(function(doc, html, url) {
var widget = doc.createElement("div");
widget.innerHTML = html;
var script = doc.currentScript; // e = a.currentScript;
if (!script) {
var scripts = doc.scripts;
for (var i = 0; i < scripts.length; ++i) {
script = scripts[i];
if (script.src && script.src.indexOf(url) != -1) break;
}
}
script.parentElement.replaceChild(widget, script);
}(document, '
Using machine learning for categorizing speech errors from people with aphasia
What is it about?
Paraphasias are speech errors often made by people with aphasia. The Philadelphia Naming Test (PNT) is a "picture-naming test", where individuals are shown commonplace pictures and asked to name them. Any paraphasias they make are categorized by a clinician. One of those categories is whether or not the error is semantically similar to the picture (e.g. saying "scissors" when the picture is of a ruler). We have software called ParAlg (paraphasia algorithms) which has tools for categorizing these errors automatically. In this paper, we fine-tune a modern machine learning based language model called BERT for the semantic classifier, to replace the original model we used (word2vec). With this new model, we reduce incorrect classifications by almost half.
Why is it important?
The PNT provides important information about a person with aphasia's speech, but it is very time-consuming for a clinician to administer and score. One hope with developing ParAlg is to automate some of that process. This paper improves upon an important aspect of that automation, the semantic similarity classifier. This work brings us closer to the goal of establishing ParAlg as a useful tool for assessing the speech of people with aphasia in research and clinical practice.