| """Coreference / pronoun resolution — elemental reading comprehension. |
| |
| Tests the model's ability to resolve pronouns and other referring expressions |
| to their antecedents. This is a prerequisite for WinoGrande-style benchmarks |
| and general reading comprehension. |
| |
| Categories: |
| - pronoun_simple: Unambiguous pronoun resolution (he/she/they) |
| - pronoun_hard: Winograd-style where world knowledge is needed |
| |
| All cases are *unambiguous* — there is exactly one correct answer. |
| |
| Format (ICL): |
| Sentence: "Alice told Bob that she would be late." |
| Who does "she" refer to? |
| Answer: Alice |
| """ |
|
|
| import random |
| from typing import Dict, List, Any |
|
|
| from tasks.base_task import BaseTask, TaskConfig |
|
|
|
|
| class CoreferenceTask(BaseTask): |
| """Elemental coreference / pronoun resolution task.""" |
|
|
| TASK_NAME = "coreference" |
|
|
| CATEGORY_DATA: Dict[str, List[Dict[str, str]]] = { |
| "pronoun_simple": [ |
| {"input": 'Sentence: "Alice told Bob that she would be late."\nWho does "she" refer to?', "output": "Alice"}, |
| {"input": 'Sentence: "David asked Sarah to help him with the project."\nWho does "him" refer to?', "output": "David"}, |
| {"input": 'Sentence: "Maria finished her homework before dinner."\nWho does "her" refer to?', "output": "Maria"}, |
| {"input": 'Sentence: "The teacher praised James because he scored the highest."\nWho does "he" refer to?', "output": "James"}, |
| {"input": 'Sentence: "Emma called her mother and told her the good news."\nWho does the first "her" refer to?', "output": "Emma"}, |
| {"input": 'Sentence: "Tom lent his car to his brother Mike."\nWho does "his" refer to?', "output": "Tom"}, |
| {"input": 'Sentence: "When Lisa entered the room, she noticed the window was open."\nWho does "she" refer to?', "output": "Lisa"}, |
| {"input": 'Sentence: "The doctor told the patient that he needed more rest."\nWho does "he" refer to?', "output": "the patient"}, |
| {"input": 'Sentence: "Kevin and his dog went for a walk in the park."\nWho does "his" refer to?', "output": "Kevin"}, |
| {"input": 'Sentence: "Rachel sent a letter to her grandmother."\nWho does "her" refer to?', "output": "Rachel"}, |
| {"input": 'Sentence: "Mr. Johnson asked his secretary to reschedule the meeting."\nWho does "his" refer to?', "output": "Mr. Johnson"}, |
| {"input": 'Sentence: "Sophie baked a cake and brought it to the party."\nWhat does "it" refer to?', "output": "the cake"}, |
| {"input": 'Sentence: "After Mark finished his speech, he received a standing ovation."\nWho does "he" refer to?', "output": "Mark"}, |
| {"input": 'Sentence: "Diana wrote a novel and published it last year."\nWhat does "it" refer to?', "output": "the novel"}, |
| {"input": 'Sentence: "The manager told the employee that she had been promoted."\nWho does "she" refer to?', "output": "the employee"}, |
| {"input": 'Sentence: "Paul took his children to the zoo on Saturday."\nWho does "his" refer to?', "output": "Paul"}, |
| {"input": 'Sentence: "Olivia practiced the piano until she could play the piece perfectly."\nWho does "she" refer to?', "output": "Olivia"}, |
| {"input": 'Sentence: "The chef prepared a special dish and served it to the guests."\nWhat does "it" refer to?', "output": "the dish"}, |
| {"input": 'Sentence: "Laura thanked her friend for helping her move."\nWho does the first "her" refer to?', "output": "Laura"}, |
| {"input": 'Sentence: "Ben fixed the car and drove it to work."\nWhat does "it" refer to?', "output": "the car"}, |
| ], |
|
|
| "pronoun_hard": [ |
| {"input": 'Sentence: "The trophy didn\'t fit in the suitcase because it was too big."\nWhat was too big?', "output": "the trophy"}, |
| {"input": 'Sentence: "The trophy didn\'t fit in the suitcase because it was too small."\nWhat was too small?', "output": "the suitcase"}, |
| {"input": 'Sentence: "The city council refused the demonstrators a permit because they feared violence."\nWho feared violence?', "output": "the city council"}, |
| {"input": 'Sentence: "The city council refused the demonstrators a permit because they advocated violence."\nWho advocated violence?', "output": "the demonstrators"}, |
| {"input": 'Sentence: "Sam pulled the wagon and Chris pushed it, even though he was very tired."\nWho was tired?', "output": "Sam"}, |
| {"input": 'Sentence: "The painting was sold to the collector because it was very valuable."\nWhat was valuable?', "output": "the painting"}, |
| {"input": 'Sentence: "The plant didn\'t grow well in the pot because it was too dry."\nWhat was too dry?', "output": "the pot"}, |
| {"input": 'Sentence: "The cat caught the mouse because it was too slow."\nWhat was too slow?', "output": "the mouse"}, |
| {"input": 'Sentence: "The cat caught the mouse because it was very quick."\nWhat was quick?', "output": "the cat"}, |
| {"input": 'Sentence: "Joan made sure to thank Susan for all the help she had given."\nWho gave help?', "output": "Susan"}, |
| {"input": 'Sentence: "The table won\'t fit through the door because it is too wide."\nWhat is too wide?', "output": "the table"}, |
| {"input": 'Sentence: "The table won\'t fit through the door because it is too narrow."\nWhat is too narrow?', "output": "the door"}, |
| {"input": 'Sentence: "Bob collapsed on the sidewalk, and Jim called an ambulance for him."\nWho needed the ambulance?', "output": "Bob"}, |
| {"input": 'Sentence: "The doctor told the nurse that she had made an error in the chart."\nWho made an error?', "output": "the nurse"}, |
| {"input": 'Sentence: "The truck hit the pole and it broke into pieces."\nWhat broke?', "output": "the pole"}, |
| {"input": 'Sentence: "The actress used to be the director\'s favorite, but she replaced her."\nWho was replaced?', "output": "the actress"}, |
| {"input": 'Sentence: "The ball broke the window because it was fragile."\nWhat was fragile?', "output": "the window"}, |
| {"input": 'Sentence: "The ball broke the window because it was moving fast."\nWhat was moving fast?', "output": "the ball"}, |
| {"input": 'Sentence: "The guests enjoyed the meal that the chef prepared because it was delicious."\nWhat was delicious?', "output": "the meal"}, |
| {"input": 'Sentence: "The lock couldn\'t be opened by the key because it was rusty."\nWhat was rusty?', "output": "the lock"}, |
| ], |
| } |
|
|
| CATEGORY_DEMOS: Dict[str, List[str]] = { |
| "pronoun_simple": [ |
| 'Sentence: "John asked Mary if she could help."\nWho does "she" refer to?\nAnswer: Mary', |
| 'Sentence: "Carlos lost his wallet at the store."\nWho does "his" refer to?\nAnswer: Carlos', |
| 'Sentence: "The professor graded the papers and returned them to the students."\nWhat does "them" refer to?\nAnswer: the papers', |
| 'Sentence: "Anna brought her dog to the vet."\nWho does "her" refer to?\nAnswer: Anna', |
| 'Sentence: "The boy kicked the ball and it flew over the fence."\nWhat does "it" refer to?\nAnswer: the ball', |
| ], |
| "pronoun_hard": [ |
| 'Sentence: "The bottle didn\'t fit in the bag because it was too large."\nWhat was too large?\nAnswer: the bottle', |
| 'Sentence: "The town council denied the miners a parade permit because they worried about damage."\nWho worried?\nAnswer: the town council', |
| 'Sentence: "The book fell off the shelf because it was uneven."\nWhat was uneven?\nAnswer: the shelf', |
| 'Sentence: "Bill passed the salt to John because he needed it for his soup."\nWho needed the salt?\nAnswer: John', |
| 'Sentence: "The vase broke when the mover dropped it because it was slippery."\nWhat was slippery?\nAnswer: the vase', |
| ], |
| } |
|
|
| def __init__(self, config: TaskConfig): |
| super().__init__(config) |
|
|
| def _load_data(self): |
| import pandas as pd |
|
|
| if self.config.in_memory_data: |
| self.data = pd.DataFrame(self.config.in_memory_data) |
| return |
|
|
| rows: List[Dict[str, str]] = [] |
| for category, examples in self.CATEGORY_DATA.items(): |
| for ex in examples: |
| rows.append({ |
| "input": ex["input"], |
| "output": ex["output"], |
| "category_name": category, |
| }) |
| self.data = pd.DataFrame(rows) |
|
|
| def build_prompt(self, instance: Dict[str, Any], num_shots: int = 5) -> str: |
| category = instance.get("category_name", "pronoun_simple") |
| demos = self.CATEGORY_DEMOS.get(category, []) |
|
|
| inst_input = instance.get("input", "") |
| demos = [d for d in demos if inst_input not in d][:num_shots] |
|
|
| prompt = "" |
| for d in demos: |
| prompt += d + "\n\n" |
| prompt += f"{inst_input}\nAnswer:" |
| return prompt |
|
|
| def evaluate(self, predictions: List[str], split: str = "test", **kwargs) -> Dict[str, float]: |
| ground_truth = self.get_ground_truth(split) |
| task_data = self.get_split(split) |
|
|
| if len(predictions) != len(ground_truth): |
| raise ValueError( |
| f"Prediction count ({len(predictions)}) != ground truth ({len(ground_truth)})" |
| ) |
|
|
| processed = [self.preprocess_prediction(p) for p in predictions] |
|
|
| correct = sum( |
| 1 for p, g in zip(processed, ground_truth) |
| if p.lower().strip() == g.lower().strip() |
| ) |
| results: Dict[str, Any] = { |
| "accuracy": correct / len(ground_truth), |
| "correct": correct, |
| "total": len(ground_truth), |
| } |
|
|
| cat_stats: Dict[str, Dict[str, int]] = {} |
| for p, g, item in zip(processed, ground_truth, task_data): |
| cat = item.get("category_name", "unknown") |
| cat_stats.setdefault(cat, {"correct": 0, "total": 0}) |
| cat_stats[cat]["total"] += 1 |
| if p.lower().strip() == g.lower().strip(): |
| cat_stats[cat]["correct"] += 1 |
|
|
| for cat, s in cat_stats.items(): |
| results[f"accuracy_{cat}"] = s["correct"] / s["total"] |
|
|
| return results |
|
|
| def get_ground_truth(self, split: str = "test") -> List[str]: |
| rows = self.get_split(split) |
| return [str(r.get("output", "")) for r in rows] |
|
|
|
|
| def create_coreference_task( |
| category: str = None, |
| name: str = "coreference", |
| ) -> CoreferenceTask: |
| """Create a CoreferenceTask, optionally filtered to one category.""" |
| data = None |
| if category and category in CoreferenceTask.CATEGORY_DATA: |
| data = [ |
| {**ex, "category_name": category} |
| for ex in CoreferenceTask.CATEGORY_DATA[category] |
| ] |
| name = f"coreference:{category}" |
|
|
| config = TaskConfig( |
| name=name, |
| description="Coreference / pronoun resolution", |
| data_format="memory", |
| in_memory_data=data, |
| input_column="input", |
| output_column="output", |
| evaluation_metrics=["accuracy"], |
| metadata={"task_type": "coreference", "category": category}, |
| ) |
| return CoreferenceTask(config) |
|
|
