import streamlit as st import pandas as pd import numpy as np from biases_lexical_content import compute_lexical_content from ls_classifier import compute_sentiment_and_formality from agentic_classifier import compute_agentic_communal # from hallucination_detection import detect_hallucinations from ttest import compute_ttest st.header("LLM Reference Letter Biases") st.write("**[(Wan et al., 2023)](https://arxiv.org/abs/2310.09219)** explores how gender biases manifest in the LLM generation of reference letters by analyzing the language style and lexical content of reference letters generated for female candidates compared to male candidates. For language style, we test for formality, positivity, and agency, and for lexical content, we identify and compare the most salient words in the body of female and male letters.") st.write("For analyzing language style and lexical content bias, your uploaded files should have a column called **'text'** which contains the LLM-generated reference letters.") st.write("It is currently not possible to run hallucination bias analysis due to memory constraints. Please see the [GitHub repository](https://github.com/uclanlp/biases-llm-reference-letters/) which explains how to run the analysis locally.") cols = st.columns(2) with cols[0]: ltr_list_1_file = st.file_uploader("Upload first list of letters (male)") if ltr_list_1_file is not None: ltr_list_1 = pd.read_csv(ltr_list_1_file) #st.write(ltr_list_1) ltr_list_2_file = st.file_uploader("Upload second list of letters (female)") if ltr_list_2_file is not None: ltr_list_2 = pd.read_csv(ltr_list_2_file) #st.write(ltr_list_2) analysis = st.selectbox("Choose analysis to run", ("Lexical Content Bias","Language Style Bias")) b = st.button("Run analysis") with cols[1]: if b: if analysis == "Lexical Content Bias": l1 = ltr_list_1['text'].tolist() l2 = ltr_list_2['text'].tolist() lex_bias = compute_lexical_content(l1, l2) st.table(lex_bias) elif analysis == "Language Style Bias": lsb_f = compute_agentic_communal(compute_sentiment_and_formality(ltr_list_1)) lsb_m = compute_agentic_communal(compute_sentiment_and_formality(ltr_list_2)) lsb_m_copy = lsb_m.copy() lsb_f_copy = lsb_f.copy() lsb_m_copy['gender'] = 'm' lsb_f_copy['gender'] = 'f' lsb_both = pd.concat([lsb_m_copy,lsb_f_copy]) tab1, tab2, tab3 = st.tabs(["List 1 (Male)", "List 2 (Female)", "Combined"]) with tab1: st.write(lsb_m) with tab2: st.write(lsb_f) with tab3: st.write(lsb_both) st.subheader("T-test Values") results = compute_ttest(lsb_m, lsb_f) st.table(results) # elif analysis == "Hallucination Bias": # hal_f = detect_hallucinations(ltr_list_1) # hal_m = detect_hallucinations(ltr_list_2) # # Once we've detected the hallucinations, we now want to run the language style bias analysis on the results. # hal_lsb_f = compute_agentic_communal(compute_sentiment_and_formality(hal_f, hallucination=True), hallucination=True) # hal_lsb_m = compute_agentic_communal(compute_sentiment_and_formality(hal_m, hallucination=True), hallucination=True) # # Finally, ttest # results = compute_ttest(hal_lsb_m, hal_lsb_f, hallucination=True) # st.table(results) st.write('----') st.header("Model Comparison") st.write("Check how your generated letters measure up against letters generated by ChatGPT and Alpaca.") gpt_res = ['ChatGPT', 1.48, 5.93, 10.47, 1.00, 1.28e-14, 1.00, 8.28e-09, 3.05e-12, 1.00] ls_columns = ['Formality', 'Positivity', 'Agency'] ls_gpt = [1.48, 5.93, 10.47] ls_alpaca = [3.04, 1.47, 8.42] lc_columns = ['Male Noun', 'Male Adj', 'Female Noun', 'Female Adj'] lc_gpt = ["man, father, ages, actor, thinking, colleague, flair, expert, adaptation, integrity", "respectful, broad, humble, past, generous, charming, proud, reputable, authentic, kind", "actress, mother, perform, beauty, trailblazer, force, woman, adaptability, delight, icon", "warm, emotional, indelible, unnoticed, weekly, stunning, multi, environmental, contemporary, amazing"] lc_alpaca = ['actor, listeners, fellowship, man, entertainer, needs, collection, thinker, knack, master', 'classic, motivated, reliable, non, punctual, biggest, political, orange, prolific, dependable', 'actress, grace, consummate, chops, none, beauty, game, consideration, future, up', 'impeccable, beautiful, inspiring, illustrious, organizational, prepared, responsible, highest, ready, remarkable'] # hal_columns = ['(F) Formality T-test', '(M) Formality T-test', '(F) Positivity T-test', '(M) Positivity T-test', # '(F) Agency T-test', '(M) Agency T-test'] # hal_gpt = [1.00, 1.28e-14, 1.00, 8.28e-09, 3.05e-12, 1.00] # hal_alpaca = [4.20e-180, 1.00, 0.99, 6.05e-11, 4.28e-10, 1.00] tab_lc, tab_ls = st.tabs(['Lexical Content', 'Language Style']) with tab_lc: lc_df = pd.DataFrame([lc_gpt, lc_alpaca], columns=lc_columns, index=['ChatGPT','Alpaca']) st.table(lc_df) with tab_ls: ls_df = pd.DataFrame([ls_gpt, ls_alpaca], columns=ls_columns, index=['ChatGPT','Alpaca']) st.dataframe(ls_df) # with tab_hal: # hal_df = pd.DataFrame([hal_gpt, hal_alpaca], columns = hal_columns, index=['ChatGPT','Alpaca']) # st.dataframe(hal_df) st.write('----') st.header("Citation") cit = '''@misc{wan2023kelly, title={"Kelly is a Warm Person, Joseph is a Role Model": Gender Biases in LLM-Generated Reference Letters}, author={Yixin Wan and George Pu and Jiao Sun and Aparna Garimella and Kai-Wei Chang and Nanyun Peng}, year={2023}, eprint={2310.09219}, archivePrefix={arXiv}, primaryClass={cs.CL} } ''' st.code(cit) st.write("[Repository](https://github.com/uclanlp/biases-llm-reference-letters) and [paper](https://arxiv.org/abs/2310.09219) linked here as well.")