Added initial setup for SEM-F1 metric

README.md

@@ -1,28 +1,49 @@
 ---
 title: SemF1
-datasets:
--  
 tags:
 - evaluate
 - metric
-description: "TODO: add a description here"
 sdk: gradio
 sdk_version: 3.19.1
 app_file: app.py
 pinned: false
+description: >-
+  SEM-F1 metric leverages the pre-trained contextual embeddings and evaluates the model generated semantic overlap 
+  summary with the reference overlap summary. It evaluates the semantic overlap summary at the sentence level and 
+  computes precision, recall and F1 scores.
+  
+  Refer to the paper `SEM-F1: an Automatic Way for Semantic Evaluation of Multi-Narrative Overlap Summaries at Scale` 
+  for more details. 
 ---
 
 # Metric Card for SemF1
 
-***Module Card Instructions:*** *Fill out the following subsections. Feel free to take a look at existing metric cards if you'd like examples.*
-
 ## Metric Description
-*Give a brief overview of this metric, including what task(s) it is usually used for, if any.*
+SEM-F1 metric leverages the pre-trained contextual embeddings and evaluates the model generated semantic overlap 
+summary with the reference overlap summary. It evaluates the semantic overlap summary at the sentence level and 
+computes precision, recall and F1 scores.
 
 ## How to Use
-*Give general statement of how to use the metric*
+SEM-F1 takes 2 mandatory arguments: 
+    `predictions`: (a list of system generated documents in the form of sentences i.e. List[List[str]]),
+    `references`: (a list of ground-truth documents in the form of sentences i.e. List[List[str]])
+
+```python
+from evaluate import load
+predictions = [
+    ["I go to School.", "You are stupid."],
+    ["I love adventure sports."],
+]
+references = [
+    ["I go to School.", "You are stupid."],
+    ["I love adventure sports."],
+]
+metric = load("semf1")
+results = metric.compute(predictions=predictions, references=references)
+```
 
-*Provide simplest possible example for using the metric*
+It also accepts multiple optional arguments:
+TODO: List optional arguments
 
 ### Inputs
 *List all input arguments in the format below*
@@ -44,7 +65,27 @@ pinned: false
 *Note any known limitations or biases that the metric has, with links and references if possible.*
 
 ## Citation
-*Cite the source where this metric was introduced.*
+```bibtex
+@inproceedings{bansal-etal-2022-sem,
+    title = "{SEM}-F1: an Automatic Way for Semantic Evaluation of Multi-Narrative Overlap Summaries at Scale",
+    author = "Bansal, Naman  and
+      Akter, Mousumi  and
+      Karmaker Santu, Shubhra Kanti",
+    editor = "Goldberg, Yoav  and
+      Kozareva, Zornitsa  and
+      Zhang, Yue",
+    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
+    month = dec,
+    year = "2022",
+    address = "Abu Dhabi, United Arab Emirates",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2022.emnlp-main.49",
+    doi = "10.18653/v1/2022.emnlp-main.49",
+    pages = "780--792",
+    abstract = "Recent work has introduced an important yet relatively under-explored NLP task called Semantic Overlap Summarization (SOS) that entails generating a summary from multiple alternative narratives which conveys the common information provided by those narratives. Previous work also published a benchmark dataset for this task by collecting 2,925 alternative narrative pairs from the web and manually annotating 411 different reference summaries by engaging human annotators. In this paper, we exclusively focus on the automated evaluation of the SOS task using the benchmark dataset. More specifically, we first use the popular ROUGE metric from text-summarization literature and conduct a systematic study to evaluate the SOS task. Our experiments discover that ROUGE is not suitable for this novel task and therefore, we propose a new sentence-level precision-recall style automated evaluation metric, called SEM-F1 (Semantic F1). It is inspired by the benefits of the sentence-wise annotation technique using overlap labels reported by the previous work. Our experiments show that the proposed SEM-F1 metric yields a higher correlation with human judgment and higher inter-rater agreement compared to the ROUGE metric.",
+}
+```
 
 ## Further References
-*Add any useful further references.*
+ - [Paper](https://aclanthology.org/2022.emnlp-main.49/)
+ - [Presentation Slides]()

requirements.txt

@@ -1 +1,3 @@
-git+https://github.com/huggingface/evaluate@main
+git+https://github.com/huggingface/evaluate@main
+scikit-learn
+sentence-transformers

semf1.py

@@ -11,58 +11,137 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""TODO: Add a description here."""
+# TODO: Add test cases
+"""SEM-F1 metric"""
 
-import evaluate
-import datasets
 
+import abc
+import sys
+from typing import List, Optional, Tuple
+
+import datasets
+import evaluate
+import numpy as np
+from numpy.typing import NDArray
+from sentence_transformers import SentenceTransformer
+from sklearn.metrics.pairwise import cosine_similarity
 
-# TODO: Add BibTeX citation
 _CITATION = """\
-@InProceedings{huggingface:module,
-title = {A great new module},
-authors={huggingface, Inc.},
-year={2020}
+@inproceedings{bansal-etal-2022-sem,
+    title = "{SEM}-F1: an Automatic Way for Semantic Evaluation of Multi-Narrative Overlap Summaries at Scale",
+    author = "Bansal, Naman  and
+      Akter, Mousumi  and
+      Karmaker Santu, Shubhra Kanti",
+    editor = "Goldberg, Yoav  and
+      Kozareva, Zornitsa  and
+      Zhang, Yue",
+    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
+    month = dec,
+    year = "2022",
+    address = "Abu Dhabi, United Arab Emirates",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2022.emnlp-main.49",
+    doi = "10.18653/v1/2022.emnlp-main.49",
+    pages = "780--792",
+    abstract = "Recent work has introduced an important yet relatively under-explored NLP task called Semantic Overlap Summarization (SOS) that entails generating a summary from multiple alternative narratives which conveys the common information provided by those narratives. Previous work also published a benchmark dataset for this task by collecting 2,925 alternative narrative pairs from the web and manually annotating 411 different reference summaries by engaging human annotators. In this paper, we exclusively focus on the automated evaluation of the SOS task using the benchmark dataset. More specifically, we first use the popular ROUGE metric from text-summarization literature and conduct a systematic study to evaluate the SOS task. Our experiments discover that ROUGE is not suitable for this novel task and therefore, we propose a new sentence-level precision-recall style automated evaluation metric, called SEM-F1 (Semantic F1). It is inspired by the benefits of the sentence-wise annotation technique using overlap labels reported by the previous work. Our experiments show that the proposed SEM-F1 metric yields a higher correlation with human judgment and higher inter-rater agreement compared to the ROUGE metric.",
 }
 """
 
-# TODO: Add description of the module here
 _DESCRIPTION = """\
-This new module is designed to solve this great ML task and is crafted with a lot of care.
+SEM-F1 metric leverages the pre-trained contextual embeddings and evaluates the model generated 
+semantic overlap summary with the reference overlap summary. It evaluates the semantic overlap summary at the 
+sentence level and computes precision, recall and F1 scores.
 """
 
-
-# TODO: Add description of the arguments of the module here
 _KWARGS_DESCRIPTION = """
-Calculates how good are predictions given some references, using certain scores
+SEM-F1 compares the system generated overlap summary with ground truth reference overlap.
+
 Args:
-    predictions: list of predictions to score. Each predictions
-        should be a string with tokens separated by spaces.
-    references: list of reference for each prediction. Each
+    predictions: List[List(str)] - List of  predictions where each prediction is a list of sentences.
+    references: List[List(str)] - List of references where each reference is a list of sentences.
         reference should be a string with tokens separated by spaces.
+    model_type: str - Model to use. [pv1, stsb, use] 
+        Options: 
+            pv1 - paraphrase-distilroberta-base-v1 
+            stsb - stsb-roberta-large 
+            use - Universal Sentence Encoder
 Returns:
     accuracy: description of the first score,
     another_score: description of the second score,
 Examples:
-    Examples should be written in doctest format, and should illustrate how
-    to use the function.
 
-    >>> my_new_module = evaluate.load("my_new_module")
-    >>> results = my_new_module.compute(references=[0, 1], predictions=[0, 1])
-    >>> print(results)
-    {'accuracy': 1.0}
+    >>> import evaluate
+    >>> predictions = [
+        ["I go to School.", "You are stupid."],
+        ["I love adventure sports."],
+    ]
+    >>> references = [
+        ["I go to School.", "You are stupid."],
+        ["I love adventure sports."],
+    ]
+    >>> metric = evaluate.load("semf1")
+    >>> results = metric.compute(predictions=predictions, references=references)
+    >>> print([round(v, 2) for v in results["f1"]])
+    [0.77, 0.56]
 """
 
-# TODO: Define external resources urls if needed
-BAD_WORDS_URL = "http://url/to/external/resource/bad_words.txt"
+
+class Encoder(metaclass=abc.ABCMeta):
+    @abc.abstractmethod
+    def encode(self, prediction: List[str]) -> NDArray:
+        pass
+
+
+class USE(Encoder):
+    def __init__(self):
+        pass
+
+    def encode(self, prediction: List[str]) -> NDArray:
+        pass
+
+
+class SBertEncoder(Encoder):
+    def __init__(self, model_name: str):
+        self.model = SentenceTransformer(model_name)
+
+    def encode(self, prediction: List[str]) -> NDArray:
+        return self.model.encode(prediction)
+
+
+def _get_encoder(model_name: str):
+    if model_name == "use":
+        return USE()
+    else:
+        return SBertEncoder(model_name)
+
+
+def _compute_f1(p, r, eps=sys.float_info.epsilon):
+    '''
+    Computes F1 value
+    :param p: Precision Value
+    :param r: Recall Value
+    :return:
+    '''
+    f1 = 2 * p * r / (p + r + eps)
+    return f1
+
+
+def _compute_cosine_similarity(pred_embeds: NDArray, ref_embeds: NDArray) -> Tuple[float, float]:
+    cosine_scores = cosine_similarity(pred_embeds, ref_embeds)
+    precision_per_sentence_sim = np.max(cosine_scores, axis=-1)
+    recall_per_sentence_sim = np.max(cosine_scores, axis=0)
+    return np.mean(precision_per_sentence_sim).item(), np.mean(recall_per_sentence_sim).item()
 
 
 @evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
 class SemF1(evaluate.Metric):
-    """TODO: Short description of my evaluation module."""
+    _MODEL_TYPE_TO_NAME = {
+        "pv1": "paraphrase-distilroberta-base-v1",
+        "stsb": "stsb-roberta-large",
+        "use": "use",
+    }
 
     def _info(self):
-        # TODO: Specifies the evaluate.EvaluationModuleInfo object
         return evaluate.MetricInfo(
             # This is the description that will appear on the modules page.
             module_type="metric",
@@ -71,25 +150,41 @@ class SemF1(evaluate.Metric):
             inputs_description=_KWARGS_DESCRIPTION,
             # This defines the format of each prediction and reference
             features=datasets.Features({
-                'predictions': datasets.Value('int64'),
-                'references': datasets.Value('int64'),
+                'predictions': datasets.Sequence(datasets.Value("string", id="sequence"), id="predictions"),
+                'references': datasets.Sequence(datasets.Value("string", id="sequence"), id="references"),
             }),
-            # Homepage of the module for documentation
-            homepage="http://module.homepage",
+            # # Homepage of the module for documentation
             # Additional links to the codebase or references
-            codebase_urls=["http://github.com/path/to/codebase/of/new_module"],
-            reference_urls=["http://path.to.reference.url/new_module"]
+            reference_urls=["https://aclanthology.org/2022.emnlp-main.49/"]
         )
 
-    def _download_and_prepare(self, dl_manager):
-        """Optional: download external resources useful to compute the scores"""
-        # TODO: Download external resources if needed
-        pass
+    def _get_model_name(self, model_type: Optional[str] = None) -> str:
+        # TODO: make it work with USE as well
+        if model_type not in self._MODEL_TYPE_TO_NAME.keys():
+            raise ValueError(f"Provide a correct model_type.\n"
+                             f"Options: {self._MODEL_TYPE_TO_NAME.keys()}\n"
+                             f"Currently provided: {model_type}")
+
+        if model_type is None:
+            model_type = "pv1"  # Change it to use
+
+        return self._MODEL_TYPE_TO_NAME[model_type]
+
+    def _compute(self, predictions, references, model_type: Optional[str] = None):
+        model_name = self._get_model_name(model_type)
+        encoder = _get_encoder(model_name)
+
+        precisions = [0] * len(predictions)
+        recalls = [0] * len(predictions)
+        f1_scores = [0] * len(predictions)
+
+        for idx, (preds, refs) in enumerate(zip(predictions, references)):
+            pred_embeddings = encoder.encode(preds)
+            ref_embeddings = encoder.encode(refs)
+            p, r = _compute_cosine_similarity(pred_embeddings, ref_embeddings)
+            f1 = _compute_f1(p, r)
+            precisions[idx] = p
+            recalls[idx] = r
+            f1_scores[idx] = f1
 
-    def _compute(self, predictions, references):
-        """Returns the scores"""
-        # TODO: Compute the different scores of the module
-        accuracy = sum(i == j for i, j in zip(predictions, references)) / len(predictions)
-        return {
-            "accuracy": accuracy,
-        }
+        return {"precision": precisions, "recall": recalls, "f1": f1_scores}