Task: Function to match construct names between codebook and human coded file

[camillastrauss]

Description

Make a function that matches the construct names between the uploaded codebook and the human coded file.

I ran into a problem where my human coded file contained constructs that were similar to my codebook constructs, but not the same. I wrote a function that called an LLM to find the most specific codebook construct that matched each human coded construct. The human coded file contained boolean columns for each construct, so the function renamed each of those columns with the matching codebook construct.

Then the function should make a list of column names for which that row has a value of True. This list will be inputted as the construct_col kwarg when load_human_dataframe() is called later.

If a human code does not match any codebook constructs closely, that human code should be printed and should be excluded from the rest of the analysis (or maybe an option for the user to change it before continuing? idk)

Tasks

• Task 1

Freeform Notes

I know this is confusing so we can work on it together and I can send you the code I've written

[camillastrauss]

rough first draft for the function:

metadata = {}
model = "google/gemini-3.1-flash-lite-preview"

def match_col_names_llm(human_df, codebook, prompt_template = None):

    constructs = list(codebook.keys())
    human_constructs = [x for x in list(human_df.columns) if 'Code: ' in x]
    
    if prompt_template == None:
        prompt_template = """
        return the most specific phrase in the following list of constructs that best matches this column name. 
        construct list: {constructs}
        column name: {col_name}
        
        return only the matching phrase from the construct list, formatted as a string."""
    
   
    for original_human_construct in human_constructs:
        stripped = original_human_construct.split('Code: ')[1].split(' Applied')[0]

        llm_prompt = prompt_template.format(
            constructs = constructs,
            col_name = stripped
        )
        
        match, metadata[stripped] = openrouter_request(llm_prompt, api_key, model = model, temperature=0)     
        rename_cols = {f"{original_human_construct}": f"Code: {match}"}
        human_df.rename(columns = rename_cols, inplace = True)
        # print(match, "      ", stripped)

[camillastrauss]

make list of True human code columns for each row

code_cols = human_df[[*human_df.filter(like='Code: ').columns]]
code_cols.columns = code_cols.columns.str.replace("Code: ", "", regex=False)
code_list_messy = code_cols.dot(code_cols.columns + ", ") # for each row, get list of cols with value True

code_list = [", ".join(list(set(x.replace('"', '').rstrip(", ").split(', ')))) for x in code_list_messy] # removes weird quotes inside strings, removes trailing commas, formats into list, removes duplicates
human_df['Code_List'] = code_list