umls_similarity_article_p3.py

# Create evaluation DataFrame containing BIOSSES pairings and additional information for evaluation
evaluation = pd.DataFrame(biosses_meta.loc[:,['Text1', 'Text2', 'Avg', 'Var']])
# Add CUI information
evaluation['Text1_CUIs'] = evaluation['Text1'].apply(lambda x: biosses_texts.loc[x,'UMLS_CUIs'])
evaluation['Text2_CUIs'] = evaluation['Text2'].apply(lambda x: biosses_texts.loc[x,'UMLS_CUIs'])
# Add UMLS terms
evaluation['Text1_UMLS_TERMS'] = evaluation['Text1'].apply(lambda x: biosses_texts.loc[x,'UMLS_Terms'])
evaluation['Text2_UMLS_TERMS'] = evaluation['Text2'].apply(lambda x: biosses_texts.loc[x,'UMLS_Terms'])
# Add texts for evaluation purposes
evaluation['Text1'] = evaluation['Text1'].apply(lambda x: biosses_texts.loc[x,'Text'])
evaluation['Text2'] = evaluation['Text2'].apply(lambda x: biosses_texts.loc[x,'Text'])

# Helper function for multiprocessing
def split_process(function, dataframe, params, num_processes=1):
    """Splits DataFrame into multiple chunks and processes them in multiple processes
      Args:
          function (function): Function to be called
          dataframe (DataFrame): DataFrame to be chunked and processed
          params (dict): Dictionary containing settings for function
          num_processes (int): Number of processes
      Returns:
          pandas.DataFrame: DataFrame containing the result
    """
    len_data = dataframe.shape[0]
    num_rows = int(len_data/num_processes)
    chunks = [dataframe.loc[dataframe.index[i:i + num_rows]] for i in range(0, dataframe.shape[0], num_rows)]
    pool = mp.Pool(processes=num_processes)
    func = partial(function, params)
    results = pool.map(func, chunks)
    return pd.concat(results)
  
# Text similarity function
def calc_cui_similarities(params, data):
    """Calculates average of similarities between all CUIs in a text
      Args:
          params (dict): Settings
          data (DataFrame): DataFrame to be processed
      Returns:
          pandas.DataFrame: DataFrame containing the result
    """
    for idx,row in data.iterrows():
        similarities = []
        pairings = list(itertools.product(row[params['col_0_name']], row[params['col_1_name']]))
        for pairing in pairings:
            if pairing[0] == pairing[1]:
                similarities.append(1)
            else:
                try:
                    similarities.append(cui_table.at[pairing, params['similarity']])
                except:
                    pass
        if len(similarities) > 0:
            data.at[idx, "{}_cnt".format(params['output_name'])] = len(similarities)
            data.at[idx, params['output_name']] = mean(similarities)
        else:
            data.at[idx, "{}_cnt".format(params['output_name'])] = 0
            data.at[idx, params['output_name']] = 0.0

    return data

# Now calculate similarity and the error, defined as | expert rating - similarity | for each pair of texts in the BIOSSES dataset.
for similarity in ['path', 'lch', 'wup']:
    output_col = "{}_similarity".format(similarity)
    error_col = "{}_error".format(similarity)
    parameters = {
        'col_0_name':  'Text1_CUIs', 
        'col_1_name':  'Text2_CUIs', 
        'similarity':  similarity,
        'output_name': output_col
    }
    evaluation = split_process(calc_cui_similarities, evaluation, parameters, num_processes=10)
    evaluation[error_col] = abs(evaluation['Avg'] - evaluation[output_col])