feat: better evaluation scripts

.github/workflows/tests.yml

@@ -69,6 +69,10 @@ jobs:
       if: matrix.python-version != '3.6' && matrix.python-version != '3.7'
       run: pip install -r requirements-dev.txt
 
+    - name: Install training dependencies
+      if: matrix.python-version != '3.6' && matrix.python-version != '3.7'
+      run: pip install -r training/requirements.txt
+
     - name: Install dependencies (legacy versions)
       if: matrix.python-version == '3.6' || matrix.python-version == '3.7'
       run: |

.gitignore

@@ -116,3 +116,4 @@ Makefile
 
 # eval
 UD/
+training/**/data

simplemma/language_detector.py

@@ -198,18 +198,14 @@ def proportion_in_target_languages(
         Returns:
             float: The proportion of text in the target language(s).
         """
-        tokens = self._token_sampler.sample_text(text)
-        if len(tokens) == 0:
-            return 0
-
-        in_target = 0
-        for token in tokens:
-            for lang_code in self._lang:
-                candidate = self._lemmatization_strategy.get_lemma(token, lang_code)
-                if candidate is not None:
-                    in_target += 1
-                    break
-        return in_target / len(tokens)
+        return sum(
+            percentage
+            for (
+                lang_code,
+                percentage,
+            ) in self.proportion_in_each_language(text).items()
+            if lang_code != "unk"
+        )
 
     def main_language(
         self,

tests/test_language_detector.py

@@ -108,15 +108,6 @@ def test_in_target_language() -> None:
         == 1.0
     )
 
-    langs = ("en", "de")
-    text = "It was a true gift"
-    assert (
-        LanguageDetector(lang=langs).proportion_in_target_languages(text)
-        == in_target_language(text, lang=langs)
-        == 1.0
-    )
-    in_target_language("It was a true gift", lang=("en", "de"))
-
 
 def test_main_language():
     text = "Dieser Satz ist auf Deutsch."

training/README.rst

@@ -0,0 +1,15 @@
+Instructions to run the evaluation
+----------------------------------
+
+The scores are calculated on `Universal Dependencies <https://universaldependencies.org/>`_ treebanks on single word tokens (including some contractions but not merged prepositions). They can be reproduced by the following steps:
+
+1. Install the evaluation dependencies (``pip install -r training/requirements.txt``)
+2. Update ``DATA_URL`` in ``training/download-eval-data.py`` to point to the latest treebanks archive from `Universal Dependencies <https://universaldependencies.org/#download>` (or the version that you which to use).
+3. Run ``python3 training/download-eval-data.py`` which will
+  1. Download the archive
+  2. Extract relevant data (language and if applicable specific treebank, see notes in the results table)
+  3. Concatenate the train, dev and test data into a single file (e.g. ``cat de_gsd*.conllu > de-gsd-all.conllu``)
+  4. Store the files at the expected location (``training/data/UD/``)
+4. Run the script, e.g. from the home directory ``python3 training/evaluate_simplema.py``
+5. Results are stored at ``training/data/results/results_summary.csv``. Also, errors are written in a CSV file for each dataset under the ``data/results``folder. 
+

training/download-eval-data.py

@@ -0,0 +1,71 @@
+from typing import Iterable, List, Tuple
+from os import mkdir, path, scandir
+import re
+import logging
+import tarfile
+import requests
+from glob import glob
+
+from simplemma.strategies.dictionaries.dictionary_factory import SUPPORTED_LANGUAGES
+
+log = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+DATA_URL = "https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-5150/ud-treebanks-v2.12.tgz?sequence=1&isAllowed=y"
+DATA_FOLDER = path.join(path.dirname(__file__), "data")
+DATA_FILE = path.join(DATA_FOLDER, "ud-treeebanks.tgz")
+CLEAN_DATA_FOLDER = path.join(DATA_FOLDER, "UD")
+
+
+def get_dirs(file_name: str) -> List[str]:
+    return [dir.name for dir in scandir(file_name) if dir.is_dir()]
+
+
+def get_files(file_name: str) -> List[str]:
+    return [dir.name for dir in scandir(file_name) if dir.is_file()]
+
+
+def get_relevant_language_data_folders(data_folder) -> Iterable[Tuple[str, str, str]]:
+    for lang_folder in get_dirs(data_folder):
+        lang_data_folder = path.join(uncompressed_data_folder, lang_folder)
+        conllu_file = glob(path.join(lang_data_folder, "*.conllu"))[0]
+        matches_files = re.search("^.*/(.*)-ud.*$", conllu_file)
+        if matches_files is not None:
+            dataset_name = matches_files.groups()[0]
+            lang = dataset_name.split("_")[0]
+
+            if lang in SUPPORTED_LANGUAGES:
+                yield (lang, dataset_name, lang_data_folder)
+
+
+if path.exists(DATA_FOLDER) or path.exists(CLEAN_DATA_FOLDER):
+    raise Exception(
+        "Data folder seems to be already present. Delete it before creating new data."
+    )
+
+mkdir(DATA_FOLDER)
+mkdir(CLEAN_DATA_FOLDER)
+
+log.info("Downloading evaluation data...")
+response = requests.get(DATA_URL)
+open(DATA_FILE, "wb").write(response.content)
+
+log.info("Uncompressing evaluation data...")
+with tarfile.open(DATA_FILE) as tar:
+    tar.extractall(DATA_FOLDER)
+uncompressed_data_folder = path.join(
+    DATA_FOLDER, glob(f"{DATA_FOLDER}/ud-treebanks-*")[0]
+)
+
+log.info("Filtering files...")
+for lang, dataset_name, dataset__folder in get_relevant_language_data_folders(
+    uncompressed_data_folder
+):
+    log.info(lang + " - " + dataset__folder)
+    # Concatenate the train, dev and test data into a single file (e.g. ``cat de_gsd*.conllu > de-gsd-all.conllu``)
+    lang_clean_data_file = path.join(CLEAN_DATA_FOLDER, f"{dataset_name}.conllu")
+    log.debug(f"Procressing data for {dataset_name}")
+    with open(lang_clean_data_file, "w") as outfile:
+        for file in glob(path.join(dataset__folder, "*.conllu")):
+            with open(file) as infile:
+                for line in infile:
+                    outfile.write(line)