import gradio as gr import torch from transformers import AutoModelForTokenClassification, AutoTokenizer import pandas as pd import numpy as np # Play with me, consts CONDITIONING_VARIABLES = ["none", "birth_place", "birth_date", "name"] FEMALE_WEIGHTS = [1.5, 5] # About 5x more male than female tokens in dataset # Internal consts START_YEAR = 1800 STOP_YEAR = 1999 SPLIT_KEY = "DATE" DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") MAX_TOKEN_LENGTH = 128 NON_LOSS_TOKEN_ID = -100 NON_GENDERED_TOKEN_ID = 30 # Picked an int that will pop out visually LABEL_DICT = {"female": 9, "male": -9} # Picked an int that will pop out visually CLASSES = list(LABEL_DICT.keys()) # Fire up the models models_paths = dict() models = dict() base_path = "emilylearning/" for var in CONDITIONING_VARIABLES: for f_weight in FEMALE_WEIGHTS: if f_weight == 1.5: models_paths[(var, f_weight)] = ( base_path + f"finetuned_cgp_added_{var}__female_weight_{f_weight}__test_run_False__p_dataset_100" ) else: models_paths[(var, f_weight)] = ( base_path + f"finetuned_cgp_add_{var}__f_weight_{f_weight}__p_dataset_100__test_False" ) models[(var, f_weight)] = AutoModelForTokenClassification.from_pretrained( models_paths[(var, f_weight)] ) # Tokenizers same for each model, so just grabbing one of them tokenizer = AutoTokenizer.from_pretrained( models_paths[(CONDITIONING_VARIABLES[0], FEMALE_WEIGHTS[0])], add_prefix_space=True ) MASK_TOKEN_ID = tokenizer.mask_token_id # more static stuff gendered_lists = [ ["he", "she"], ["him", "her"], ["his", "hers"], ["male", "female"], ["man", "woman"], ["men", "women"], ["husband", "wife"], ] male_gendered_dict = {list[0]: list for list in gendered_lists} female_gendered_dict = {list[1]: list for list in gendered_lists} male_gendered_token_ids = tokenizer.convert_tokens_to_ids( list(male_gendered_dict.keys()) ) female_gendered_token_ids = tokenizer.convert_tokens_to_ids( list(female_gendered_dict.keys()) ) assert tokenizer.unk_token_id not in male_gendered_token_ids assert tokenizer.unk_token_id not in female_gendered_token_ids label_list = list(LABEL_DICT.values()) assert label_list[0] == LABEL_DICT["female"], "LABEL_DICT not an ordered dict" label2id = {label: idx for idx, label in enumerate(label_list)} # Prepare text def tokenize_and_append_metadata(text, tokenizer): tokenized = tokenizer( text, truncation=True, padding=True, max_length=MAX_TOKEN_LENGTH, ) # Finding the gender pronouns in the tokens token_ids = tokenized["input_ids"] female_tags = torch.tensor( [ LABEL_DICT["female"] if id in female_gendered_token_ids else NON_GENDERED_TOKEN_ID for id in token_ids ] ) male_tags = torch.tensor( [ LABEL_DICT["male"] if id in male_gendered_token_ids else NON_GENDERED_TOKEN_ID for id in token_ids ] ) # Labeling and masking out occurrences of gendered pronouns labels = torch.tensor([NON_LOSS_TOKEN_ID] * len(token_ids)) labels = torch.where( female_tags == LABEL_DICT["female"], label2id[LABEL_DICT["female"]], NON_LOSS_TOKEN_ID, ) labels = torch.where( male_tags == LABEL_DICT["male"], label2id[LABEL_DICT["male"]], labels ) masked_token_ids = torch.where( female_tags == LABEL_DICT["female"], MASK_TOKEN_ID, torch.tensor(token_ids) ) masked_token_ids = torch.where( male_tags == LABEL_DICT["male"], MASK_TOKEN_ID, masked_token_ids ) tokenized["input_ids"] = masked_token_ids tokenized["labels"] = labels return tokenized # Run inference def predict_gender_pronouns( num_points, conditioning_variables, f_weights, input_text, return_preds=False ): text_portions = input_text.split(SPLIT_KEY) years = np.linspace(START_YEAR, STOP_YEAR, int(num_points)).astype(int) dfs = [] dfs.append(pd.DataFrame({"year": years})) for f_weight in f_weights: for var in conditioning_variables: prefix = f"w{f_weight}_{var}" model = models[(var, f_weight)] p_female = [] p_male = [] for b_date in years: target_text = f"{b_date}".join(text_portions) tokenized_sample = tokenize_and_append_metadata( target_text, tokenizer=tokenizer, ) ids = tokenized_sample["input_ids"] atten_mask = torch.tensor(tokenized_sample["attention_mask"]) toks = tokenizer.convert_ids_to_tokens(ids) labels = tokenized_sample["labels"] with torch.no_grad(): outputs = model(ids.unsqueeze(dim=0), atten_mask.unsqueeze(dim=0)) preds = torch.argmax(outputs[0][0].cpu(), dim=1) was_masked = labels.cpu() != -100 preds = torch.where(was_masked, preds, -100) num_preds = torch.sum(was_masked).item() p_female.append(len(torch.where(preds==0)[0])/num_preds*100) p_male.append(len(torch.where(preds==1)[0])/num_preds*100) dfs.append(pd.DataFrame({f"%f_{prefix}": p_female, f"%m_{prefix}": p_male})) results = pd.concat(dfs, axis=1).set_index("year") female_df = results.filter(regex=".*f_") female_df_for_plot = ( female_df.reset_index() ) # Gradio timeseries requires x-axis as column? male_df = results.filter(regex=".*m_") male_df_for_plot = ( male_df.reset_index() ) # Gradio timeseries requires x-axis as column? return ( target_text, female_df_for_plot, female_df, male_df_for_plot, male_df, ) title = "Changing Gender Pronouns" description = """ This is a demo for a project exploring possible spurious correlations in training datasets that can be exploited and manipulated to achieve alternative outcomes. In this case, manipulating `DATE` to change the predicted gender pronouns for both the BERT base model and a model fine-tuned with a specific pronoun predicting task using the [wiki-bio](https://huggingface.co/datasets/wiki_bio) dataset. One way to explain phenomena is by looking at a likely data generating process for biographical-like data in both the main BERT training dataset as well as the `wiki_bio` dataset, in the form of a causal DAG. In the DAG, we can see that `birth_place`, `birth_date` and `gender` are all independent elements that have no common cause with the other covariates in the DAG. However `birth_place`, `birth_date` and `gender` may all have a role in causing one's `access_to_resources`, with the general trend that `access_to_resources` has become less gender-dependent over time, but not in every `birth_place`, with recent events in Afghanistan providing a stark counterexample to this trend. `access_to_resources` further determines how or if at all, you may appear in the dataset’s `context_words`. We also argue that although there are complex causal interactions between words in a segment, the `context_words` are more likely to cause the `gender_pronouns`, rather than vice versa. For example, if the subject is a famous doctor and the object is her wealthy father, these context words will determine which person is being referred to, and thus which gendered-pronoun to use. In this graph, any pink path between `context_words` and `gender_pronouns` will allow the flow of statistical correlation (regardless of direction of the causal arrow), inviting confounding and thus spurious correlations into the trained model.
DAG of possible data generating process for datasets used in training.
Those familiar with causal DAGs may note when can simply condition on `gender` to block any confounding between the `context_words` and the `gender_pronouns`. However, this is not always possible, particularly in generative or mask-filling tasks, like those common in language models. Here, we automatically mask (for prediction) the following tokens (and they will also be automatically masked if you use them below.) ``` gendered_lists = [ ['he', 'she'], ['him', 'her'], ['his', 'hers'], ['male', 'female'], ['man', 'woman'], ['men', 'women'], ["husband", "wife"], ] ``` In this demo we are looking for a dose-response relationship between: - our treatment: the text, - and our outcome: the predicted gender of pronouns in the text. Specifically we are seeing if making larger magnitude intervention: an older `DATE` in the text will result in a larger magnitude effect in the outcome: higher percentage of predicted female pronouns. In the demo below you can select among 4 different fine-tuning methods: - which, if any, conditioning variable was appended to the text. And two different weighting schemes that were used in the loss function to nudge more toward the minority class in the dataset: - female pronouns. """ article = "Check out [main colab notebook](https://colab.research.google.com/drive/14ce4KD6PrCIL60Eng-t79tEI1UP-DHGz?usp=sharing#scrollTo=Mg1tUeHLRLaG) \ with a lot more details about this method and implementation." gr.Interface( fn=predict_gender_pronouns, inputs=[ gr.inputs.Number( default=10, label="Number of points (years) plotted -- select fewer if slow.", ), gr.inputs.CheckboxGroup( CONDITIONING_VARIABLES, default=["none", "birth_date"], type="value", label="Pick model(s) that were trained with the following conditioning variables", ), gr.inputs.CheckboxGroup( FEMALE_WEIGHTS, default=[5], type="value", label="Pick model(s) that were trained with the following loss function weight on female predictions", ), gr.inputs.Textbox( lines=7, label="Input Text. Include one of more instance of the word 'DATE' below, to be replace with a range of dates in demo.", default="Born DATE, she was a computer scientist. Her work was greatly respected, and she was well-regarded in her field.", ), ], outputs=[ gr.outputs.Textbox(type="auto", label="Sample target text fed to model"), gr.outputs.Timeseries( x="year", label="Precent pred female pronoun vs year, per model trained with conditioning and with weight for female preds", ), gr.outputs.Dataframe( overflow_row_behaviour="show_ends", label="Precent pred female pronoun vs year, per model trained with conditioning and with weight for female preds", ), gr.outputs.Timeseries( x="year", label="Precent pred male pronoun vs year, per model trained with conditioning and with weight for female preds", ), gr.outputs.Dataframe( overflow_row_behaviour="show_ends", label="Precent pred male pronoun vs year, per model trained with conditioning and with weight for female preds", ), ], title = title, description = description, article = article ).launch()