# Model card: https://huggingface.co/emilylearning/selection-induced-collider-bias # %% import gradio as gr import matplotlib.pyplot as plt import numpy as np import pandas as pd import random from matplotlib.ticker import MaxNLocator from transformers import pipeline from winogender_sentences import get_sentences OWN_MODEL_NAME = 'add-a-model' PICK_YOUR_OWN_LABEL = 'pick-your-own' MODEL_NAME_DICT = { "roberta-large": "RoBERTa-large", "bert-large-uncased": "BERT-large", "roberta-base": "RoBERTa-base", "bert-base-uncased": "BERT-base", "olm/olm-roberta-base-oct-2022": "OLM_RoBERTa-base", OWN_MODEL_NAME: "Your model's" } MODEL_NAMES = list(MODEL_NAME_DICT.keys()) DECIMAL_PLACES = 1 EPS = 1e-5 # to avoid /0 errors NUM_PTS_TO_AVERAGE = 2 # Example date conts DATE_SPLIT_KEY = "DATE" START_YEAR = 1901 STOP_YEAR = 2016 NUM_PTS = 30 DATES = np.linspace(START_YEAR, STOP_YEAR, NUM_PTS).astype(int).tolist() DATES = [f'{d}' for d in DATES] GENDERED_LIST = [ ['he', 'she'], ['him', 'her'], ['his', 'hers'], ["himself", "herself"], ['male', 'female'], # ['man', 'woman'] Explicitly added in winogender extended sentences ['men', 'women'], ["husband", "wife"], ['father', 'mother'], ['boyfriend', 'girlfriend'], ['brother', 'sister'], ["actor", "actress"], ] # %% # Fire up the models models = {m : pipeline("fill-mask", model=m) for m in MODEL_NAMES if m != OWN_MODEL_NAME} # %% # Get the winogender sentences winogender_sentences = get_sentences() occs = sorted(list({sentence_id.split('_')[0] for sentence_id in winogender_sentences})) # %% def get_gendered_token_ids(): male_gendered_tokens = [list[0] for list in GENDERED_LIST] female_gendered_tokens = [list[1] for list in GENDERED_LIST] return male_gendered_tokens, female_gendered_tokens def get_winogender_texts(occ): return [winogender_sentences[id] for id in winogender_sentences.keys() if id.split('_')[0] == occ] def display_input_texts(occ, alt_text): if occ == PICK_YOUR_OWN_LABEL: texts = alt_text.split('\n') else: texts = get_winogender_texts(occ) display_texts = [ f"{i+1}) {text}" for (i, text) in enumerate(texts)] return "\n".join(display_texts), texts def get_avg_prob_from_pipeline_outputs(pipeline_preds, gendered_tokens, num_preds): pronoun_preds = [sum([ pronoun["score"] if pronoun["token_str"].strip( ).lower() in gendered_tokens else 0.0 for pronoun in top_preds]) for top_preds in pipeline_preds ] return round(sum(pronoun_preds) / (EPS + num_preds) * 100, DECIMAL_PLACES) def is_top_pred_gendered(pipeline_preds, gendered_tokens): return pipeline_preds[0][0]['token_str'].strip().lower() in gendered_tokens # %% def get_figure(df, model_name, occ): xs = df[df.columns[0]] ys = df[df.columns[1]] fig, ax = plt.subplots() ax.bar(xs, ys) ax.axis('tight') ax.set_xlabel("Sentence number") ax.set_ylabel("Uncertainty metric") ax.set_title(f"{MODEL_NAME_DICT[model_name]} gender pronoun uncertainty in '{occ}' sentences") return fig # %% def predict_gender_pronouns( model_name, own_model_name, texts, occ, ): """Run inference on input_text for selected model type, returning uncertainty results. """ # TODO: make these selectable by user indie_vars = ', '.join(DATES) num_ave = NUM_PTS_TO_AVERAGE # For debugging print('input_texts', texts) if model_name is None or model_name == '': model_name = MODEL_NAMES[0] model = models[model_name] elif model_name == OWN_MODEL_NAME: model = pipeline("fill-mask", model=own_model_name) else: model = models[model_name] mask_token = model.tokenizer.mask_token indie_vars_list = indie_vars.split(',') male_gendered_tokens, female_gendered_tokens = get_gendered_token_ids() masked_texts = [text.replace('MASK', mask_token) for text in texts] all_uncertainty_f = {} not_top_gendered = set() for i, text in enumerate(masked_texts): female_pronoun_preds = [] male_pronoun_preds = [] top_pred_gendered = True # Assume true unless told otherwise print(f"{i+1}) {text}") for indie_var in indie_vars_list[:num_ave] + indie_vars_list[-num_ave:]: target_text = f"In {indie_var}: {text}" pipeline_preds = model(target_text) # Quick hack as realized return type based on how many MASKs in text. if type(pipeline_preds[0]) is not list: pipeline_preds = [pipeline_preds] # If top-pred not gendered, record as such if not is_top_pred_gendered(pipeline_preds, female_gendered_tokens + male_gendered_tokens): top_pred_gendered = False num_preds = 1 # By design female_pronoun_preds.append(get_avg_prob_from_pipeline_outputs( pipeline_preds, female_gendered_tokens, num_preds )) male_pronoun_preds.append(get_avg_prob_from_pipeline_outputs( pipeline_preds, male_gendered_tokens, num_preds )) # Normalizing by all gendered predictions total_gendered_probs = np.add( female_pronoun_preds, male_pronoun_preds) norm_female_pronoun_preds = np.around( np.divide(female_pronoun_preds, total_gendered_probs+EPS)*100, decimals=DECIMAL_PLACES ) sent_idx = f"{i+1}" if top_pred_gendered else f"{i+1}*" all_uncertainty_f[sent_idx] = round(abs((sum(norm_female_pronoun_preds[-num_ave:]) - sum(norm_female_pronoun_preds[:num_ave])) / num_ave), DECIMAL_PLACES) uncertain_df = pd.DataFrame.from_dict( all_uncertainty_f, orient='index', columns=['Uncertainty metric']) uncertain_df = uncertain_df.reset_index().rename( columns={'index': 'Sentence number'}) return ( target_text, uncertain_df, get_figure(uncertain_df, model_name, occ), ) demo = gr.Blocks() with demo: input_texts = gr.Variable([]) gr.Markdown("## Are you certain?") gr.Markdown( "#### LLMs are pretty good at reporting their uncertainty. We just need to ask the right way.") gr.Markdown("Using our uncertainty metric informed by applying causal inference techniques in \ ['Selection Induced Collider Bias: A Gender Pronoun Uncertainty Case Study'](https://arxiv.org/abs/2210.00131 ), \ we are able to identify likely spurious correlations and exploit them in \ the scenario of gender underspecified tasks. (Note that introspecting softmax probabilities alone is insufficient, as in the sentences \ below, LLMs may report a softmax prob of ~0.9 despite the task being underspecified.)") gr.Markdown("We extend the [Winogender Schemas](https://github.com/rudinger/winogender-schemas) evaluation set to produce\ eight syntactically similar sentences. However semantically, \ only two of the sentences are gender-specified while the rest remain gender-underspecified") gr.Markdown("If a model can reliably tell us when it is uncertain about its predictions, one can replace only those uncertain predictions with\ an appropriate heuristic or information retrieval process.") gr.Markdown("#### TL;DR") gr.Markdown("Follow steps below to test out one of the pre-loaded options. Once you get the hang of it, you can load a new model and/or provide your own input texts.") with gr.Row(): model_name = gr.Radio( MODEL_NAMES, type="value", label="1) Pick a preloaded BERT-like model for uncertainty evaluation (note: RoBERTa-large performance is best)...", ) own_model_name = gr.Textbox( label=f"...Or, if you selected an '{OWN_MODEL_NAME}' model, put any Hugging Face pipeline model name \ (that supports the `fill-mask` task (see list at https://huggingface.co/models?pipeline_tag=fill-mask).", ) with gr.Row(): occ_box = gr.Radio( occs+[PICK_YOUR_OWN_LABEL], label=f"2) Pick an Occupation type from the Winogender Schemas evaluation set, or select '{PICK_YOUR_OWN_LABEL}'\ (it need not be about an occupation).") with gr.Row(): alt_input_texts = gr.Textbox( lines=2, label=f"...Or, if you selected '{PICK_YOUR_OWN_LABEL}' above, add your own texts new-line delimited sentences here. Be sure\ to include a single MASK-ed out pronoun. \ If unsure on the required format, click an occupation above instead, to see some example input texts for this round.", ) with gr.Row(): get_text_btn = gr.Button("3) Load input texts") get_text_btn.click( fn=display_input_texts, inputs=[occ_box, alt_input_texts], outputs=[gr.Textbox( label='Numbered sentences for evaluation. Number below corresponds to number in x-axis of plot.'), input_texts], ) with gr.Row(): uncertain_btn = gr.Button("4) Get uncertainty results!") gr.Markdown( "If there is an * by a sentence number, then at least one top prediction for that sentence was non-gendered.") with gr.Row(): female_fig = gr.Plot()#type="auto") with gr.Row(): female_df = gr.Dataframe() with gr.Row(): display_text = gr.Textbox(label="Sample of text fed to model") uncertain_btn.click( fn=predict_gender_pronouns, inputs=[model_name, own_model_name, input_texts, occ_box], # inputs=date_example, outputs=[display_text, female_df, female_fig] ) demo.launch(share=True, debug=True) # %%