👋 Hi, I’m Melissa and welcome to my biweekly Note, Language Processor. Every other Tuesday, I dig deep into language & behavior, the limits of technologies, and the connection between what people say and do. Once a month, Signal Lab takes over and reports on these topics for startups and early-stage investors.
Predicting violence using any method of risk assessment or set of risk factors presents serious pitfalls for inaccuracy or prejudice.
Violence is a complex behavior linked to a host of factors. Mental and personality disorders, demographics, and dispositional characteristics all contribute.1
What has long been perceived but not elucidated is that language captures and demonstrates many of the deeper factors involved in violence. Peer-reviewed academic research establishes links between linguistic choices and psychological disposition, and from psychological disposition to behavior, but the chain between these had not been joined, in part because of the methodological difficulty entailed.
Language is a goldmine of insights into behavior — past, present, and future.
We asked whether language could be used as a medium to better understand and quantitatively measure criminogenic needs (i.e., dynamic, changeable, and person-centered risk factors directly linked to criminal behavior and recidivism) otherwise inaccessible to risk assessment. Could inmate language in parole hearings be used as a predictor for future violence?
A New Parole Assessment Using Language
The question was all the more appropriate because inmate language in parole hearings was already being used as a predictor for future violence, as the key factor in parole release decisions. The predictive value of inmate hearing performance was either an enduring misconception or an acutely underdeveloped intuition.
In either case, parole commissioners were sorely lacking evidence-based assistance in their most central function.
The application of behavioral linguistic intelligence recognized the intuitive and widely-accepted importance of inmate statements in assessment, respected the inmates themselves by taking them on their own terms, and tested the validity of language as an indication of success or failure on parole.
It was the first evidence-based assessment of parole hearing data in the United States.
Linguistic metrics were sourced and operationalized from criminological, social-cognitive, and psychological theories of violence, and tested in a computational model. The model assessed the same inmate statements considered by parole board members.
By quantifying language structures, we found that how inmates presented their crimes2 and future prospects correlated with parole outcome at a level of statistical significance.
What Matters?
Language-based correlations with non-recidivism (i.e., success on parole) were very different than what the parole board expected and assumed. For example, the framework used in California for parole hearing performance evaluation showed almost no accuracy when we tested it.3 In other states, qualities parole boards had favored, such as expressions of remorse and factual consistency, were shown not to be indicative of success on parole.
Our project uncovered strong linguistic correlations with parole outcome that had never before been observed or reported. Some of the linguistic patterns display behaviorally-linked psycho-social deficiencies. For example:
COLLECTIVIZING PRO-FORMS: “Everyone went” v. “Some friends went”
AMBIGUOUS REFERENCE: Untraceable connection between signifier and sign (what does “it” refer to?)
SEMANTIC BLEACHING: “I used it” v. “I stole the car”
PRESUPPOSITION: “I know I won’t do it again” v. “I don’t want to do it again”
Accuracy
While recidivism predictors have been tried and tested over decades of criminological research in a field devoted purely to such enquiries, our project using linguistic variables — the first of its kind — demonstrated a remarkable accuracy rate.
Our approach was twice as predictive as methods in place, exceeding even the gold standard for prediction thresholds,4 and represents the most accurate dynamic parole assessment in over 80 years.
Ultimately, the study’s findings supported the value of assessment by interview, as the study statistically demonstrated that linguistic choices can reflect psychological and criminological insights into recidivism with high levels of accuracy.
It made the stronger point that parole assessment requires new approaches to improve accuracy in the evaluation of inmate performance.
See you in two weeks.
Inmate: The other thing that I want to, I want to say is, um, I understand you looking all the way back, and I’ve heard the board say any number of times that my past is my best indicator of what I’ll do in the future. To me, that seems – while I understand the value of doing that, I don’t necessarily agree entirely with that. That’s like saying you’re never going to change. This is what I was 30 years ago, and that’s what I’m always gonna be for the rest of my life. I did a terrible thing 30 years ago, and, yeah, I’ve had my problems and I’ve done a lot of things that I’m not proud of. But that was me 30 years ago. If you want to look at my past up to there, I wholeheartedly agree you should. But I also ask that you will look also what I’ve done after that up to this point. I’ve worked hard. I try to stay out of trouble. I do what I believe is the right thing to do, and I’ve worked very hard to try to improve myself. And I do it because I want to. I, I recognise that, yeah, there’s things that I’ve needed to do, but I did it because it’s what I wanted to do. I want to be a better person. I want that for me, and I want that for the people that I love, and I’ll do anything for that, and that’s all I ask. That’s all I have to say.
1 McNiel, Borum, Douglas, Hart, Lyon, Sullivan, & Hemphill, 2002, p. 8
2 Violent crimes, i.e., murder, rape or sexual assault, and/or robbery
3 Stanford Criminal Justice Center. Weisberg, Mukamal, Segall, 2011.
4 70% accuracy (Latessa & Lovins, 2010)