Support for multilingual generative metrics (#293)

---------

Co-authored-by: Nathan Habib <30601243+NathanHB@users.noreply.github.com>
Co-authored-by: Hynek Kydlicek <kydliceh.hynek@gmail.com>
Co-authored-by: Clémentine Fourrier <22726840+clefourrier@users.noreply.github.com>

.github/workflows/tests.yaml

@@ -26,7 +26,7 @@ jobs:
          cache: 'pip'
      - name: Install lighteval in editable mode
        run: |
-         pip install -e .[dev,extended_tasks]
+         pip install -e .[dev,extended_tasks,multilingual]
      - name: Get cached files
        uses: actions/cache@v2
        id: get-cache

community_tasks/_template.py

@@ -33,7 +33,7 @@
 
 from lighteval.metrics import Metrics
 from lighteval.metrics.metrics import SampleLevelMetric
-from lighteval.metrics.utils import MetricCategory, MetricUseCase
+from lighteval.metrics.utils.metric_utils import MetricCategory, MetricUseCase
 from lighteval.tasks.default_prompts import LETTER_INDICES
 from lighteval.tasks.lighteval_task import LightevalTaskConfig
 from lighteval.tasks.requests import Doc

pyproject.toml

@@ -97,6 +97,12 @@ extended_tasks = [
   "openai", # llm as a judge using openai models
 ]
 s3 = ["s3fs"]
+multilingual = [
+    "stanza",
+    "spacy[ja,ko,th]",
+    "jieba", # for chinese tokenizer
+    "pyvi", # for vietnamese tokenizer
+]
 
 [project.urls]
 Homepage = "https://github.com/huggingface/lighteval"

src/lighteval/metrics/dynamic_metrics.py

@@ -20,13 +20,25 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
-from typing import Callable
+from typing import Callable, Literal
 
 import numpy as np
 
-from lighteval.metrics.metrics_sample import LoglikelihoodAcc, NormalizedMultiChoiceProbability, Probability
-from lighteval.metrics.normalizations import LogProbNormalization, LogProbPMINorm, LogProbTokenNorm
-from lighteval.metrics.utils import MetricCategory, MetricUseCase, SampleLevelMetric
+from lighteval.metrics.metrics_sample import (
+    ExactMatches,
+    F1_score,
+    LoglikelihoodAcc,
+    NormalizedMultiChoiceProbability,
+    Probability,
+)
+from lighteval.metrics.normalizations import (
+    LogProbNormalization,
+    LogProbPMINorm,
+    LogProbTokenNorm,
+    get_multilingual_normalizer,
+)
+from lighteval.metrics.utils.metric_utils import MetricCategory, MetricUseCase, SampleLevelMetric
+from lighteval.utils.language import Language
 
 
 def loglikelihood_acc_metric(normalization: LogProbNormalization | None = None) -> SampleLevelMetric:
@@ -92,3 +104,67 @@ def probability_metric(
         corpus_level_fn=np.mean,
         higher_is_better=True,
     )
+
+
+def multilingual_quasi_f1_score_metric(
+    language: Language, aggregation_function: Callable[[list[float]], float] = max
+) -> SampleLevelMetric:
+    """
+    Creates a language-aware F1 score metric, which returns the F1 score.
+
+    Args:
+        language: The language of the samples.
+        aggregation_function: Aggregation samples to use when multiple golds are present.
+
+    Returns:
+        F1 score metric.
+    """
+    metric_name = f"f1_{language}"
+
+    multilang_normalizer = get_multilingual_normalizer(language)
+    return SampleLevelMetric(
+        metric_name=metric_name,
+        sample_level_fn=F1_score(
+            normalize_gold=multilang_normalizer,
+            normalize_pred=multilang_normalizer,
+            aggregation_function=aggregation_function,
+        ).compute,
+        category=MetricCategory.GENERATIVE,
+        use_case=MetricUseCase.ACCURACY,
+        corpus_level_fn=np.mean,
+        higher_is_better=True,
+    )
+
+
+def multilingual_quasi_exact_match_metric(
+    language: Language,
+    match_type: Literal["prefix", "suffix", "full"] = "full",
+    aggregation_function: Callable[[list[float]], float] = max,
+) -> SampleLevelMetric:
+    """
+    Creates a language-aware exact match metric, which returns the exact match score
+    Args:
+        language: The language of the samples.
+        match_type: The type of match to use
+            - "prefix": Prefixes must match
+            - "suffix": Suffixes must match
+            - "full": Full strings must match
+        aggregation_function: Aggregation samples to use when multiple golds are present.
+    Returns:
+        Exact match metric.
+    """
+    metric_name = f"exact_match_{language}_{match_type}"
+    multilang_normalizer = get_multilingual_normalizer(language)
+    return SampleLevelMetric(
+        metric_name=metric_name,
+        sample_level_fn=ExactMatches(
+            normalize_gold=multilang_normalizer,
+            normalize_pred=multilang_normalizer,
+            aggregation_function=aggregation_function,
+            type_exact_match=match_type,
+        ).compute,
+        category=MetricCategory.GENERATIVE,
+        use_case=MetricUseCase.ACCURACY,
+        corpus_level_fn=np.mean,
+        higher_is_better=True,
+    )

src/lighteval/metrics/metrics.py

@@ -61,7 +61,7 @@
     remove_braces_and_strip,
 )
 from lighteval.metrics.sample_preparator import GenerativePreparator, LoglikelihoodPreparator, PerplexityPreparator
-from lighteval.metrics.utils import (
+from lighteval.metrics.utils.metric_utils import (
     CorpusLevelMetric,
     CorpusLevelMetricGrouping,
     Metric,

src/lighteval/metrics/metrics_sample.py

@@ -56,9 +56,9 @@
 class ExactMatches:
     def __init__(
         self,
-        aggregation_function: callable = None,
-        normalize_gold: callable = None,
-        normalize_pred: callable = None,
+        aggregation_function: Callable[[list[float]], float] = max,
+        normalize_gold: Callable[[str], str] | None = None,
+        normalize_pred: Callable[[str], str] | None = None,
         strip_strings: bool = False,
         type_exact_match: str = "full",
     ):
@@ -78,8 +78,6 @@ def __init__(
                 `suffix` if the prediction ends with the gold,
                 `full` if the prediction and gold are equal
         """
-        if aggregation_function is None:
-            aggregation_function = max
         self.aggregation_function = aggregation_function
         self.normalize_gold = normalize_gold
         self.normalize_pred = normalize_pred
@@ -145,9 +143,9 @@ def compute_one_item(
 class F1_score:
     def __init__(
         self,
-        aggregation_function: callable = None,
-        normalize_gold: callable = None,
-        normalize_pred: callable = None,
+        aggregation_function: Callable[[list[float]], float] = max,
+        normalize_gold: Callable[[str], str] | None = None,
+        normalize_pred: Callable[[str], str] | None = None,
         strip_strings: bool = False,
     ):
         """An F1 score class. F1 is computed over the bag of words of the golds and predictions.
@@ -163,8 +161,8 @@ def __init__(
         """
         if aggregation_function is None:
             aggregation_function = max
-        self.aggregation_function = aggregation_function
 
+        self.aggregation_function = aggregation_function
         self.normalize_gold = normalize_gold
         self.normalize_pred = normalize_pred
         self.strip_strings = strip_strings

src/lighteval/metrics/normalizations.py

@@ -22,7 +22,13 @@
 
 import re
 import string
+import sys
+import unicodedata
 from dataclasses import dataclass
+from typing import Callable
+
+from lighteval.metrics.utils.linguistic_tokenizers import get_word_tokenizer
+from lighteval.utils.language import Language
 
 
 # From HELM
@@ -355,6 +361,61 @@ def gsm8k_normalizer(text: str) -> str:
         return INVALID_ANS
 
 
+PUNCT = {chr(i) for i in range(sys.maxunicode) if unicodedata.category(chr(i)).startswith("P")}.union(
+    string.punctuation
+)
+
+_ARTICLE_PATTERNS = {
+    Language.ENGLISH: r"\b(a|an|the)\b",
+    Language.SPANISH: r"\b(el|la|los|las|un|una|unos|unas)\b",
+    Language.PORTUGUESE: r"\b(o|a|os|as|um|uma|uns|umas)\b",
+    Language.ITALIAN: r"\b(il|lo|la|i|gli|le|un|uno|una)\b",
+    Language.FRENCH: r"\b(le|la|les|l'|un|une|des)\b",
+    Language.GERMAN: r"\b(der|die|das|den|dem|des|ein|eine|einer|eines|einem|einen)\b",
+    Language.FINNISH: r"\b(se|yksi|yks)\b",
+    Language.GREEK: r"\b(ὁ|οἱ|τοῦ|τῶν|τόν|τούς|ὦ|ἡ|αἱ|τῆς|τῶν|τήν|τάς|τό|τά|τοῦ|τῶν|τό|τά)\b",
+    Language.NORWEGIAN: r"\b(en|ei|et|den|det|de)\b",
+    Language.SWEDISH: r"\b(en|ett|den|det|de)\b",
+    Language.TURKISH: r"\b(bir)\b",
+    Language.DUTCH: r"\b(de|het|een)\b",
+    Language.HUNGARIAN: r"\b(a|az|egy)\b",
+    Language.CATALAN: r"\b(el|la|els|les|un|una|uns|unes)\b",
+    Language.HEBREW: r"\b(ה)\b",
+    Language.GALICIAN: r"\b(o|a|os|as|un|unha|uns|unhas)\b",
+}
+
+
+def remove_articles(text: str, lang: Language) -> str:
+    """
+    Removes definite and indefinite articles from the text.
+    Generated using LLM then manually checked by non-expert.
+    We currently only support languages that don't blend articles.
+    If you are a native speaker of a language where articles are blended,
+    we would appreciate your contribution!
+    """
+    pattern = _ARTICLE_PATTERNS.get(lang)
+    return re.sub(pattern, " ", text) if pattern else text
+
+
+def remove_punc(text: str) -> str:
+    return "".join(ch for ch in text if ch not in PUNCT)
+
+
+def get_multilingual_normalizer(lang: Language, lower: bool = True) -> Callable[[str], str]:
+    tokenizer = get_word_tokenizer(lang)
+
+    def _inner_normalizer(text: str) -> str:
+        text = remove_articles(text, lang)
+        text = remove_punc(text)
+        if lower:
+            text = text.lower()
+
+        tokens = tokenizer.word_tokenize(text)
+        return " ".join(tokens)
+
+    return _inner_normalizer
+
+
 # Loglikelihood normalization
 @dataclass
 class LogProbPMINorm: