/aukafall/tala/pers*
-        FsRunaEftirSögn? LoTengtSögn_nf/tala/kyn? NlBeintAndlag/aukafall NhFylling?
+        SagnÖgn? FsRunaEftirSögn? LoTengtSögn_nf/tala/kyn? NlBeintAndlag/aukafall NhFylling?
 
     # (Ég tel/segi/álít) Pál vera dagfarsprúðan mann
     | So_1_þf_nh/tala/pers OgSögn_1_þf_nh/tala/pers* SagnarBotn
@@ -2701,6 +2719,8 @@ Sögn_1/tala/pers/kyn →
     # Ég læt þess getið að Jón er frábær
     | 'láta:so'_gm/tala/pers SögnÞessGetið/tala/kyn
 
+$tag(apply_particle_bonus) Sögn_1/tala/pers/kyn
+
 FáAndlag_þf →
     # Páll fékk bílinn strax endurgreiddan
     NlBeintAndlag_þf/tala/pers/kyn AtvFs? so_lhþt_sb_þf/tala/kyn
@@ -3474,9 +3494,9 @@ $score(+3) NhSagnInnskot
 NhSögn →
     so_0_nh so_0_gm_vh_p3_et? OgNhSögn*                                              # að kjósa [þurfi] og telja (upp á nýtt)
     | so_1_nf_nh so_0_gm_vh_p3_et? OgNhSögn_nf* FsRunaEftirSögn? NlSagnfylling_nf    # að hrjósa [muni] hugur
-    | so_1_þf_nh so_0_gm_vh_p3_et? OgNhSögn_þf* FsRunaEftirSögn? NlBeintAndlag_þf    # að selja [skyldi] í snatri hestinn
-    | so_1_þgf_nh so_0_gm_vh_p3_et? OgNhSögn_þgf* FsRunaEftirSögn? NlBeintAndlag_þgf # að hætta [ætti] baráttunni
-    | so_1_ef_nh so_0_gm_vh_p3_et? OgNhSögn_ef* FsRunaEftirSögn? NlBeintAndlag_ef    # að freista [eigi] gæfunnar
+    | so_1_þf_nh so_0_gm_vh_p3_et? OgNhSögn_þf* SagnÖgn? FsRunaEftirSögn? NlBeintAndlag_þf    # að selja [skyldi] í snatri hestinn / að búa til mat
+    | so_1_þgf_nh so_0_gm_vh_p3_et? OgNhSögn_þgf* SagnÖgn? FsRunaEftirSögn? NlBeintAndlag_þgf # að hætta [ætti] baráttunni / að skila inn skýrslunni
+    | so_1_ef_nh so_0_gm_vh_p3_et? OgNhSögn_ef* SagnÖgn? FsRunaEftirSögn? NlBeintAndlag_ef    # að freista [eigi] gæfunnar
     | so_2_þf_þf_nh so_0_gm_vh_p3_et? OgNhSögn_þf_þf* FsRunaEftirSögn? NlÓbeintAndlag_þf NlBeintAndlag_þf    # að lýsa auðlindir þjóðareign
     | so_2_þf_þgf_nh so_0_gm_vh_p3_et? OgNhSögn_þf_þgf* FsRunaEftirSögn? NlÓbeintAndlag_þf NlBeintAndlag_þgf # að kalla [ætti] Pál illum nöfnum
     | so_2_þf_ef_nh so_0_gm_vh_p3_et? OgNhSögn_þf_ef* FsRunaEftirSögn? NlÓbeintAndlag_þf NlBeintAndlag_ef    # að spyrja og krefja Pál svars
@@ -3506,6 +3526,8 @@ NhSögn →
     # (Hann tók á öllu sínu til að) verða ekki fúll
     | NhSögnLo
 
+$tag(apply_particle_bonus) NhSögn
+
 NhSögnLo →
     HjSögnVeraVerða SögnErLoBotn/tala/kyn
 
diff --git a/src/reynir/config/Verbs.conf b/src/reynir/config/Verbs.conf
index 37bc1b1a..546a62a3 100644
--- a/src/reynir/config/Verbs.conf
+++ b/src/reynir/config/Verbs.conf
@@ -2252,6 +2252,7 @@ búa              /um sig *upp
 búa              /yfir illur_hk_et_þgf
 búa              /að þgf /með þgf *vel
 búa              /við þf # vel, illa
+búa              |þf *til # Hún bjó til mat
 #búa             *afskekkt # afskekkt, rúmt, þröngt
 búa              *um
 búa              *saman
diff --git a/src/reynir/reducer.py b/src/reynir/reducer.py
index 0fee98af..7ea1b89c 100644
--- a/src/reynir/reducer.py
+++ b/src/reynir/reducer.py
@@ -119,6 +119,10 @@ class ResultDict(TypedDict, total=False):
     # Verb scope information
     so: VerbList
     sl: VerbList
+    # Text of an adverb matched by a literal terminal (only at token nodes)
+    ao: str
+    # Text of a verb particle (SagnÖgn), passed to the enclosing verb phrase
+    pcl: str
 
 
 VerbStack = List[Optional[VerbList]]
@@ -138,6 +142,12 @@ class ResultDict(TypedDict, total=False):
 
 _VERB_PREP_BONUS = 7  # Give 7 extra points for a verb/preposition match
 _VERB_PREP_PENALTY = -2  # Subtract 2 points for a non-match
+# Bonus for a verb particle (SagnÖgn) preceding the object, e.g. 'bjó til hús',
+# if the particle is listed for the verb in Verbs.conf (*til); otherwise
+# a penalty, which must outweigh _VERB_PREP_PENALTY so that a preposition
+# reading is preferred for verbs that do not take the particle
+_VERB_PARTICLE_BONUS = 7
+_VERB_PARTICLE_PENALTY = -6
 # Penalty for a verb terminal whose argument cases are only licensed by
 # frames with a reflexive pronoun ('eiga sér |þf' -> eiga_þgf_þf), so that
 # a reading with arbitrary objects in those cases is discouraged
@@ -159,7 +169,7 @@ class _ReductionScope:
     """Class to accumulate information about a nonterminal and its
     child productions during reduction"""
 
-    __slots__ = ("reducer", "sc", "pushed_prep_bonus", "start_verb")
+    __slots__ = ("reducer", "sc", "pushed_prep_bonus", "start_verb", "nt")
 
     def __init__(self, reducer: "ParseForestReducer", node: Node) -> None:
         self.reducer = reducer
@@ -167,6 +177,7 @@ def __init__(self, reducer: "ParseForestReducer", node: Node) -> None:
         self.sc: ChildDict = defaultdict(lambda: {"sc": 0})
         # We are only interested in completed nonterminals
         nt = node.nonterminal if node.is_completed else None
+        self.nt = nt
         # Verb/preposition matching stuff
         self.pushed_prep_bonus = False
         verb = reducer.get_current_verb()
@@ -196,6 +207,13 @@ def add_child(self, ix: int, rd: ResultDict) -> None:
         where the parent family has index ix (0..n)"""
         d = self.sc[ix]
         d["sc"] += rd.get("sc", 0)
+        if "ao" in rd and self.nt is not None and self.nt.has_tag("verb_particle"):
+            # A verb particle (SagnÖgn): note its text so that the enclosing
+            # verb phrase can check it against the verb's frames
+            d["pcl"] = rd["ao"]
+        elif "pcl" in rd:
+            # Carry the particle up to the enclosing verb phrase
+            d["pcl"] = rd["pcl"]
         # Carry information about contained verbs ("so") up the tree
         for key in ("so", "sl"):
             if key in rd:
@@ -222,6 +240,14 @@ def process(self, node: Node) -> Tuple[ResultDict, int]:
 
             nt = node.nonterminal if node.is_completed else None
 
+            if nt is not None and nt.has_tag("apply_particle_bonus"):
+                # Sögn_1 & co.: adjust the score of each family containing
+                # a verb particle (SagnÖgn) by whether the verb(s) take it
+                for d in csc.values():
+                    pcl = d.pop("pcl", None)
+                    if pcl is not None:
+                        d["sc"] += self.reducer.verb_particle_bonus(pcl, d.get("so"))
+
             if len(csc) == 1:
                 # Not ambiguous: only one result, do a shortcut
                 # Will raise an exception if not exactly one item
@@ -265,6 +291,7 @@ def process(self, node: Node) -> Tuple[ResultDict, int]:
                     # and Setning have this tag
                     sc.pop("so", None)  # Simpler than if "so" in sc: del sc["so"]
                     sc.pop("sl", None)
+                    sc.pop("pcl", None)
 
             return sc, ix
 
@@ -349,12 +376,31 @@ def verb_prep_bonus(
         # If no match, discourage
         return _VERB_PREP_PENALTY
 
+    def verb_particle_bonus(self, particle: str, verbs: Optional[VerbList]) -> int:
+        """Return a bonus if any of the given verbs takes the particle
+        according to its frames in Verbs.conf, otherwise a penalty"""
+        if verbs:
+            for verb_terminal, verb_token in verbs:
+                m = verb_token.match_with_meaning(verb_terminal)
+                assert isinstance(m, BIN_Tuple)
+                verb = m.stofn
+                if "MM" in m.beyging:
+                    verb = BIN_Token.mm_verb_stem(verb)
+                verb_with_cases = verb + verb_terminal.verb_cases
+                if VerbFrame.matches_particle(verb_with_cases, "*" + particle):
+                    return _VERB_PARTICLE_BONUS
+        return _VERB_PARTICLE_PENALTY
+
     def visit_token(self, node: Node) -> ResultDict:
         """At token node"""
         # Return the score of this token/terminal match
         d: ResultDict = {"sc": 0}
         nt = cast(BIN_Terminal, node.terminal)
         sc = self._scores[node.start][nt]
+        if nt.matches_category("ao") and nt.is_literal:
+            # Adverb matched by a literal terminal such as "upp:ao":
+            # note its text, in case it is a verb particle (SagnÖgn)
+            d["ao"] = cast(BIN_Token, node.token).lower
         if nt.matches_category("fs"):
             # Preposition terminal - this is either a normal fs_case terminal
             # or a literal terminal such as "á:fs"
diff --git a/test/test_parse.py b/test/test_parse.py
index ad0ac789..530d0286 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2481,7 +2481,7 @@ def test_rel_clause_pp_attachment(r) -> None:
         s.tree.flat_with_all_variants
         == "S0 S-MAIN IP NP-SUBJ pfn_et_hk_nf_p3 /NP-SUBJ VP VP "
         "so_2_þgf_þf_et_fh_gm_p3_þt /VP NP-IOBJ abfn_þgf /NP-IOBJ "
-        "NP-OBJ no_et_kk_þf PP P fs_þf /P NP no_et_hk_þf "
+        "NP-OBJ no_et_kk_þf PP P fs_þgf /P NP no_et_hk_þgf "
         "NP-POSS no_ef_et_kvk to_ft_hk_nf /NP-POSS /NP /PP p "
         "CP-REL C stt /C IP VP VP so_0_et_fh_gm_nt_p3 /VP "
         "PP P fs_þf /P NP no_et_kk_þf C st /C no_et_kk_þf /NP /PP "
@@ -2528,6 +2528,37 @@ def test_reflexive_verb_frames(r) -> None:
     assert s.tree.verbs == ["láta"]
 
 
+def test_verb_particle_before_object(r) -> None:
+    """A verb particle may precede the object ('bjó til hús',
+    'kastaði upp boltanum'); the particle is accepted if Verbs.conf lists
+    it for the verb (*til, *upp), and the object is parsed as a direct
+    object rather than as the object of a preposition."""
+    for sentence, pattern in (
+        ("Hann bjó til hús.", "VP so_1_þf_et_p3 /VP ao NP-OBJ no_et_þf_hk /NP-OBJ"),
+        ("Hann tók upp bókina.", "VP so_1_þf_et_p3 /VP ao NP-OBJ no_et_þf_kvk /NP-OBJ"),
+        ("Hann kastaði upp boltanum.", "VP so_1_þgf_et_p3 /VP ao NP-OBJ no_et_þgf_kk /NP-OBJ"),
+        ("Hann skilaði inn skýrslunni.", "VP so_1_þgf_et_p3 /VP ao NP-OBJ no_et_þgf_kvk /NP-OBJ"),
+        ("Hann sagði upp störfum.", "VP so_1_þgf_et_p3 /VP ao NP-OBJ no_ft_þgf_hk /NP-OBJ"),
+        ("Hún hefur búið til mat.", "VP so_1_þf_sagnb /VP ao NP-OBJ no_et_þf_kk /NP-OBJ"),
+        ("Hún ætlar að búa til mat.", "VP so_1_þf_nh /VP ao NP-OBJ no_et_þf_kk /NP-OBJ"),
+        ("Hann ætlar að skila inn skýrslunni.", "VP so_1_þgf_nh /VP ao NP-OBJ no_et_þgf_kvk /NP-OBJ"),
+        ("Búðu til mat!", "VP so_1_þf_bh /VP ao NP-OBJ no_et_þf_kk /NP-OBJ"),
+    ):
+        s = r.parse_single(sentence)
+        assert s is not None and s.tree is not None, sentence
+        assert pattern in s.tree.flat, (sentence, s.tree.flat)
+    # Prepositional phrases are still preferred where the verb
+    # does not take the particle
+    for sentence in (
+        "Hann bjó í húsinu.",
+        "Hann kastaði upp á þakið.",
+        "Hann ætlar að kasta upp á þakið.",
+    ):
+        s = r.parse_single(sentence)
+        assert s is not None and s.tree is not None, sentence
+        assert " PP " in s.tree.flat and "NP-OBJ" not in s.tree.flat, (sentence, s.tree.flat)
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2589,4 +2620,5 @@ def test_reflexive_verb_frames(r) -> None:
     test_skommu_eftir_ad(g)
     test_rel_clause_pp_attachment(g)
     test_reflexive_verb_frames(g)
+    test_verb_particle_before_object(g)
     g.__class__.cleanup()

From b7fda5be04425da2b33a528ce2f1e176917f41c5 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Thu, 27 Aug 2026 19:08:02 +0000
Subject: [PATCH 03/10] Break reduction ties by grammar order instead of parser
 completion order
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Equally scored families of children were resolved by family index,
i.e. by the order in which the Earley parser completed the derivations
(the C++ forest prepends families, so the last completed derivation
won). That order shifts with unrelated grammar edits, so every change
to a common production flipped a large number of tied parses.

Ties are now broken by a content-based key: the production that
appears first in the grammar wins, and between derivations of the
same production, the one with the longest leading children wins.
Terminal/nonterminal indices are assigned by sorted name and the
production index by grammar position, so the outcome only changes
when an alternative is inserted between two competing alternatives.

Accordingly, give "í" + dative a small bonus over "í" + accusative,
so that locative readings win when the noun form is the same in
both cases ("í umdæmi", "í Reykjavík", "í höfn"); directional
readings governed by verb frames are unaffected.

Measured on 1000 parsed news sentences: 313 trees change once, all
score ties except 84 that improve via the í + þgf bonus; no
measurable speed difference. Engine and GreynirCorrect test suites
pass; PYTHONHASHSEED was verified not to affect results.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/reducer.py |  48 +++++++++++++++++---
 test/test_parse.py    | 103 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 142 insertions(+), 9 deletions(-)

diff --git a/src/reynir/reducer.py b/src/reynir/reducer.py
index 7ea1b89c..9e1c90eb 100644
--- a/src/reynir/reducer.py
+++ b/src/reynir/reducer.py
@@ -92,7 +92,17 @@
 
 """
 
-from typing import Dict, DefaultDict, List, Set, Tuple, Optional, Any, cast
+from typing import (
+    Dict,
+    DefaultDict,
+    Iterable,
+    List,
+    Set,
+    Tuple,
+    Optional,
+    Any,
+    cast,
+)
 from typing_extensions import TypedDict, Required
 
 from collections import defaultdict
@@ -169,12 +179,14 @@ class _ReductionScope:
     """Class to accumulate information about a nonterminal and its
     child productions during reduction"""
 
-    __slots__ = ("reducer", "sc", "pushed_prep_bonus", "start_verb", "nt")
+    __slots__ = ("reducer", "sc", "keys", "pushed_prep_bonus", "start_verb", "nt")
 
     def __init__(self, reducer: "ParseForestReducer", node: Node) -> None:
         self.reducer = reducer
         # Child tree scores
         self.sc: ChildDict = defaultdict(lambda: {"sc": 0})
+        # Content-based tie-break keys, by family index
+        self.keys: Dict[int, Any] = {}
         # We are only interested in completed nonterminals
         nt = node.nonterminal if node.is_completed else None
         self.nt = nt
@@ -195,11 +207,24 @@ def __init__(self, reducer: "ParseForestReducer", node: Node) -> None:
         reducer.push_current_verb(verb)
         self.start_verb = verb
 
-    def start_family(self, ix: int, prod: Production) -> None:
+    def start_family(
+        self, ix: int, prod: Production, children: Iterable[Optional[Node]]
+    ) -> None:
         """Start the processing of a production (numbered ix) of a nonterminal"""
         # Initialize the score of this family of children, so that productions
         # with higher priorities (more negative prio values) get a starting bonus
         self.sc[ix]["sc"] = -10 * prod.priority
+        # Remember a content-based key for this family, used to break
+        # ties between equally scored families deterministically.
+        # The family index itself reflects the order in which the Earley
+        # parser completed the derivations, which shifts with unrelated
+        # grammar edits. Instead, the production that appears first in
+        # the grammar wins, and between derivations of the same
+        # production, the one with the longest leading children wins.
+        self.keys[ix] = (
+            -prod.index,
+            tuple((ch.start, ch.end) for ch in children if ch is not None),
+        )
         self.reducer.set_current_verb(self.start_verb)
 
     def add_child(self, ix: int, rd: ResultDict) -> None:
@@ -253,9 +278,13 @@ def process(self, node: Node) -> Tuple[ResultDict, int]:
                 # Will raise an exception if not exactly one item
                 [(ix, sc)] = csc.items()
             else:
-                # Find the best scoring family, using the lowest
-                # family index as a tie-breaker for determinism
-                ix, sc = max(csc.items(), key=lambda x: (x[1]["sc"], -x[0]))
+                # Find the best scoring family, breaking ties by the
+                # content-based family key (see start_family()) and
+                # only as a last resort by the family index
+                keys = self.keys
+                ix, sc = max(
+                    csc.items(), key=lambda x: (x[1]["sc"], keys[x[0]], -x[0])
+                )
 
             if nt is not None:
                 # Adjust the winning family's score. Note that sc is this
@@ -506,7 +535,7 @@ def calc_score(w: Node, sig: Any) -> ResultDict:
                 fam_children: List[List[Tuple[Node, Any]]] = []
                 # Go through each family and calculate its score
                 for family_ix, (prod, children) in enumerate(w._families):
-                    scope.start_family(family_ix, prod)
+                    scope.start_family(family_ix, prod, children)
                     this_fam: List[Tuple[Node, Any]] = []
                     for ch in children:
                         if ch is not None:
@@ -744,6 +773,11 @@ def _calc_terminal_scores(self, w: Node) -> ScoreDict:
                     elif txt == "á" and t.has_variant("þgf"):
                         # Larger bonus for á + þgf to resolve conflict with verb 'eiga'
                         sc[t] += 4
+                    elif txt == "í" and t.has_variant("þgf"):
+                        # Slightly larger bonus for í + þgf (location) than
+                        # í + þf (direction), to resolve ties where the noun
+                        # form is the same in both cases ("í umdæmi")
+                        sc[t] += 3
                     else:
                         # Else, give a bonus for each matched preposition
                         sc[t] += 2
diff --git a/test/test_parse.py b/test/test_parse.py
index 530d0286..1284364c 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2482,11 +2482,12 @@ def test_rel_clause_pp_attachment(r) -> None:
         == "S0 S-MAIN IP NP-SUBJ pfn_et_hk_nf_p3 /NP-SUBJ VP VP "
         "so_2_þgf_þf_et_fh_gm_p3_þt /VP NP-IOBJ abfn_þgf /NP-IOBJ "
         "NP-OBJ no_et_kk_þf PP P fs_þgf /P NP no_et_hk_þgf "
-        "NP-POSS no_ef_et_kvk to_ft_hk_nf /NP-POSS /NP /PP p "
+        "NP-POSS no_ef_et_kvk to_ft_hk_nf p "
         "CP-REL C stt /C IP VP VP so_0_et_fh_gm_nt_p3 /VP "
         "PP P fs_þf /P NP no_et_kk_þf C st /C no_et_kk_þf /NP /PP "
-        "/VP /IP /CP-REL /NP-OBJ /VP /IP /S-MAIN p /S0"
+        "/VP /IP /CP-REL /NP-POSS /NP /PP /NP-OBJ /VP /IP /S-MAIN p /S0"
     )
+    # The relative clause attaches to the nearest noun (lögreglustöðvar)
     # A preposition that matches the main verb (fresta e-u vegna e-s)
     # must still attach to the verb phrase, not the object noun phrase
     s = r.parse_single("Dómarinn frestaði mótinu vegna veðurs.")
@@ -2559,6 +2560,102 @@ def test_verb_particle_before_object(r) -> None:
         assert " PP " in s.tree.flat and "NP-OBJ" not in s.tree.flat, (sentence, s.tree.flat)
 
 
+def test_deterministic_tie_break() -> None:
+    """Equally scored families of children are resolved by the position
+    of the production in the grammar (first alternative wins) and then
+    by the spans of the child nodes, never by the order in which the
+    parser happened to complete the derivations."""
+    from reynir.reducer import _ReductionScope
+
+    class FakeReducer:
+        """Stand-in for ParseForestReducer with no verb context"""
+
+        def get_current_verb(self):
+            return None
+
+        def set_current_verb(self, verb):
+            pass
+
+        def push_current_verb(self, verb):
+            pass
+
+        def pop_current_verb(self):
+            pass
+
+    class FakeNode:
+        is_completed = False
+        nonterminal = None
+        start = 0
+        end = 3
+
+        def __init__(self, start=0, end=3):
+            self.start = start
+            self.end = end
+
+    class FakeProd:
+        priority = 0
+
+        def __init__(self, index):
+            self.index = index
+
+    def choose(families):
+        """families: list of (prod, children); all scored equally.
+        Returns the index of the winning family."""
+        reducer = FakeReducer()
+        scope = _ReductionScope(reducer, FakeNode())  # type: ignore
+        for ix, (prod, children) in enumerate(families):
+            scope.start_family(ix, prod, children)  # type: ignore
+            for _ in children:
+                scope.add_child(ix, {"sc": 1})  # type: ignore
+        _, chosen = scope.process(FakeNode())  # type: ignore
+        return chosen
+
+    a, b = FakeNode(0, 1), FakeNode(1, 3)
+    # The production appearing first in the grammar wins,
+    # regardless of family order
+    assert choose([(FakeProd(10), [a, b]), (FakeProd(5), [a, b])]) == 1
+    assert choose([(FakeProd(5), [a, b]), (FakeProd(10), [a, b])]) == 0
+    # Same production: the derivation whose first child spans more wins
+    p = FakeProd(7)
+    left_heavy = [FakeNode(0, 2), FakeNode(2, 3)]
+    right_heavy = [FakeNode(0, 1), FakeNode(1, 3)]
+    assert choose([(p, right_heavy), (p, left_heavy)]) == 1
+    assert choose([(p, left_heavy), (p, right_heavy)]) == 0
+    # A better score still beats a better key
+    reducer = FakeReducer()
+    scope = _ReductionScope(reducer, FakeNode())  # type: ignore
+    scope.start_family(0, FakeProd(5), [a, b])  # type: ignore
+    scope.add_child(0, {"sc": 1})  # type: ignore
+    scope.start_family(1, FakeProd(10), [a, b])  # type: ignore
+    scope.add_child(1, {"sc": 2})  # type: ignore
+    assert scope.process(FakeNode())[1] == 1  # type: ignore
+
+
+def test_locative_i_prefers_dative(r) -> None:
+    """When the noun form is the same in accusative and dative, 'í' is
+    read as locative (dative) unless a verb frame says otherwise"""
+    for sentence in (
+        "Áfram óvissa meðan ástandið er svona í Úkraínu.",
+        "Regína er sviðsstjóri velferðarsviðs í Reykjavík.",
+        "Samkomulagið er í höfn.",
+        "Slysið varð í umdæmi lögreglunnar.",
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        assert "P fs_þgf /P" in s.tree.flat, sentence
+        assert "P fs_þf /P" not in s.tree.flat, sentence
+    # Directional readings governed by the verb frame stay accusative
+    for sentence in (
+        "Chelsea kom í heimsókn.",
+        "Hann fór í bæinn.",
+        "Hún setti bókina í hillu.",
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        assert "P fs_þf /P" in s.tree.flat, sentence
+        assert "P fs_þgf /P" not in s.tree.flat, sentence
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2621,4 +2718,6 @@ def test_verb_particle_before_object(r) -> None:
     test_rel_clause_pp_attachment(g)
     test_reflexive_verb_frames(g)
     test_verb_particle_before_object(g)
+    test_deterministic_tie_break()
+    test_locative_i_prefers_dative(g)
     g.__class__.cleanup()

From a339bd5d60ce5b02c44d30c0e05743189002736b Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Thu, 27 Aug 2026 19:36:06 +0000
Subject: [PATCH 04/10] Parse verbs with a cliticized second person subject
 (ertu, viltu)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

BÍN lists interrogative verb forms with a cliticized subject
('spurnarmyndir': ertu, viltu, geturðu, veistu, komstu) tagged with
SP, and binparser already reserved a restrictive 'sp' variant for
them, but the grammar never used it, so sentences such as "Ertu
búinn?", "Viltu koma?" or "Hvernig komstu í kynni við hana?" failed
to parse (or, for 'ertu', parsed as an imperative of 'erta').

Add 'sp' as a fourth value of the /pers variant. The existing verb
phrase productions then yield Sagnliður_et_sp whose finite verb
terminals (so_..._et_sp) only match SP forms. A new SagnHluti_et_p2
alternative, SagnliðurSníkill, uses it as a verb-first clause with
an included subject, which covers questions, fronted adverbials
("Þá viltu það ekki") and the other verb-first contexts. Noun
phrases of the 'sp' person map to those of the second person.

Measured on the unparsed-sentence corpus: 802 of 1017 sentences
containing such forms now parse. On 1000 previously parsed news
sentences, 5 trees change (all containing clitic forms, all
improved). The grammar grows by ~10% (nonterminals 6808 -> 7513)
and parsing is ~5% slower. Engine and GreynirCorrect test suites pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar | 19 ++++++++++++++++---
 src/reynir/simpletree.py   |  1 +
 test/test_parse.py         | 38 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 55 insertions(+), 3 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index 3596145b..33f182ba 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -112,7 +112,7 @@
 /tala = et ft
 
 # Person (1st, 2nd, 3rd) variant
-/pers = p1 p2 p3
+/pers = p1 p2 p3 sp   # sp = spurnarmynd með viðskeyttu frumlagi (ertu, viltu)
 
 # Hacks to generate single variants only
 /hk = hk
@@ -401,8 +401,7 @@ Spurnarsetning →
     | SpurnarForsetningarliður BeygingarliðurSögnFremst? Spurningarmerki       # Í hvaða fötum er barnið?
     | SpurnarAtviksorð BeygingarliðurSögnFremst? Spurningarmerki               # Hvernig missti tröllið lífið?
     | SpurnarFrumlag/tala/kyn BlTagl/tala/p3/kyn? Spurningarmerki              # Hvað er í pokanum?
-    | BeygingarliðurSögnFremst Spurningarmerki                                 # Getur þú hlaupið hratt?
-    #| BeygingarliðurSpurnarsníkill Spurningarmerki                            # Ferðu á ballið?
+    | BeygingarliðurSögnFremst Spurningarmerki                                 # Getur þú hlaupið hratt? / Ferðu á ballið?
 
 # Spurnarorðið er frumlag setningarinnar, stendur utan beygingarliðar
 BlTagl/tala/pers/kyn →
@@ -891,6 +890,15 @@ SagnHlutiMegiSkuli →
     # [Ætíð] skal fara með skjölin sem trúnaðarmál
     HjSögnMegiSkuli_et_p3 FsAtv? EinSögn_nh_et_hk SagnInnskot?
 
+# Sögn með viðskeyttu frumlagi ('spurnarmynd'): Ertu búinn? Viltu koma?
+# (Hvernig) komstu í kynni við hana? (Þá) viltu það ekki
+# Sagnmyndin sjálf inniheldur frumlagið 'þú', sem því vantar í setninguna
+SagnHluti_et_p2/kyn →
+    SagnliðurSníkill/kyn SagnInnskot? SagnViðhengi_et_p2/kyn?
+
+SagnliðurSníkill/kyn →
+    Sagnliður_et_sp/kyn
+
 $tag(begin_prep_scope) SagnHluti/tala/pers/kyn # Nýtt undirtré fyrir tengingu sagnar og forsetningar
 
 FrumlagsInnskot/tala/pers/kyn →
@@ -3926,6 +3934,11 @@ NlStak_p2_ft/fall/kyn →
 
 NlStak_ft_p2/fall/kyn → NlStak_p2_ft/fall/kyn FsRunaEftirNl?
 
+# Persónan 'sp' (spurnarmynd sagnar með viðskeyttu frumlagi, 'ertu')
+# er í raun 2. persóna eintölu; nafnliðir hennar eru þeir sömu
+NlStak_et_sp/fall/kyn → NlStak_et_p2/fall/kyn
+NlStak_ft_sp/fall/kyn → NlStak_ft_p2/fall/kyn
+
 # !!! ATH: reglurnar hér fyrir ofan leyfa 'ég og þú fóruð til Frakklands'
 
 # Sameiginlegar reglur fyrir 3p. eintölu og fleirtölu
diff --git a/src/reynir/simpletree.py b/src/reynir/simpletree.py
index a63100e7..c4ae0d35 100644
--- a/src/reynir/simpletree.py
+++ b/src/reynir/simpletree.py
@@ -244,6 +244,7 @@ class SimpleTreeNode(CanonicalTokenDict, total=False):
     # "NhSögn": "VP",
     "NhEinfaldur": "VP",
     "SagnliðurÁnF": "VP",
+    "SagnliðurSníkill": "VP",
     "ÖfugurSagnliður": "VP",
     "SagnliðurVh": "VP",
     "HjSögnLhÞt": "VP",  # Auxiliary verb, hjálparsögn
diff --git a/test/test_parse.py b/test/test_parse.py
index 1284364c..1338cb71 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2656,6 +2656,43 @@ def test_locative_i_prefers_dative(r) -> None:
         assert "P fs_þgf /P" not in s.tree.flat, sentence
 
 
+def test_verb_with_clitic_subject(r) -> None:
+    """Interrogative verb forms with a cliticized second person subject
+    ('ertu', 'viltu', 'geturðu') parse as verbs with an included subject"""
+    for sentence, lemma in (
+        ("Ertu búinn?", "vera"),
+        ("Viltu koma út að leika?", "vilja"),
+        ("Geturðu hjálpað mér?", "geta"),
+        ("Veistu hvað klukkan er?", "vita"),
+        ("Hefurðu heyrt eitthvað frá íbúunum?", "hafa"),
+        ("Ertu duglegur að hreyfa þig?", "vera"),
+        ("Hvernig komstu í kynni við vinnuna?", "koma"),
+        ("Ferðu á ballið?", "fara"),
+        ("Áttu einhverja sögu af vandræðalegu stefnumóti?", "eiga"),
+        ("Þá viltu það ekki.", "vilja"),
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        flat = s.tree.flat
+        assert "NP-SUBJ" not in flat, sentence
+        assert "_sp" in flat, sentence
+        assert lemma in s.tree.verbs, sentence
+    # The clitic form is not an imperative of 'erta'
+    s = r.parse_single("Ertu búinn?")
+    assert s and s.tree
+    assert s.tree.flat == (
+        "S0 S-QUE IP VP VP-AUX so_et_sp /VP-AUX VP so_lhþt_sb_nf_et_kk /VP "
+        "/VP /IP p /S-QUE /S0"
+    )
+    # A regular second person form does not get an implied subject
+    s = r.parse_single("Ert búinn?")
+    assert s and s.tree is None
+    # An explicit subject still parses normally
+    s = r.parse_single("Ert þú búinn?")
+    assert s and s.tree
+    assert "NP-SUBJ pfn_et_nf /NP-SUBJ" in s.tree.flat
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2720,4 +2757,5 @@ def test_locative_i_prefers_dative(r) -> None:
     test_verb_particle_before_object(g)
     test_deterministic_tie_break()
     test_locative_i_prefers_dative(g)
+    test_verb_with_clitic_subject(g)
     g.__class__.cleanup()

From 1bc92773cd5ff0cefe250ea5c17bdbc798c98c3f Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Thu, 27 Aug 2026 19:57:28 +0000
Subject: [PATCH 05/10] Discourage clitic verb readings of ambiguous forms
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

BÍN lists interrogative clitic forms (SP) for practically every verb,
including obscure ones whose form coincides with a common adjective
('næstu' = ná + þú, 'mestu' = meta + þú, 'elstu' = ala + þú) or with
a regular verb form ('áttu', 'yrðu', 'máttu' as 3rd person plural).
On 1500 parsed news sentences containing any such token, about 30
switched to a wrong clitic reading after the previous commit.

Penalize an sp terminal in the reducer when the token also has an
adjective reading (-40) or a non-clitic verb reading (-6), and give
SagnliðurSníkill a small negative score so that a regular analysis
wins whenever one is available. This brings the wrong readings down
to a handful, all in sentences whose previous parse was poor, while
genuine clitic questions ('Ertu ...?', 'Áttu von á ...?', 'Hvenær
áttu að eiga?') keep their new analysis. The cluster recovery
(802/1017) and the 1k baseline (5 improved trees) are unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar |  3 +++
 src/reynir/reducer.py      | 19 +++++++++++++++++++
 2 files changed, 22 insertions(+)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index 33f182ba..e1345447 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -899,6 +899,9 @@ SagnHluti_et_p2/kyn →
 SagnliðurSníkill/kyn →
     Sagnliður_et_sp/kyn
 
+# Kjósum fremur venjulega greiningu ef hún er tæk
+$score(-8) SagnliðurSníkill/kyn
+
 $tag(begin_prep_scope) SagnHluti/tala/pers/kyn # Nýtt undirtré fyrir tengingu sagnar og forsetningar
 
 FrumlagsInnskot/tala/pers/kyn →
diff --git a/src/reynir/reducer.py b/src/reynir/reducer.py
index 9e1c90eb..989a74f6 100644
--- a/src/reynir/reducer.py
+++ b/src/reynir/reducer.py
@@ -166,6 +166,11 @@ class ResultDict(TypedDict, total=False):
 
 _CASES_SET = BIN_Token.CASES_SET
 
+# Penalties for verb forms with a cliticized subject ('ertu', 'viltu')
+# when the token also has an adjective reading ('næstu', 'mestu') or
+# a regular verb reading ('áttu', 'yrðu')
+_CLITIC_ADJECTIVE_PENALTY = -40
+_CLITIC_VERB_PENALTY = -6
 _CONTAINED_VERBS_SET = frozenset(("begin_prep_scope", "purge_verb"))
 
 # BÍN categories ('fl') of person and entity names
@@ -822,6 +827,20 @@ def _calc_terminal_scores(self, w: Node) -> ScoreDict:
                         sc[t] += adj + (adjmax or 0)
                         if weak_only:
                             sc[t] += _VERB_WEAK_ONLY_PENALTY
+                    if t.has_variant("sp"):
+                        # Interrogative form with a cliticized subject
+                        # ('ertu', 'viltu'). BÍN lists such forms for many
+                        # verbs whose clitic form coincides with a common
+                        # adjective ('næstu', 'mestu', 'elstu') or with a
+                        # regular verb form ('áttu', 'yrðu', 'máttu'): in
+                        # those cases, discourage the clitic reading
+                        if any(m.ordfl == "lo" for m in token.meanings):
+                            sc[t] += _CLITIC_ADJECTIVE_PENALTY
+                        elif any(
+                            m.ordfl == "so" and "SP" not in m.beyging
+                            for m in token.meanings
+                        ):
+                            sc[t] += _CLITIC_VERB_PENALTY
                     if t.is_bh:
                         # Discourage 'boðháttur'
                         sc[t] -= 4

From 2bc6cde329b36ef90966a83f90a9df813657cb33 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Thu, 27 Aug 2026 20:11:58 +0000
Subject: [PATCH 06/10] =?UTF-8?q?Parse=20the=20free=20article=20with=20a?=
 =?UTF-8?q?=20nominalized=20adjective=20(hi=C3=B0=20sama,=20hi=C3=B0=20opi?=
 =?UTF-8?q?nbera)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The free article 'hinn/hin/hið' followed by an adjective and no noun
('hið sama', 'hið rétta', 'hið opinbera', 'hið versta', 'hið góða og
hið illa') had no analysis unless the article form happened to have
a pronoun reading in BÍN (as 'hinu', 'hins' do), so most such
sentences failed to parse.

Add NlGreinirLo (Greinir + LoLiður) as a NlStak_p3 alternative, and
HiðFyrsta, an adverbial phrase for the fixed superlative expressions
'hið fyrsta', 'hið snarasta', 'hið bráðasta', 'hið minnsta', 'hið
mesta' (with optional 'allra'), so that they don't become verb objects.
Also let the adjective reading of 'sama' survive the Phrases.conf
disambiguation of 'sama gildir/gilti/gildi', which otherwise blocks
'Hið sama gildir um ...'.

Measured on the unparsed-sentence corpus: 317 of 517 sentences with
such phrases now parse. On 1000 previously parsed news sentences,
3 trees change, all now using the article rather than the pronoun.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar     | 26 ++++++++++++++++++++++++++
 src/reynir/config/Phrases.conf |  6 +++---
 test/test_parse.py             | 26 ++++++++++++++++++++++++++
 3 files changed, 55 insertions(+), 3 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index e1345447..c7fb4621 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -4170,6 +4170,15 @@ $score(+24) GreinirMeðLo/tala/fall/kyn
 NlStak_p3/tala/fall/kyn →
     AtviksliðurEinkunn* LoSemNafnliður/tala/fall/kyn
 
+# Lýsingarorð með lausum greini sem stendur sem nafnliður:
+# 'Hið sama á við um starfsfólkið', 'Hið rétta er að ...',
+# 'hið opinbera greiðir', 'björgunarsveitir óttast hið versta'
+NlStak_p3/tala/fall/kyn →
+    NlGreinirLo/tala/fall/kyn
+
+NlGreinirLo/tala/fall/kyn →
+    Greinir/tala/fall/kyn LoLiður/tala/fall/kyn
+
 # 'Fleiri/færri komu í veisluna en ég bjóst við'
 # 'Fleirum fannst kvikmyndin leiðinleg'
 NlStak_p3_ft/fall/kyn →
@@ -6392,6 +6401,22 @@ FsUndirÁÍ/fall →
 
 # Atviksliður
 
+# Atviksliður úr lausum greini og efsta stigi lýsingarorðs:
+# (leysa) hið fyrsta / (yfirgefa) hið snarasta / (koma) hið allra fyrsta
+HiðFyrsta →
+    'hinn:gr'_et_þf_hk "allra:ao"? HiðFyrstaLo
+
+HiðFyrstaLo →
+    'fyrstur:lo'_vb_et_þf_hk      # hið fyrsta
+    | 'snar:lo'_vb_et_þf_hk      # hið snarasta
+    | 'bráður:lo'_vb_et_þf_hk    # hið bráðasta
+    | 'fljótur:lo'_vb_et_þf_hk   # hið fljótasta
+    | 'skjótur:lo'_vb_et_þf_hk   # hið skjótasta
+    | 'lítill:lo'_vb_et_þf_hk    # hið minnsta
+    | 'mikill:lo'_vb_et_þf_hk    # hið mesta
+
+$score(+16) HiðFyrsta
+
 Al → EinnAl+
 
 EinnAl →
@@ -6407,6 +6432,7 @@ EinnAl →
     | AlHvortSemUmErAðRæða  # Hvort sem um er að ræða X eða Y
     | AðNýju
     | FjöldiSaman           # Tugþúsundum saman
+    | HiðFyrsta             # hið fyrsta, hið snarasta, hið allra bráðasta
 
     # Fjölyrtir atviksliðir sem ekki er unnt að búa til fasta frasa úr
 
diff --git a/src/reynir/config/Phrases.conf b/src/reynir/config/Phrases.conf
index deb91f6e..7515f2ee 100644
--- a/src/reynir/config/Phrases.conf
+++ b/src/reynir/config/Phrases.conf
@@ -1677,9 +1677,9 @@ meaning = fs frasi -
 "lokið er við" so so fs/ao
 "eiga* afturkvæmt" so ao
 "nýta* til uppbyggingar" so fs kvk
-"sama gilti" fn so
-"sama gildi" fn so
-"sama gildir" fn so
+"sama gilti" fn/lo so
+"sama gildi" fn/lo so
+"sama gildir" fn/lo so
 "nefna* sem dæmi" so fs hk
 "taka* sem dæmi" so fs hk
 "vera* boðið upp á" so so ao fs
diff --git a/test/test_parse.py b/test/test_parse.py
index 1338cb71..30c1ab88 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2693,6 +2693,31 @@ def test_verb_with_clitic_subject(r) -> None:
     assert "NP-SUBJ pfn_et_nf /NP-SUBJ" in s.tree.flat
 
 
+def test_article_with_nominalized_adjective(r) -> None:
+    """The free article 'hinn/hin/hið' followed by an adjective forms
+    a noun phrase on its own ('hið sama', 'hið opinbera', 'hið versta'),
+    or an adverbial phrase with certain superlatives ('hið fyrsta')"""
+    for sentence, phrase in (
+        ("Hið sama á við um starfsfólk deildarinnar.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk /NP-SUBJ"),
+        ("Hið rétta er að það kviknaði í bátnum.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk /NP-SUBJ"),
+        # 'sama gildir' is disambiguated in Phrases.conf; the adjective
+        # reading must survive so that the article can attach to it
+        ("Hið sama gildir um fleiri tegundir.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk /NP-SUBJ"),
+        ("Hið opinbera greiðir kostnaðinn.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk /NP-SUBJ"),
+        ("Björgunarsveitir óttast hið versta.", "NP-OBJ gr_et_þf_hk lo_þf_et_hk /NP-OBJ"),
+        ("Eftir útskýringar kom hið sanna í ljós.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk /NP-SUBJ"),
+        ("Hann sá hið góða og hið illa.", "gr_et_þf_hk lo_þf_et_hk C st /C gr_et_þf_hk lo_þf_et_hk"),
+        ("Verkefnið þarf að leysa hið fyrsta.", "ADVP gr_et_þf_hk lo_vb_et_þf_hk /ADVP"),
+        ("Íbúum var gert að yfirgefa heimili sín hið snarasta.", "ADVP gr_et_þf_hk lo_vb_et_þf_hk /ADVP"),
+        # The article still attaches to a following noun
+        ("Hið nýja hús er stórt.", "NP-SUBJ gr_et_nf_hk lo_nf_et_hk no_et_nf_hk /NP-SUBJ"),
+        ("Hann fékk hið fyrsta verkefni.", "NP-OBJ gr_et_þf_hk lo_þf_et_hk no_et_þf_hk /NP-OBJ"),
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        assert phrase in s.tree.flat, (sentence, s.tree.flat)
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2758,4 +2783,5 @@ def test_verb_with_clitic_subject(r) -> None:
     test_deterministic_tie_break()
     test_locative_i_prefers_dative(g)
     test_verb_with_clitic_subject(g)
+    test_article_with_nominalized_adjective(g)
     g.__class__.cleanup()

From f92943129bdb1588ef59c12260980310195478d5 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 28 Aug 2026 12:54:53 +0000
Subject: [PATCH 07/10] Parse a quoted question or exclamation followed by an
 attribution
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Direct speech ending in a question mark or exclamation mark is not
followed by a comma before the attribution ("Komst hann ekki í
hópinn?" sagði hann / "Við erum að falla á tíma!" skrifar Limbourg),
but Staðhæfing required one, so such sentences did not parse.

Add SagtSpurning (a question clause, or a sentence ending in "!") and
SegirÁnKommu (attribution with optional comma) as a second Staðhæfing
alternative, with an optional "og bætti við: ..." continuation. Add
the attribution verbs spyrja, svara, hrópa, kalla, tísta, hugsa and
útskýra, "bætir X við" as an attribution on its own, and gesture
continuations ("segir KK og hlær", "spurði hún og kinkaði kolli").
A proper noun as the attributed source (Heimild) is now third person
singular only, instead of getting an arbitrary person by tie-break.

Measured on the unparsed-sentence corpus: 241 of 448 sentences with a
quoted question/exclamation and attribution now parse. On 1000
previously parsed news sentences, 5 trees change (proper-noun
attributions now p3; one statement now analyzed as a quote), 0 lost.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar | 48 +++++++++++++++++++++++++++++++++++---
 src/reynir/simpletree.py   |  1 +
 test/test_parse.py         | 29 +++++++++++++++++++++++
 3 files changed, 75 insertions(+), 3 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index c7fb4621..08b04ee4 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -155,6 +155,21 @@ MgrInnihald →
 # "Allt er klárt," segir Jón og bætir við: "Komum þessu í gang."
 Staðhæfing →
     Sagt Segjandi ÁframSagt? Lokatákn?
+    # Spurning eða upphrópun í beinni ræðu, án kommu á undan segjanda:
+    # "Komst hann ekki í hópinn?" sagði hann. / "Við erum að falla á tíma!" skrifar Limbourg.
+    | SagtSpurning SegirÁnKommu SegirÁframSagt? Lokatákn?
+
+# og bætti við: "Þetta er mjög gott."
+SegirÁframSagt →
+    SegirÁfram ":"? Sagt?
+
+SagtSpurning →
+    Forskeyti? Upphrópun* Spurnarsetning
+    | Forskeyti? Yfirsetning StLiður* "!"
+
+SegirÁnKommu →
+    ","? SögnSegja/tala/pers Heimild_nf/tala/pers AtvFsAuka?
+    | ","? SögnBætaVið/tala/pers
 
 ÁframSagt →
     Segir SegirÁfram Sagt
@@ -197,11 +212,16 @@ StrikEðaKomma → "-:em" | ","
 $score(-6) Forskeyti # Ekki freistast í forskeyti ef annað er í boði
 
 Heimild_nf/tala/pers →
-    Nl_nf/tala/pers | Sérnafn
+    Nl_nf/tala/pers
 Heimild_þgf/tala/pers →
-    Nl_þgf/tala/pers | Sérnafn
+    Nl_þgf/tala/pers
 Heimild_ef/tala/pers →
-    Nl_ef/tala/pers | Sérnafn
+    Nl_ef/tala/pers
+
+# Sérnafn sem heimild er alltaf 3. persóna eintölu: segir Limbourg
+Heimild_nf_et_p3 → Sérnafn
+Heimild_þgf_et_p3 → Sérnafn
+Heimild_ef_et_p3 → Sérnafn
 
 # Koma í veg fyrir að þessir hnútar séu bestaðir burt
 $tag(keep) Heimild_nf/tala/pers
@@ -213,10 +233,25 @@ Sagt → Forskeyti? Yfirsetning StLiður*
 Segir →
     # , segir/fullyrðir Jón Ásmundsson, fulltrúi samtakanna, í dag í útvarpinu
     "," SögnSegja/tala/pers Heimild_nf/tala/pers AtvFsAuka?
+    # , bætir Jón við
+    | "," SögnBætaVið/tala/pers
+
+# bætir Jón við / bætir hún við í færslu sinni
+SögnBætaVið/tala/pers →
+    'bæta:so'_0_gm_fh/tala/pers Heimild_nf/tala/pers "við:ao" AtvFsAuka?
 
 SegirÁfram →
     "og:st" SögnSegja/tala/pers SegirLíka?
     | "og:st" SögnSegjaÁfram/tala/pers
+    | "og:st" SögnSegjaLátbragð/tala/pers   # spurði Þorgerður og hló
+
+# Látbragð sem fylgir tilvitnun: segir KK og hlær / brosir / andvarpar
+SögnSegjaLátbragð/tala/pers →
+    'hlæja:so'_0_gm_fh/tala/pers
+    | 'brosa:so'_0_gm_fh/tala/pers
+    | 'andvarpa:so'_0_gm_fh/tala/pers
+    | 'kinka:so'_0_gm_fh/tala/pers 'kollur:kk'_þgf_et
+    | 'hrista:so'_0_gm_fh/tala/pers 'höfuð:hk'_þf_et_gr
 
 $score(+999) SegirÁfram
 
@@ -232,6 +267,13 @@ SögnSegja/tala/pers →
     | 'ítreka:so'_0_gm_fh/tala/pers
     | 'viðurkenna:so'_0_gm_fh/tala/pers
     | 'álíta:so'_0_gm_fh/tala/pers
+    | 'spyrja:so'_0_gm_fh/tala/pers      # "Er þetta satt?" spyr Jón
+    | 'svara:so'_0_gm_fh/tala/pers
+    | 'hrópa:so'_0_gm_fh/tala/pers
+    | 'kalla:so'_0_gm_fh/tala/pers
+    | 'tísta:so'_0_gm_fh/tala/pers
+    | 'hugsa:so'_0_gm_fh/tala/pers
+    | 'útskýra:so'_0_gm_fh/tala/pers
 
 SögnSegjaÁfram/tala/pers →
     'bæta:so'_0_gm_fh/tala/pers "við:ao"
diff --git a/src/reynir/simpletree.py b/src/reynir/simpletree.py
index c4ae0d35..79c76232 100644
--- a/src/reynir/simpletree.py
+++ b/src/reynir/simpletree.py
@@ -138,6 +138,7 @@ class SimpleTreeNode(CanonicalTokenDict, total=False):
     "Afleiðing": "S-CONS",
     "Spurnarsetning": "S-QUE",
     "Sagt": "CP-QUOTE",
+    "SagtSpurning": "CP-QUOTE",
     # "Segjandi": "CP-SOURCE",
     "Forskeyti": "S-PREFIX",
     "Tíðarsetning": "CP-ADV-TEMP",
diff --git a/test/test_parse.py b/test/test_parse.py
index 30c1ab88..9c65dead 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2718,6 +2718,34 @@ def test_article_with_nominalized_adjective(r) -> None:
         assert phrase in s.tree.flat, (sentence, s.tree.flat)
 
 
+def test_quoted_question_attribution(r) -> None:
+    """A quoted question or exclamation followed directly by an
+    attribution ('spyr hann', 'sagði hún', 'bætir hún við') without
+    an intervening comma"""
+    for sentence, attr in (
+        ("„Komst hann ekki í hópinn?“ sagði hann.", "segja"),
+        ("„Við erum að falla á tíma!“ skrifar Limbourg.", "skrifa"),
+        ("„Áttu mynd?“ spyr sá þriðji.", "spyrja"),
+        ("„Er þetta verk dagskrárstjórans?“ spyr hann.", "spyrja"),
+        ("„Er það ekki bara flott fyrirsögn?“ svarar Ottó hlæjandi.", "svara"),
+        ("„Hvað er það sem Miðflokkurinn er að gera?“ bætir hún við.", "bæta"),
+        ("„Er þetta gott?“ sagði Sindri og bætti við: „Þetta er mjög gott.“", "segja"),
+        ("„Er þátturinn ekki að verða búinn?“ spurði Þorgerður og hló.", "hlæja"),
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        flat = s.tree.flat
+        assert flat.startswith("S0 S-QUOTE IP CP-QUOTE "), (sentence, flat)
+        assert attr in s.tree.verbs, sentence
+    # A proper noun as the attributed source is third person singular
+    s = r.parse_single("„Við erum að falla á tíma!“ skrifar Limbourg.")
+    assert s and s.tree
+    assert "VP so_0_gm_fh_et_p3 /VP sérnafn" in s.tree.flat
+    s = r.parse_single("„Þetta er gott,“ segir Limbourg.")
+    assert s and s.tree
+    assert "VP so_0_gm_fh_et_p3 /VP sérnafn" in s.tree.flat
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2784,4 +2812,5 @@ def test_article_with_nominalized_adjective(r) -> None:
     test_locative_i_prefers_dative(g)
     test_verb_with_clitic_subject(g)
     test_article_with_nominalized_adjective(g)
+    test_quoted_question_attribution(g)
     g.__class__.cleanup()

From 7ba6ea3893d55afeaf7466768ee98d971f2248dc Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 28 Aug 2026 16:22:05 +0000
Subject: [PATCH 08/10] =?UTF-8?q?Recognize=20"=C3=BEess=20=C3=AD=20sta?=
 =?UTF-8?q?=C3=B0"=20as=20a=20fixed=20adverbial=20phrase?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The static phrase "þess í stað" was commented out in Phrases.conf, so
the three tokens were analyzed separately ('þess' as a genitive
object, 'í stað' as a prepositional phrase), which rarely fits the
sentence; sentences such as "Þess í stað fór hann heim" or "Hann fór
þess í stað að læra" did not parse. Enable the phrase, which then
matches an 'ao' terminal like the other fixed adverbials.

Measured on the unparsed-sentence corpus: 171 of 250 sentences with
'þess í stað' now parse. No change on 1000 previously parsed
news sentences.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/config/Phrases.conf |  2 +-
 test/test_parse.py             | 18 ++++++++++++++++++
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/src/reynir/config/Phrases.conf b/src/reynir/config/Phrases.conf
index 7515f2ee..0df66e54 100644
--- a/src/reynir/config/Phrases.conf
+++ b/src/reynir/config/Phrases.conf
@@ -1365,7 +1365,7 @@ meaning = ao frasi -
 # "þess efnis", "fahee nhee", "sá efni"
 #"þess utan", "fphee ae", "það utan"
 #"þess vegna", "fphee ae", "það vegna"
-#"þess í stað", "fphee ao nkeo", "það í staður"
+"þess í stað", "fphee ao nkeo", "það í staður"
 "þvert á móti", "lhensf aa aa", "þver á móti"
 "því miður", "fpheþ aam", "það lítt"
 "fyrir slysni", "ao nveo", "fyrir slysni"
diff --git a/test/test_parse.py b/test/test_parse.py
index 9c65dead..21b34035 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -2746,6 +2746,23 @@ def test_quoted_question_attribution(r) -> None:
     assert "VP so_0_gm_fh_et_p3 /VP sérnafn" in s.tree.flat
 
 
+def test_thess_i_stad(r) -> None:
+    """'þess í stað' is a fixed adverbial phrase"""
+    for sentence in (
+        "Þess í stað fór hann heim.",
+        "Hann fór þess í stað í bæinn.",
+        "Þess í stað hafi flugfélagið fengið leiguvélar.",
+        "Það er ekki rétt, þess í stað er loftinu hleypt úr dekkjunum.",
+        "Hann gaf upp ferilinn og fór þess í stað að læra.",
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        assert any(
+            t.text.lower() == "þess í stað" and t.terminal == "ao"
+            for t in s.tree.leaves
+        ), sentence
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2813,4 +2830,5 @@ def test_quoted_question_attribution(r) -> None:
     test_verb_with_clitic_subject(g)
     test_article_with_nominalized_adjective(g)
     test_quoted_question_attribution(g)
+    test_thess_i_stad(g)
     g.__class__.cleanup()

From 81261d6b16bf04184a3d335e270168101e8d545d Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 28 Aug 2026 16:26:40 +0000
Subject: [PATCH 09/10] =?UTF-8?q?Parse=20the=20raising=20construction=20"X?=
 =?UTF-8?q?=20er=20sag=C3=B0ur/talinn=20hafa=20gert=20Y"?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The past participles sagður, talinn, álitinn and látinn followed by an
infinitive ("Maðurinn er sagður hafa stolið bílnum", "Hún er talin
vera í haldi") only had an analysis for "sögð vera + predicate", so
the common news pattern "er sagður hafa + supine" failed to parse
("sagður vera í haldi" happened to parse with 'vera' as a noun).

Add three SögnErLoBotn alternatives after SögnLhÞtVera: 'hafa' + a
supine verb phrase, a plain infinitive verb phrase, and vera/hafa
verið + a past participle, each with an optional adverb ("var ekki
talinn hafa misnotað aðstöðu sína").

Measured on the unparsed-sentence corpus: 241 of 369 sentences with
such phrases now parse. On 1000 previously parsed news sentences,
6 trees change (3 ties, 3 improved: "voru talin skekkja ímynd",
"var látinn vinna", "látin vita"), 0 lost.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar |  6 ++++++
 test/test_parse.py         | 24 +++++++++++++++++++++---
 2 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index 08b04ee4..56af4c2b 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -2031,6 +2031,12 @@ SögnErLoBotn/tala/kyn →
     | SögnErLoMM
     # (Hún er/var/hefur verið) talin/álitin/sögð vera lykillinn [að því að málið leystist]
     | SögnLhÞtVera_nf/tala/kyn HjSögnVera NlSagnfylling_nf
+    # (Maðurinn er/var) sagður/talinn [ekki] hafa stolið bílnum / hafa verið vopnaður
+    | SögnLhÞtVera_nf/tala/kyn FsAtv? 'hafa:so'_nh HreinSögn_sagnb/tala/kyn
+    # (Maðurinn er) sagður/talinn [ekki] vera í haldi lögreglu / búa í Reykjavík
+    | SögnLhÞtVera_nf/tala/kyn FsAtv? HreinSögn_nh/tala/kyn
+    # (Maðurinn er) sagður/talinn [ekki] vera/hafa verið vopnaður
+    | SögnLhÞtVera_nf/tala/kyn FsAtv? HjSögnEkkiVera so_lhþt_sb_nf/tala/kyn
 
 SögnErLoMM →
     FsAtv? so_0_mm_sagnb
diff --git a/test/test_parse.py b/test/test_parse.py
index 21b34035..77ec5a68 100644
--- a/test/test_parse.py
+++ b/test/test_parse.py
@@ -249,9 +249,8 @@ def test_parse(r: Greynir, verbose: bool = False) -> None:
         "borða",
     ]
     assert results[12].tree.verbs == ["horfa", "borða"]
-    assert (results[32].tree.verbs == ["finna", "segja", "gefa", "hafa", "deyja"]) or (
-        results[32].tree.verbs == ["finna", "gefa", "hafa", "deyja"]
-    )
+    # 'er sögð gefa í skyn' is a raising construction: vera + sögð + gefa
+    assert results[32].tree.verbs == ["finna", "vera", "segja", "gefa", "hafa", "deyja"]
     # Test that the parser finds the correct word lemmas
     assert results[0].tree.lemmas == [
         "hér",
@@ -2763,6 +2762,24 @@ def test_thess_i_stad(r) -> None:
         ), sentence
 
 
+def test_sagdur_hafa(r) -> None:
+    """Raising construction: X er sagður/talinn (ekki) hafa gert Y / vera Z"""
+    for sentence, verbs in (
+        ("Maðurinn er sagður hafa stolið bílnum.", ["vera", "segja", "hafa", "stela"]),
+        ("Konan er sögð hafa stolið bílnum.", ["vera", "segja", "hafa", "stela"]),
+        ("Mennirnir eru sagðir hafa stolið bílnum.", ["vera", "segja", "hafa", "stela"]),
+        ("Maðurinn er talinn hafa stolið bílnum.", ["vera", "telja", "hafa", "stela"]),
+        ("Því var hann ekki talinn hafa misnotað aðstöðu sína.", ["vera", "telja", "hafa", "misnota"]),
+        ("Hann er sagður vera í haldi lögreglu.", ["vera", "segja", "vera"]),
+        ("Hún er sögð hafa verið í haldi lögreglu.", ["vera", "segja", "hafa", "vera"]),
+        ("Hann er sagður búa í Reykjavík.", ["vera", "segja", "búa"]),
+    ):
+        s = r.parse_single(sentence)
+        assert s and s.tree, sentence
+        assert s.tree.verbs == verbs, (sentence, s.tree.verbs)
+        assert "lhþt" in s.tree.flat, sentence
+
+
 if __name__ == "__main__":
     # When invoked as a main module, do a verbose test
     from reynir import Greynir
@@ -2831,4 +2848,5 @@ def test_thess_i_stad(r) -> None:
     test_article_with_nominalized_adjective(g)
     test_quoted_question_attribution(g)
     test_thess_i_stad(g)
+    test_sagdur_hafa(g)
     g.__class__.cleanup()

From b58f67b9d85cc8768c8a24c8bef722ea1cade8c2 Mon Sep 17 00:00:00 2001
From: Vilhjalmur Thorsteinsson <vt@extrada.com>
Date: Fri, 28 Aug 2026 18:20:40 +0000
Subject: [PATCH 10/10] Factor shared prefixes out of the quote attribution
 productions
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Per review: move Forskeyti? and the optional comma up into the
Staðhæfing alternative, and share the attribution core (SegirKjarni)
between the comma-prefixed Segir and the question/exclamation case,
instead of repeating the prefixes in each alternative. No change in
parse results on the 1000-sentence baseline.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/reynir/Greynir.grammar | 21 ++++++++++-----------
 1 file changed, 10 insertions(+), 11 deletions(-)

diff --git a/src/reynir/Greynir.grammar b/src/reynir/Greynir.grammar
index 56af4c2b..7dbe6c2b 100644
--- a/src/reynir/Greynir.grammar
+++ b/src/reynir/Greynir.grammar
@@ -157,19 +157,15 @@ Staðhæfing →
     Sagt Segjandi ÁframSagt? Lokatákn?
     # Spurning eða upphrópun í beinni ræðu, án kommu á undan segjanda:
     # "Komst hann ekki í hópinn?" sagði hann. / "Við erum að falla á tíma!" skrifar Limbourg.
-    | SagtSpurning SegirÁnKommu SegirÁframSagt? Lokatákn?
+    | Forskeyti? SagtSpurning ","? SegirKjarni SegirÁframSagt? Lokatákn?
 
 # og bætti við: "Þetta er mjög gott."
 SegirÁframSagt →
     SegirÁfram ":"? Sagt?
 
 SagtSpurning →
-    Forskeyti? Upphrópun* Spurnarsetning
-    | Forskeyti? Yfirsetning StLiður* "!"
-
-SegirÁnKommu →
-    ","? SögnSegja/tala/pers Heimild_nf/tala/pers AtvFsAuka?
-    | ","? SögnBætaVið/tala/pers
+    Upphrópun* Spurnarsetning
+    | Yfirsetning StLiður* "!"
 
 ÁframSagt →
     Segir SegirÁfram Sagt
@@ -231,10 +227,13 @@ $tag(keep) Heimild_ef/tala/pers
 Sagt → Forskeyti? Yfirsetning StLiður* 
 
 Segir →
-    # , segir/fullyrðir Jón Ásmundsson, fulltrúi samtakanna, í dag í útvarpinu
-    "," SögnSegja/tala/pers Heimild_nf/tala/pers AtvFsAuka?
-    # , bætir Jón við
-    | "," SögnBætaVið/tala/pers
+    "," SegirKjarni
+
+SegirKjarni →
+    # segir/fullyrðir Jón Ásmundsson, fulltrúi samtakanna, í dag í útvarpinu
+    SögnSegja/tala/pers Heimild_nf/tala/pers AtvFsAuka?
+    # bætir Jón við
+    | SögnBætaVið/tala/pers
 
 # bætir Jón við / bætir hún við í færslu sinni
 SögnBætaVið/tala/pers →