Published: Wed 28 September 2016
By Daniel Walter
In blog .
from collections import Counter
import matplotlib.pyplot as plt
from nltk.tokenize import RegexpTokenizer
import pandas as pd
from pprint import pprint
import seaborn as sns
% matplotlib inline
A Presidential Debate in N-grams
While listening to the first presidential debate between Hillary Clinton and Donald Trump, I had a suspicion that a simple n-gram analysis would reveal some interesting insights into their respective styles of speech.
So I grabbed a transcript off the internet (not the one here, but one like it) and got to n-gramming.
(N-grams are phrases of length n from a text or speech. When n is 2, it's called a bigram ; when n is 3 it's called a trigram . We use the term n-gram because it generalizes to whichever N we want.)
with open ( "/data/data/sentiment/debate_transcript.txt" , "r" ) as f :
all_lines = [ line for line in f if line is not ' \n ' ]
What we get at the end of the function below is a list of 'tokens', or words, that a candidate has spoken. It takes care of a few formatting niceties: removing apostrophes so contractions are only one 'token', removing the candidates' names from their own speeches, changing the newline character to the word "NEWLINE", changing everything to lower case, and removing all punctuation.
def tokenize_by_user ( line_list , identifying_string ):
"""Reads in lines, filters based on the identifying string, and returns a list of tokens."""
speaker_lines = [ line for line in line_list if identifying_string in line ]
speaker_text = "" . join ( speaker_lines )
strings_to_drop = [ identifying_string , "’" ]
for element in strings_to_drop :
speaker_text = speaker_text . replace ( element , "" )
speaker_text = speaker_text . lower ()
speaker_text = speaker_text . replace ( " \n " , " NEWLINE " )
tokenizer = RegexpTokenizer ( r '\w+' )
speaker_tokens = tokenizer . tokenize ( speaker_text )
return speaker_tokens
We need a list of tokens for both candidates.
trump_tokens = tokenize_by_user ( all_lines , "Donald Trump:" )
clinton_tokens = tokenize_by_user ( all_lines , "Hillary Clinton:" )
trump_counter = Counter ( trump_tokens )
print ( "Total number of words spoken by Trump: {}" . format ( len ( trump_tokens )))
print ( "Number of unique words spoken by Trump: {}" . format ( len ( trump_counter . keys ())))
print ( "Ratio of unique words to total words: {}" . format ( round ( len ( trump_counter . keys ()) / len ( trump_tokens ), 3 )))
print ( " \n " )
clinton_counter = Counter ( clinton_tokens )
print ( "Total number of words spoken by Clinton: {}" . format ( len ( clinton_tokens )))
print ( "Number of unique words spoken by Clinton: {}" . format ( len ( clinton_counter . keys ())))
print ( "Ratio of unique words to total words: {}" . format ( round ( len ( clinton_counter . keys ()) / len ( clinton_tokens ), 3 )))
The disparity between their respective words spoken is sizable given the relatively tight structure of the debate. Mr. Trump seems to speak more quickly than Secretary Clinton, which is borne out in this dataset. The New York Times reports he spoke for 44:23 compared to her 41:21, so while he spoke for more time overall, he must speak more quickly as well.
pprint ( trump_counter . most_common ( 20 ))
pprint ( clinton_counter . most_common ( 20 ))
As would be expected, these lists are mostly dominated by stop-words. Mr. Trump's most commonly spoken word, 'I', is Secretary Clinton's fourth most commonly spoken.
trump_df = pd . DataFrame ({ 'word' : trump_tokens })
trump_df [ 'wordlength' ] = trump_df [ 'word' ] . apply ( len )
clinton_df = pd . DataFrame ({ 'word' : clinton_tokens })
clinton_df [ 'wordlength' ] = clinton_df [ 'word' ] . apply ( len )
trump_df [ 'wordlength' ] . hist ( bins = 16 , alpha = 0.6 , color = 'red' )
plt . title ( "A histogram of Donald Trump's Word Lengths" )
clinton_df [ 'wordlength' ] . hist ( bins = 16 , alpha = 0.6 , color = 'blue' )
plt . title ( "A histogram of Word Lengths (colors are based on party lines)" )
In general, Mr. Trump used shorter words than Secretary Clinton.
The n-grams part
What we wanted to know was whether there were phrases that popped up frequently for each candidate, so we analyzed for n=2...10 whether there were repeated phrases. My implementation is naive, as it considers the entire spoken corpus of each candidate as one document. The find_ngrams function is a nice little piece of code I found on Scott Triglia's blog.
def find_ngrams ( input_list , n ):
return zip ( * [ input_list [ i :] for i in range ( n )])
def print_ngrams ( input_tokens , n_grams , candidate_name ):
"""Starts with tokens and ends with only the top 15 n-grams that have been repeated
where available"""
print ( "{}'s {}-grams" . format ( candidate_name , n_grams ))
found_ngrams = find_ngrams ( input_tokens , n_grams )
ngramcount = Counter ( found_ngrams )
repeat_keys = [ key for key in ngramcount . keys () if ngramcount [ key ] >= 2 ]
relevant_entries = sorted ([[ key , ngramcount [ key ]] for key in repeat_keys ], key = lambda x : x [ 1 ], reverse = True )
print ( "{} {}-grams were repeated" . format ( len ( relevant_entries ), n_grams ))
for entry in relevant_entries [: 15 ]:
print ( "'{}' : {} times " . format ( " " . join ( entry [ 0 ]) . replace ( "NEWLINE" , "-" ), entry [ 1 ]))
#pprint(relevant_entries[:15])
print ( " \n " )
return len ( relevant_entries )
trump_repeats , clinton_repeats = {}, {}
for n_gram in range ( 2 , 10 ):
clinton_repeats [ n_gram ] = print_ngrams ( clinton_tokens , n_gram , 'Hillary Clinton' )
A few phrases seem notably rehearsed:
Then there is the repeated phrase:
which may have come during an interruption. It's interesting to note that there are no 8 or 9-grams.
for n_gram in range ( 2 , 10 ):
trump_repeats [ n_gram ] = print_ngrams ( trump_tokens , n_gram , 'Donald Trump' )
A few phrases seem notably rehearsed:
Mr. Trump had two repeated 9-grams, which are pretty sizable pieces of text. The last four n-grams all sound like they were soundbites that had been chosen specifically for this debate.
Finally, I thought it would be worth looking at the total number of n-grams for each speaker.
trump_repeat = pd . Series ( trump_repeats )
trump_repeat = trump_repeat / ( len ( trump_tokens ) / len ( clinton_tokens ))
trump_repeat . plot ( kind = 'bar' , color = 'red' , alpha = 0.7 , title = 'Number of repeated n-grams for Donald Trump' )
plt . ylim ( 0 , 500 )
plt . show ()
clinton_repeat = pd . Series ( clinton_repeats )
clinton_repeat . plot ( kind = 'bar' , color = 'blue' , alpha = 0.7 , title = 'Number of repeated n-grams for Hillary Clinton' )
plt . ylim ( 0 , 500 )
plt . show ()
For the images above, I accounted for the difference in total words spoken between the two by scaling Mr. Trump's counts by the same factor his total tokens outnumbered Secretary Clinton's. It is immmediately clear that Mr. Trump had more repeated phrase chunks of every size.
Conclusions?
From this cursory n-gram-based method of analysis, it appears that Mr. Trump had more repeated phrases, even relatively longer phrases, than Secretary Clinton. One explanation could be that Mr. Trump's debate preparation may have emphasized word-for-word talking points more than Secretary Clinton's.