Which news publications have the most positive headlines? (simple sentiment analysis with Textblob)

Inspiration/base dataset

Kaggle provides a great dataset containing news headlines for most major publications. I decided to run some simple sentiment analysis using Textblob, a Python library for processing textual data, that comes with some pre-trained sentiment classifiers. One could of course train their own model, and probably obtain more accurate results overall, but I wasn't able to quickly fine a clean dataset of news headlines tagged with sentiment. Textblob should work fine for comparing the publications relative to each other.

To start, from Kaggle we're given a table as follows:

After poking around the n-counts by year, it looks like 2016-2017 are the only years with decent data >~2000 headlines across each publication. So, I filtered the dataset to these two years:

#years came in string.0 format, just filtering them as they are for now
years = ['2016.0','2017.0']
df = df.loc[df['year'].isin(years)]  

Running our data through Textblob to determine news headline sentiment

Great, so now we have a pretty clean dataframe containing headlines by news publication. Next, I'll create a utility function to log the polarity (ranging from -1 as most negative to 1 as most positive) of each headline

#used to analyze sentiment w/ textblob, leaving the raw output
def analize_sentiment_raw(headline):
    '''
    Function to lassify the polarity of a tweet
    using textblob.
    '''
    analysis = TextBlob(headline)
    #polarity = negative (-1) to positive (1)
    return analysis.sentiment.polarity 

Next, we'll just feed the headlines into the function we created above:

df['SA'] = np.array([ analize_sentiment_raw(headline) for headline in df['title']])

We can now see the polarity score (again ranging from -1 to 1) reflected in our dataframe:

Pivoting the data to rank news publications by headline sentiment

Now that we have sentiment attached to each news headline as reflected in col 'SA' above, we can pivot to summarize the publications with the most positive/negative headlines:

#grab a slice of the original df with just publication and polarity score
df_clean = df[['SA', 'publication']]
#pivoting to grab the average sentiment by publication
df_clean2 = df_clean.groupby(['publication']).mean().reset_index()
#preview the dataframe, sorting by most positive
df_clean2.sort_values(['SA'],ascending=False).head(15)

Below, we can see that (according to Textblob's pre-trained classifier) Business Insider has the most positive headlines on average, with Breitbart coming in at the bottom with the most negative headlines.

Next, lets check out the n-counts on these, and add a view that shows the % of articles that are positive (score >0), negative (score 0, 1, 0)

df_clean['neutral'] = np.where(df['SA'] == 0, 1, 0)
df_clean['negative'] = np.where(df['SA'] < 0, 1, 0)
df_clean2 = df_clean[['publication','positive','neutral','negative']]
df_clean3 = df_clean2.groupby(['publication']).sum().reset_index()
df_clean3.sort_values(['positive'],ascending=False).head(15)

Great, we can now see the # of headlines Textblob identified as positive/neutral/negative. Lets add some %s into this df too:

#fields for positive/neutral/negative percents
df_clean3['positive_perc'] = \
df_clean3['positive']/(df_clean3['positive']+df_clean3['neutral']+df_clean3['negative'])
df_clean3['neutral_perc'] = \
df_clean3['neutral']/(df_clean3['positive']+df_clean3['neutral']+df_clean3['negative'])
df_clean3['negative_perc'] = \
df_clean3['negative']/(df_clean3['positive']+df_clean3['neutral']+df_clean3['negative'])
#sort and preview dataframe
df_clean3.sort_values(['positive_perc'],ascending=False).head(15)

Visualizing news publication sentiment

Using plotly, making a quick bar chart to visualize the news publication sentiment shown in our pivoted data above:

#bar chart
x = df_clean2['publication']
y = df_clean2['SA']

data = [
    go.Bar(
        x=x,
        y=y,
        marker=dict(
            color='#8bffc0f2',
            line=dict(
                color='#d8ffea96',
                width=1.5
            ),
        ),
        opacity=1
    )
]
layout = go.Layout(
      title='News publications with the most/least positive headlines',
    xaxis=dict(
        title='News publication'
    ),
    yaxis=dict(
        title='Sentiment (higher = more positive',
    ),
    margin=go.Margin(
        l=140,
        r=90,
        b=140,
        t=60,
        pad=4
    ),
)
fig = go.Figure(data=data, layout=layout)
py.iplot(fig, filename='news_pub_sentiment')

Voila! We were able to quickly use Textblob to classify and rank sentiment across hundreds of thousands of news headlines.

For those interested, the jupyter notebook with the complete code can be found in our github repository.