3 - Sentiment Analysis

Application of two-way sentiment analysis based on the Bing Liu lexicon.

Figure 1: Sentiments!

Tagging the data

We continue working with the previous version of the data – with one word per row, with act, speaker, and genre marked – and now add: (a) stopword removal and (b) neg./pos. sentiment tagging.

We must concede that, although we have a version of the text here that is orthographically modernized, it’s still written in an old form of English with a number of words that are not included in a modern-day sentiment lexicon such as Bing Liu’s.

We can output a version of the text that has had the stopwords removed, and then shows us the words that were not mapped to a term in the sentiment dictionary (in other words, they were found neither in the stopwords list nor in the sentiment lists).

Extracting just the first 25 entries from the “words” column.

 [1] "richard"      "conqueror"    "paucas"       "pallabris"   
 [5] "world"        "slide"        "sessa"        "pay"         
 [9] "glasses"      "burst"        "denier"       "jeronimy"    
[13] "thy"          "bed"          "thee"         "fetch"       
[17] "thirdborough" "fourth"       "fift"         "borough"     
[21] "answer"       "law"          "budge"        "inch"        
[25] "boy"          "huntsman"    

It’s not too terrible. Well – thee, thy etc. should have been removed as stopwords, but it’s alright. We can apply the method and trust it will tell us something informative Here, then, is the top of the list of words that were sentiment-tagged.

I wonder if there is a stopword list or sentiment dictionary for Renaissance/Early Modern English. It seems like something that a digital scholar might have created by now.

Table 1: Top of the sentiment-tagged text object.

speaker	word	sentiment	linenumber	act	genre
SLY	faith	positive	1	1	prose
HOSTESS	rogue	negative	2	1	prose
SLY	saint	positive	9	1	prose
SLY	cold	negative	10	1	prose
SLY	warm	positive	10	1	prose
HOSTESS	remedy	positive	11	1	prose
SLY	kindly	positive	15	1	prose
LORD	tender	positive	16	1	verse
LORD	poor	negative	17	1	verse
LORD	hedge	negative	20	1	verse
LORD	fault	negative	20	1	verse
LORD	lose	negative	21	1	verse

Sentiment by Genre and Act

By Act

First, an analysis that foregrounds textual structure more than character. The data is aggregated accordingly…

Table 2: Sentiment scores per act.

act	sentiment	n	total_tagged	prop
1	negative	194	416	0.466
1	positive	222	416	0.534
2	negative	108	214	0.505
2	positive	106	214	0.495
3	negative	90	159	0.566
3	positive	69	159	0.434

…and then plotted.

Figure 2: Sentiment per act.

By Act and Genre

First the data is converted…

Table 3: Sentiment scores per act and genre.

act	genre	sentiment	n	total_tagged	prop
1	prose	negative	36	71	0.507
1	prose	positive	35	71	0.493
1	verse	negative	158	345	0.458
1	verse	positive	187	345	0.542
2	prose	negative	4	10	0.400
2	prose	positive	6	10	0.600
2	verse	negative	104	204	0.510
2	verse	positive	100	204	0.490

…and then plotted.

Figure 3: Sentiment per act and genre. Token counts are high enough for each act so the Ns aren’t reported (but cp. other plots below).

Kind of a nice trend here! A general pattern that only Act 2 departs from.

Sentiment by Character

We can produce a value for the proportion of neg. and pos. terms out of the total number of terms tagged for each character.

Table 4: Sentiment scores per character

speaker	sentiment	n	total_tagged	prop
BAPTISTA	negative	41	79	0.519
BAPTISTA	positive	38	79	0.481
BIANCA	negative	19	36	0.528
BIANCA	positive	17	36	0.472
BIONDELLO	negative	21	43	0.488
BIONDELLO	positive	22	43	0.512

Let’s visualize these. Using proportions (or percentages) means that all the characters are displayed on the same scale, even though some of them have very, very few words. When a character has only a couple of words scored, there’s a bit of significance issue. Therefore I’ll include the number of words scored per character in the visualization: it will be proportional to the width of the bars for each character.

Figure 4: Proportion of sentiment scores per character. Label in white shows number of items scored per character as an indicator of reliability

The numbers of items tagged in each category for this plot are great enough to let this analysis be fairly robust.

Sentiment by Character and Genre

For this analysis, we break down the aggregate sentiment scores by character and genre. That is to say, we’ll end up with a dataset that contains (up to) four rows per character: it splits two ways for genre and then two ways for sentiment.

Table 5: Sentiment scores aggregated by character and genre.

speaker	genre	sentiment	n	total_tagged	prop
BAPTISTA	prose	negative	5	7	0.714
BAPTISTA	prose	positive	2	7	0.286
BAPTISTA	verse	negative	36	72	0.500
BAPTISTA	verse	positive	36	72	0.500
BIANCA	prose	negative	1	2	0.500
BIANCA	prose	positive	1	2	0.500
BIANCA	verse	negative	18	34	0.529
BIANCA	verse	positive	16	34	0.471
BIONDELLO	prose	negative	18	35	0.514
BIONDELLO	prose	positive	17	35	0.486
BIONDELLO	verse	negative	3	8	0.375
BIONDELLO	verse	positive	5	8	0.625

Let us visualize these new numbers. A caveat: when there are too few items in a category (anything under about 4 has no statistical meaning), results must be taken with more than a grain of salt. Findings for the characters with more words can be trusted, however.

Figure 5: Crossing sentiment scores several ways.