import logging from datasets import load_dataset from imblearn.metrics import macro_averaged_mean_absolute_error from sklearn.metrics import f1_score from evaluate import load import numpy as np import argparse from collections import defaultdict import json logging.basicConfig(format='%(asctime)s %(levelname)-8s %(message)s', level=logging.INFO, datefmt='%Y-%m-%d %H:%M:%S') # argument parser = argparse.ArgumentParser(description='Super TweetEval evaluation script.') parser.add_argument('-p', '--prediction-path', required=True, type=str, help="path to directory wiht that contains the model predictions on the test sets. One file per task.") parser.add_argument('-o', '--output-file', default="scores.json", type=str, help="path to the output file") parser.add_argument('--t2t-format', action="store_false", default=True, help="format prediction file in T2T format (ONLY for NER7)") opt = parser.parse_args() task_names = ['tweet_topic', 'tweet_ner7', 'tweet_qa', 'tweet_qg', 'tweet_intimacy', 'tweet_similarity', 'tempo_wic', 'tweet_hate', 'tweet_nerd', 'tweet_emoji', 'tweet_sentiment', 'tweet_emotion'] scores = defaultdict(lambda : 0) #{k:0 for k in task_names} not_found = [] for task in task_names: # load dataset data = load_dataset("cardiffnlp/super_tweeteval", task, use_auth_token=True, split="test") try: if task == 'tempo_wic': label2id = {"no": 0, "yes": 1} with open(f"{opt.prediction_path}/tempo-wic.txt") as f: _predictions = [] output = f.read().split('\n') for entry in output: if entry in label2id: _predictions.append(label2id[entry]) else: _predictions.append(-1) gold_labels = data["gold_label_binary"] eval_metric = {"accuracy": np.mean([int(a == b) for a, b in zip(_predictions, gold_labels)])} scores[task] = eval_metric["accuracy"] elif task == "tweet_emoji": # load label names with open('../data/tweet_emoji/map.txt') as f: label_classes = f.readlines() label_names = [x.strip('\n') for x in label_classes] label_names = [x.split(',')[1] for x in label_names] with open(f"{opt.prediction_path}/tweet-emoji.txt") as f: lines = f.readlines() lines = [l.strip('\n') for l in lines] predictions = [] for l in lines: pred_instance = [] # consider only top 5 predictions lines = l.split(',') if ',' in l else l.split(' ') for label in lines[:5]: label = label.strip(" ,") if label in label_names: pred_instance.append(label_names.index(label)) else: pred_instance.append(-1) # emoji not in label_names predictions.append(pred_instance) # metric: accuracy at top 5 gold_labels = np.array(data["gold_label"][:40_000]) eval_metric = {"accuracy_top5": np.mean([1 if gold_labels[i] in predictions[i] else 0 for i in range(len(gold_labels))])} scores[task] = eval_metric["accuracy_top5"] elif task == "tweet_emotion": label_names = data.features['gold_label_list'].feature.names with open(f"{opt.prediction_path}/tweet-emotion.txt") as f: lines = f.readlines() lines = [l.strip('\n') for l in lines] predictions = [] for l in lines: pred_instance = [0] * len(label_names) for label in l.split(','): label = label.strip(' ') if label in label_names: pred_instance[label_names.index(label)] = 1 predictions.append(pred_instance) # metric gold_labels = data["gold_label_list"] eval_metric = {"macro_f1": f1_score(gold_labels, predictions, average='macro')} scores[task] = eval_metric["macro_f1"] elif task == "tweet_ner7": labels = [ 'B-corporation', 'B-creative_work', 'B-event', 'B-group', 'B-location', 'B-person', 'B-product', 'I-corporation', 'I-creative_work', 'I-event', 'I-group', 'I-location', 'I-person', 'I-product', 'O' ] id2label = {i: label for i, label in enumerate(labels)} true_sequence = [[id2label[i] for i in ii] for ii in data['gold_label_sequence']] # metric metric = load("seqeval") if opt.t2t_format: # format prediction file in IOB sequence with open(f"{opt.prediction_path}/tweet-ner7.txt") as f: lines = f.read().split("\n") output = [l.strip('\n') for l in lines] output = [list(set(i.split(","))) for i in output] prediction_sequence = [] for d, o in zip(data, output): tag_seq = ['O'] * len(d['text_tokenized']) for _o in o: if len(_o.split(":")) != 2: continue entity, _type = _o.split(":") entity_tokens = entity.split(" ") try: i = d['text_tokenized'].index(entity_tokens[0]) tag_seq[i] = f"B-{_type.strip()}" if len(entity_tokens) > 1: for j in range(1, len(entity_tokens)): tag_seq[i + j] = f"I-{_type.strip()}" except: continue prediction_sequence.append(tag_seq) else: with open(opt.prediction_file) as f: prediction_sequence = [[id2label[j] if j in id2label else j for j in i.split('\t')] for i in f.read().split("\n")] eval_metric = metric.compute(predictions=prediction_sequence, references=true_sequence) eval_metric = {'overall_f1': eval_metric['overall_f1']} scores[task] = eval_metric['overall_f1'] elif task == "tweet_hate": label_names = data.features['gold_label'].names with open(f"{opt.prediction_path}/tweet-hate.txt") as f: lines = f.readlines() output = [i.strip('\n') for i in lines] predictions = [] for x in output: if x not in label_names: predictions.append(-1) else: predictions.append(label_names.index(x)) gold_labels = data["gold_label"] # do not consider not_hate class f1_multi = f1_score(gold_labels, predictions, labels=list(range(7)), average='macro') # consider all hate subclasses as one class predictions_binary = [1 if x in list(range(7)) else 0 for x in predictions] gold_labels_binary = [1 if x in list(range(7)) else 0 for x in gold_labels] f1_binary = f1_score(gold_labels_binary, predictions_binary, average='micro') eval_metric = {"combined_f1": (f1_multi+f1_binary)/2} scores[task] = eval_metric["combined_f1"] elif task == "tweet_intimacy": gold_labels = data["gold_score"] # mean_value to be used if model outputs a non-numeric value mean_value = sum(gold_labels)/len(gold_labels) # metric metric = load("spearmanr") with open(f"{opt.prediction_path}/tweet-intimacy.txt") as f: _predictions = [] lines = f.readlines() output = [l.strip('\n') for l in lines] for i in output: try: _predictions.append(float(i)) except ValueError: _predictions.append(mean_value) failed_predictions += 1 corr_spear = metric.compute(predictions=_predictions, references=gold_labels) eval_metric = {"spearmanr": corr_spear} scores[task] = eval_metric["spearmanr"]['spearmanr'] elif task == "tweet_nerd": # metric label2id = {"no": 0, "yes": 1} with open(f"{opt.prediction_path}/tweet-nerd.txt") as f: _predictions = [] output = f.read().split('\n') output = [x.lower().strip() for x in output] for entry in output: if entry in label2id: _predictions.append(label2id[entry]) else: _predictions.append(-1) gold_labels = data["gold_label_binary"] eval_metric = {"accuracy": np.mean([int(a == b) for a, b in zip(_predictions, gold_labels)])} scores[task] = eval_metric["accuracy"] elif task == "tweet_qa": metric = load("squad") with open(f"{opt.prediction_path}/tweet-qa.txt") as f: lines = f.readlines() output = [l.strip('\n') for l in lines] _predictions = [{"prediction_text": p, "id": str(_n)} for _n, p in enumerate(output)] _references = [{"answers": {"answer_start": [100], "text": [r["gold_label_str"]]}, "id": str(_n)} for _n, r in enumerate(data)] eval_metric = metric.compute(predictions=_predictions, references=_references) eval_metric.pop("exact_match") eval_metric["f1"] = eval_metric["f1"]/100 scores[task] = eval_metric["f1"] elif task == "tweet_qg": metric = load("meteor") with open(f"{opt.prediction_path}/tweet-qg.txt") as f: lines = f.readlines() _predictions = [l.strip('\n') for l in lines] _references = data["gold_label_str"] eval_metric = metric.compute(predictions=_predictions, references=_references) scores[task] = eval_metric["meteor"] elif task == "tweet_sentiment": label_names = data.features['gold_label'].names with open(f"{opt.prediction_path}/tweet-sentiment.txt") as f: lines = f.readlines() output = [l.strip('\n') for l in lines] predictions = [] # if the model outputs a label that is not in the label set, we set the label to be "neutral or negative" (2) for x in output: x = x.strip(' ') if x not in label_names: predictions.append(2) else: predictions.append(label_names.index(x)) # metric: r2 score gold_labels = data["gold_label"] macro_mae = macro_averaged_mean_absolute_error(gold_labels, predictions) macro_mae = 1 - macro_mae # set a floor of -1 for worst model macro_mae = max([-1, macro_mae]) eval_metric = {"macro_mae": macro_mae} scores[task] = eval_metric["macro_mae"] elif task == "tweet_similarity": gold_labels = data["gold_score"] # mean_value to be used if model outputs a non-numeric value mean_value = sum(gold_labels)/len(gold_labels) # metric metric = load("spearmanr") with open(f"{opt.prediction_path}/tweet-similarity.txt") as f: _predictions = [] lines = f.readlines() output = [l.strip('\n') for l in lines] for i in output: try: _predictions.append(float(i)) except ValueError: _predictions.append(mean_value) corr_spear = metric.compute(predictions=_predictions, references=gold_labels) eval_metric = {"spearmanr": corr_spear} scores[task] = eval_metric["spearmanr"]['spearmanr'] elif task == "tweet_topic": label_names = data.features['gold_label_list'].feature.names with open(f"{opt.prediction_path}/tweet-topic.txt") as f: lines = f.readlines() lines = [l.strip('\n') for l in lines] predictions = [] for l in lines: pred_instance = [0] * len(label_names) for label in l.split(','): label = label.strip(' ') if label in label_names: pred_instance[label_names.index(label)] = 1 predictions.append(pred_instance) # metric gold_labels = data["gold_label_list"] eval_metric = {"macro_f1": f1_score(gold_labels, predictions, average='macro')} scores[task] = eval_metric["macro_f1"] except FileNotFoundError: not_found.append(task) continue # clusters/groups to evaluate subgroups = { "temporal": ["tweet_ner7", "tempo_wic", "tweet_topic", "tweet_nerd"], "multi-label": ["tweet_topic", "tweet_emotion"], "multi-class": ["tweet_sentiment", "tweet_hate"], "regression": ["tweet_similarity", "tweet_intimacy", "tweet_sentiment"], "target-based": ["tweet_sentiment", "tempo_wic", "tweet_nerd"], "big-label": ["tweet_emoji", "tweet_topic"], "disambiguation": ["tempo_wic", "tweet_nerd"], "generation": ["tweet_qa", "tweet_qg"], } #scores = {k:round(v*100, 2) for k,v in scores.items()} score_avg = np.mean(list(scores.values())) subgroups_scores = {group: np.mean([scores[task] for task in subgroups[group] if task not in not_found]) for group in subgroups} logging.info(f"Tasks not found: {not_found}\n") logging.info("*** Scores ***") for sc in scores: logging.info(f"{sc}: {scores[sc]:.2f}") logging.info("*** Clusters scores ***") for group in subgroups: logging.info(f"{group}: {subgroups_scores[group]:.2f}") logging.info(f"*** Average score: {score_avg:.2f} ***") # write scores to file results = {"scores": scores, "clusters": subgroups_scores, "average": score_avg} with open(opt.output_file, "w") as f: json.dump(results, f, indent=4)