Add a pos-specific charlm map for the medical EN datasets and the one dataset which appears to be hurt by the charlm (tr_boun)pos_charlm

craft, genia -> None

3 files changed, 17 insertions, 3 deletions

diff --git a/stanza/resources/prepare_resources.py b/stanza/resources/prepare_resources.py
index 2facd144..98007a10 100644
--- a/stanza/resources/prepare_resources.py
+++ b/stanza/resources/prepare_resources.py
@@ -177,6 +177,18 @@ default_charlms = {
     "zh-hans": "gigaword"
 }
 
+pos_charlms = {
+    "en": {
+        # none of the English charlms help with craft or genia
+        "craft": None,
+        "genia": None,
+        "mimic": "mimic",
+    },
+    "tr": {   # no idea why, but this particular one goes down in dev score
+        "boun": None,
+    },
+}
+
 ner_charlms = {
     "en": {
         "conll03": "1billion",
diff --git a/stanza/utils/training/common.py b/stanza/utils/training/common.py
index 0969d20a..901c9662 100644
--- a/stanza/utils/training/common.py
+++ b/stanza/utils/training/common.py
@@ -336,7 +336,9 @@ def choose_charlm(language, dataset, charlm, language_charlms, dataset_charlms):
         return None
     elif charlm != "default":
         return charlm
-    elif specific_charlm:
+    elif dataset in dataset_charlms.get(language, {}):
+        # this way, a "" or None result gets honored
+        # thus treating "not in the map" as a way for dataset_charlms to signal to use the default
         return specific_charlm
     elif default_charlm:
         return default_charlm
diff --git a/stanza/utils/training/run_pos.py b/stanza/utils/training/run_pos.py
index 2afe76b6..6e5ede44 100644
--- a/stanza/utils/training/run_pos.py
+++ b/stanza/utils/training/run_pos.py
@@ -9,7 +9,7 @@ from stanza.resources.prepare_resources import no_pretrain_languages
 from stanza.utils.training import common
 from stanza.utils.training.common import Mode, add_charlm_args, build_charlm_args, choose_charlm
 
-from stanza.resources.prepare_resources import default_charlms
+from stanza.resources.prepare_resources import default_charlms, pos_charlms
 
 logger = logging.getLogger('stanza')
 
@@ -46,7 +46,7 @@ def run_treebank(mode, paths, treebank, short_name,
     test_gold_file = f"{pos_dir}/{short_name}.test.gold.conllu"
     test_pred_file = temp_output_file if temp_output_file else f"{pos_dir}/{short_name}.test.pred.conllu"
 
-    charlm = choose_charlm(short_language, dataset, command_args.charlm, default_charlms, {})
+    charlm = choose_charlm(short_language, dataset, command_args.charlm, default_charlms, pos_charlms)
     charlm_args = build_charlm_args(short_language, charlm)
 
     if mode == Mode.TRAIN: