What Do Death Row Inmates Say Just Before the State of Texas Kills Them?

The State of Texas hosts a repository of final words spoken by every inmate it has executed since 1982. This writer of 100+ Interesting Data Sets for Statistics suggests — tongue in cheek, I think— that it might be interesting to perform sentiment analysis on it.

Joke or not, I took him up on the challenge, and learned quite a lot — about human nature and the limits of sentiment analysis.

I scraped 422 statements from the Texas Department of Criminal Justice (TDCJ) website, which lists 526 detailed inmate profiles, not all of which include a last statement (some declined to say anything). (You can download my cleaned data set here.)

So what do these inmates — overwhelmingly male and convicted of a Texas capital crime (murder in all its forms) — say in the moments before and just after receiving a lethal injection of pentobarbital?

Getting the Gist of What Was Said

As you’ll read, sentiment analysis is going to prove itself a pretty crude tool in answering this question, so let’s start with a visualization, a random shuffle of each inmate’s last sentence atop his or her photo. This puts faces to words, gives a flavor of what was said, and gives an idea of the number of people we’re talking about.

The color coding is derived from sentiment analysis, which I’ll get to in a minute, but let’s first focus on the content of what the inmates say. What topics do they most often address in these last words?

Three word clouds (the bigger the word, the more frequently the inmates say it) follow. These show most frequent single (top), double (middle), and triple words (bottom). The idea behind including the word pairs and triplets (bigrams and trigrams, in text miner lingo) is to give a better sense of context (“don’t love”, said again and again, would obviously mean something very different than “do love”).

So, in aggregate, the inmates:

1. Tell their families that they love them (yes, using the Texas ya’ll form!) and to stay strong.
2. Apologize to their victims’ families.
3. Pray.

The tendency to pray really becomes apparent when you look for the most frequent four-word phrases. All but one are pieces of Bible verse.

Specifically, they’re from the Lord’s Prayer .

Perhaps none of this is too surprising — especially the urge to pray when there’s a syringe of pentobarbital stuck in your arm — but I was surprised that inmates did not express rage in larger numbers, but will get to those who did. That’s where the crude tool of sentiment analysis comes in.

Trying to Classify What Was Said By Emotion

As I understand it, sentiment analysis works like this. Psychologists have agreed on six universal emotions: happiness, sadness, fear, anger, surprise, and disgust. (Yes, there are only six now, but once there were only five tastes and then umami came along.)

Linguists, in turn, have created corpora of words associated with each emotion. If a document — each inmate’s last words, in this case — contains a lot of words from one corpus — the sack of sad words, for example — then perhaps that speaker is feeling sad.

There are more sophisticated methods for trying to derive meaning from the test document than this approach — ones that take into account what words actually mean in combination — and the limits of this simpler approach are revealed just by examining its picks for top five angriest sentences from all inmate speeches:

1. “Life is too short to harbor feelings of hatred and anger.”
2. “I know you all probably have bitterness and hate for what I did.”
3. “After all these years my people are still lost in hatred and anger.”
4. “I’m not mad or bitter though.”
5. “I come with no hate in my heart or bitterness.”

Yes, these inmates are speaking of anger, but as something they’ve let go of, or as something still directed at them by others. You can’t really call that anger.

So, keeping the limits of this type of sentiment analysis well in mind, here’s how the final statements break into emotion categories.

That’s a lot of joy indeed from people about to die — you’d think we were at Disneyland. These mis-categorizations come from all the expressions of love — using that very word and others like it — toward family members and friends.

I wouldn’t call that joy exactly, but if you read the statements in bulk, you will be struck by the tone of equanimity toward death — serenity you might even call it — that most express.

Where the graph is also sort of right is in showing that few inmates (around 5% — but don’t bet on that exact number) directly express much anger. Those speeches that do express anger really stand out from the more typical pattern of expressing love, apologizing, and praying.

And the sentiment analysis tool I used, simple as it is, does seem adept at pulling out these angry statements. If you use it to filter out the top ten angriest speeches, for example (analyzing each speech as a whole again), you’ll get a few false categorizations, but you’ll mostly get speeches with at least some parts that are white hot with anger.

An example:

Henry Porter #551

This is America’s equal justice. A Mexican’s life is worth nothing. When a policeman kills someone he gets a suspended sentence or probation. When a Mexican kills a police officer this is what you get. From there you call me a cold-blooded murderer. I didn’t tie anyone to a stretcher. I didn’t pump any poison into anybody’s veins from behind a locked door. You call this justice. I call this and your society a bunch of cold-blooded murderers.

Yes, that’s angry, chillingly so. The algorithm did its job here, even if it is prone to quite a lot of mis-classification elsewhere.

Conclusion

So, what have we learned? Word clouds, even if the professionals disdain them, do nicely show what inmates tend to say during execution in Texas. Sentiment analysis can sometimes suss out texts at the extremes of emotion — especially anger, in this case — but, at least in this simple form, it’s prone to a lot of mis-classification.

One last lesson, this one about human nature: Some rare individuals maintain the ability to ham it up, even when the State is in the process of killing them. No algorithm exists yet to find their quotes, but I’ve found and ranked them for you.

Five Best Ham Quotes in the Data Set

1. “Where’s my stunt double when you need one?” — Vincent Gutierrez, #999262.
2. “Tell the guys on Death Row that I’m not wearing a diaper.” — Charles Nealy #999289
3. “I am ready to teleport.” — Lee Taylor, #999344
4. “Y’all kiss my black ass. Let’s do it.”— Brian Roberson #886
5. “Kiss my proud white Irish ass. I’m ready, Warden. Send me home.” — Robert Atworth #999199

__

Open Source Software Gratefully Used in this Post

1. R and the following packages: tm (short for “text mining”), RWeka (for sentence detection), sentiment, wordcloud, gplots.
2. Python and the following modules: OpenCV (for auto-cropping inmate faces), PIL (for making the “last sentences” visualization) and BeautifulSoup (for screen scraping).