Update setup script (#4)

- address script documentation and add reference by grader (e968adefbca584126c4b4fc2c0e8f4d4c7a67ecb)
- update gitignore (ac78acc4d75fad46672f51163444b4c9abfc34f2)
- add uv refence (1fdeff28295bfef1616b594b2f025a1f96b85dee)
- add reference columns to all splits (3cf44d9b50cdc889e7c17997361645d241241f91)
- update readme (711ca56fe85c185fe6731a6b2e534044d271eab2)
- update readme (bc306613b7b163651c59ed1d8490926c6161ce8a)

.gitignore

@@ -0,0 +1,2 @@
+uv.lock
+.python-version

README.md

@@ -16,7 +16,7 @@ size_categories:
 ## Dataset Description
 - **Purpose**: Automated Essay Scoring
 - **Contents**: Student Essay Grades
-- **Source**: https://github.com/kamel-usp/aes_enem
+- **Source**: https://huggingface.co/datasets/kamel-usp/aes_enem_dataset
 - **Size**: N<1000
 
 ## Use Case and Creators
@@ -70,10 +70,11 @@ size_categories:
   - sourceAOnly: sourceA data
   - sourceAWithGraders: sourceA data augmented with Grader's review. In a nutshell, each row becomes three (the original grade plus two graders result)
   - sourceB: sourceB data
+  - PROPOR2024: Same split used in PROPOR2024 paper. The others are updated and fix some tiny bugs (eg reproducilibity issues)
 ## Data Considerations
 - **Known Limitations**:
 - **Ethical Considerations**:
 
 ## Additional Information
-- **Additional Links**: Main code is [here](https://github.com/kamel-usp/aes_enem)
+- **Additional Links**: Main code is [here](https://huggingface.co/datasets/kamel-usp/aes_enem_dataset)
 - **Related Datasets**: https://github.com/evelinamorim/aes-pt

aes_enem_dataset.py

@@ -11,9 +11,6 @@
 # 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: Address all TODOs and remove all explanatory comments
-"""TODO: Add a description here."""
-
 
 import csv
 import math
@@ -28,16 +25,45 @@ from tqdm.auto import tqdm
 
 np.random.seed(42)  # Set the seed
 
-# TODO: Add BibTeX citation
-# Find for instance the citation on arxiv or on the dataset repo/website
-_CITATION = """\
-TODO
+_CITATION = """
+@inproceedings{silveira-etal-2024-new,
+    title = "A New Benchmark for Automatic Essay Scoring in {P}ortuguese",
+    author = "Silveira, Igor Cataneo  and
+      Barbosa, Andr{\'e}  and
+      Mau{\'a}, Denis Deratani",
+    editor = "Gamallo, Pablo  and
+      Claro, Daniela  and
+      Teixeira, Ant{\'o}nio  and
+      Real, Livy  and
+      Garcia, Marcos  and
+      Oliveira, Hugo Goncalo  and
+      Amaro, Raquel",
+    booktitle = "Proceedings of the 16th International Conference on Computational Processing of Portuguese - Vol. 1",
+    month = mar,
+    year = "2024",
+    address = "Santiago de Compostela, Galicia/Spain",
+    publisher = "Association for Computational Lingustics",
+    url = "https://aclanthology.org/2024.propor-1.23/",
+    pages = "228--237"
+}
 """
 
-# TODO: Add description of the dataset here
-# You can copy an official description
 _DESCRIPTION = """\
-This new dataset is designed to solve this great NLP task and is crafted with a lot of care.
+This dataset was created as part of our work on advancing Automatic Essay Scoring for 
+Brazilian Portuguese. It comprises a large collection of publicly available essays 
+collected from websites simulating University Entrance Exams, with a subset expertly 
+annotated to provide reliable assessment indicators. The dataset includes both the raw 
+text and processed forms of the essays, along with supporting prompts and supplemental 
+texts.
+
+Key Features:
+    - A diverse corpus of essays with detailed annotations.
+    - A subset graded by expert annotators to evaluate essay quality and task difficulty.
+    - Comprehensive metadata providing provenance and context for each essay.
+    - An empirical analysis framework to support state-of-the-art predictive modeling.
+
+For further details, please refer to the paper “A New Benchmark for Automatic Essay 
+Scoring in Portuguese” available at https://aclanthology.org/2024.propor-1.23/.
 """
 
 # TODO: Add a link to an official homepage for the dataset here
@@ -81,6 +107,7 @@ CSV_HEADER = [
     "general",
     "specific",
     "essay_year",
+    "reference"
 ]
 
 CSV_HEADERPROPOR = [
@@ -90,35 +117,53 @@ CSV_HEADERPROPOR = [
     "essay",
     "grades",
     "essay_year",
+    "reference"
 ]
 
 SOURCE_A_DESC = """
-Source A have 860 essays available from August 2015 to March 2020.
-For each month of that period, a new prompt together with supporting texts were given, and the graded essays from the previous month were made available.
+SourceA have 860 essays available from August 2015 to March 2020.
+For each month of that period, a new prompt together with supporting texts were given, 
+and the graded essays from the previous month were made available.
 Of the 56 prompts, 12 had no associated essays available (at the time of download).
-Additionally, there were 3 prompts that asked for a text in the format of a letter. We removed those 15 prompts and associated texts from the corpus.
-For an unknown reason, 414 of the essays were graded using a five-point scale of either {0, 50, 100, 150, 200} or its scaled-down version going from 0 to 2.
-To avoid introducing bias, we also discarded such instances, resulting in a dataset of 386 annotated essays with prompts and supporting texts (with each component being clearly identified).
-Some of the essays used a six-point scale with 20 points instead of 40 points as the second class. As we believe this introduces minimal bias, we kept such essays and relabeled class 20 as class 40.
-The original data contains comments from the annotators explaining their per-competence scores. They are included in our dataset.
+Additionally, there were 3 prompts that asked for a text in the format of a letter. 
+We removed those 15 prompts and associated texts from the corpus.
+For an unknown reason, 414 of the essays were graded using a five-point scale of either 
+{0, 50, 100, 150, 200} or its scaled-down version going from 0 to 2.
+To avoid introducing bias, we also discarded such instances, resulting in a dataset of 
+386 annotated essays with prompts and supporting texts (with each component being clearly identified).
+Some of the essays used a six-point scale with 20 points instead of 40 points as the second class. 
+As we believe this introduces minimal bias, we kept such essays and relabeled class 20 as class 40.
+The original data contains comments from the annotators explaining their per-competence scores. 
+They are included in our dataset.
 """
 
-SOURCE_A_WITH_GRADERS = "Same as SourceA but augmented with reviwers contractors grade's. Each essay then have three grades: the downloaded one and each grader's feedback. "
+SOURCE_A_WITH_GRADERS = """
+sourceAWithGraders includes the original dataset augmented with grades from additional reviewers.
+Each essay is replicated three times:
+1. The original essay with its grades from the website.
+2. The same essay with grades from the first human grader.
+3. The same essay with grades from the second human grader.
+"""
 
 SOURCE_B_DESC = """
-Source B is very similar to Source A: a new prompt and supporting texts are made available every month along with the graded essays submitted in the previous month.
-We downloaded HTML sources from 7,700 essays from May 2009 to May 2023. Essays released prior to June 2016 were graded on a five-point scale and consequently discarded.
+SourceB is very similar to Source A: a new prompt and supporting texts are made 
+available every month along with the graded essays submitted in the previous month.
+We downloaded HTML sources from 7,700 essays from May 2009 to May 2023. Essays released 
+prior to June 2016 were graded on a five-point scale and consequently discarded.
 This resulted in a corpus of approx. 3,200 graded essays on 83 different prompts.
 
-Although in principle, Source B also provides supporting texts for students, none were available at the time the data was downloaded.
-To mitigate this, we extracted supporting texts from the Essay-Br corpus, whenever possible, by manually matching prompts between the two corpora.
-We ended up with approx. 1,000 essays containing both prompt and supporting texts, and approx. 2,200 essays containing only the respective prompt.
+Although in principle, Source B also provides supporting texts for students, none were 
+available at the time the data was downloaded.
+To mitigate this, we extracted supporting texts from the Essay-Br corpus, whenever 
+possible, by manually matching prompts between the two corpora.
+We ended up with approx. 1,000 essays containing both prompt and supporting texts, and 
+approx. 2,200 essays containing only the respective prompt.
 """
 
 PROPOR2024 = """
-Splits used for PROPOR paper. It is a variation of sourceAWithGraders dataset. Post publication we noticed that there was an issue in the reproducible setting.
-
-We fix that and set this config to keep reproducibility w.r.t. numbers reported in the paper.
+This split corresponds to the results reported in the PROPOR 2024 paper. While reproducibility was 
+fixed in the sourceAWithGraders configuration, this split preserves the original 
+distribution of prompts and scores as used in the paper.
 """
 
 
@@ -147,20 +192,34 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
     ]
 
     def _info(self):
-        features = datasets.Features(
-            {
-                "id": datasets.Value("string"),
-                "id_prompt": datasets.Value("string"),
-                "prompt": datasets.Value("string"),
-                "supporting_text": datasets.Value("string"),
-                "essay_title": datasets.Value("string"),
-                "essay_text": datasets.Value("string"),
-                "grades": datasets.Sequence(datasets.Value("int16")),
-                "essay_year": datasets.Value("int16"),
-                "general_comment": datasets.Value("string"),
-                "specific_comment": datasets.Value("string"),
-            }
-        )
+        if self.config.name=="PROPOR2024":
+            features = datasets.Features(
+                {
+                    "id": datasets.Value("string"),
+                    "id_prompt": datasets.Value("string"),
+                    "essay_title": datasets.Value("string"),
+                    "essay_text": datasets.Value("string"),
+                    "grades": datasets.Sequence(datasets.Value("int16")),
+                    "essay_year": datasets.Value("int16"),
+                    "reference": datasets.Value("string"),
+                }
+            )
+        else:
+            features = datasets.Features(
+                {
+                    "id": datasets.Value("string"),
+                    "id_prompt": datasets.Value("string"),
+                    "prompt": datasets.Value("string"),
+                    "supporting_text": datasets.Value("string"),
+                    "essay_title": datasets.Value("string"),
+                    "essay_text": datasets.Value("string"),
+                    "grades": datasets.Sequence(datasets.Value("int16")),
+                    "essay_year": datasets.Value("int16"),
+                    "general_comment": datasets.Value("string"),
+                    "specific_comment": datasets.Value("string"),
+                    "reference": datasets.Value("string"),
+                }
+            )
 
         return datasets.DatasetInfo(
             # This is the description that will appear on the datasets page.
@@ -215,11 +274,42 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
         ]  # arbitrary removal of zero graded essays
         df.to_csv(filepath, index=False)
 
+    def _preprocess_propor2024(self, base_path: str):
+        for split_case in ["train.csv", "validation.csv", "test.csv"]:
+            filepath = f"{base_path}/propor2024/{split_case}"
+            df = pd.read_csv(filepath)
+    
+            # Dictionary to track how many times we've seen each (id, id_prompt) pair
+            counts = {}
+            # List to store the reference for each row
+            references = []
+    
+            # Define the mapping for each occurrence
+            occurrence_to_reference = {
+                0: "crawled_from_web",
+                1: "grader_a",
+                2: "grader_b"
+            }
+    
+            # Iterate through rows in the original order
+            for _, row in df.iterrows():
+                key = (row["id"], row["id_prompt"])
+                count = counts.get(key, 0)
+                # Assign the reference based on the count
+                ref = occurrence_to_reference.get(count, "unknown")
+                references.append(ref)
+                counts[key] = count + 1
+    
+            # Add the reference column without changing the order of rows
+            df["reference"] = references
+            df.to_csv(filepath, index=False)
+
     def _split_generators(self, dl_manager):
         urls = _URLS[self.config.name]
         extracted_files = dl_manager.download_and_extract({self.config.name: urls})
         if "PROPOR2024" == self.config.name:
             base_path = extracted_files["PROPOR2024"]
+            self._preprocess_propor2024(base_path)
             return [
                 datasets.SplitGenerator(
                     name=datasets.Split.TRAIN,
@@ -306,7 +396,8 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
         for grader in [grader_a, grader_b]:
             grader.grades = grader.grades.apply(lambda x: x.strip("[]").split(", "))
             grader.grades = grader.grades.apply(map_list)
-
+        grader_a["reference"] = "grader_a"
+        grader_b["reference"] = "grader_b"
         return grader_a, grader_b
 
     def _generate_splits(self, filepath: str, train_size=0.7):
@@ -409,7 +500,6 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
         assert (
             len(set(val_df["id_prompt"]).intersection(set(test_df["id_prompt"]))) == 0
         ), "Overlap between val and test id_prompt"
-        #train_df['essay_year'] = train_df['essay_year'].astype(int)
         train_df.to_csv(f"{dirname}/train.csv", index=False)
         val_df.to_csv(f"{dirname}/validation.csv", index=False)
         test_df.to_csv(f"{dirname}/test.csv", index=False)
@@ -430,6 +520,7 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
                         "essay_text": row["essay"],
                         "grades": grades,
                         "essay_year": row["essay_year"],
+                        "reference": row["reference"]
                     }
         else:
             with open(filepath, encoding="utf-8") as csvfile:
@@ -449,6 +540,7 @@ class AesEnemDataset(datasets.GeneratorBasedBuilder):
                         "essay_year": row["essay_year"],
                         "general_comment": row["general"],
                         "specific_comment": row["specific"],
+                        "reference": row["reference"]
                     }
             
 
@@ -719,6 +811,7 @@ class HTMLParser:
                 general_comment = None
                 specific_comment = None
                 essay_year = None
+                reference = "crawled_from_web"
                 for prompt_folder in tqdm(
                     sub_folders,
                     desc=f"Parsing HTML files from: {key}",
@@ -761,6 +854,7 @@ class HTMLParser:
                                 general_comment,
                                 specific_comment,
                                 essay_year,
+                                reference
                             ]
                         )
                         essay_id += 1

pyproject.toml

@@ -0,0 +1,13 @@
+[project]
+name = "aes-enem-dataset"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.12"
+dependencies = [
+    "beautifulsoup4>=4.12.3",
+    "datasets>=3.2.0",
+    "ipdb>=0.13.13",
+    "pandas>=2.2.3",
+    "tqdm>=4.67.1",
+]