The SLP and ChatGPT
I’m a nerd, and when a new technology like ChatGPT comes along it gets my attention. ChatGPT developed by OpenAI in San Francisco is a kind of Artificial Intelligence termed a “large language model.” Large language models (LLMs) are trained on massive datasets such as the entire Internet. The LLM can teach itself, without being given explicit instructions. With enough data and a set of algorithms called machine learning, the LLM independently learns words, as well as the relationships between them and concepts behind them. It can, for example, learn to differentiate the two meanings of the word “bark” based on its context. Or how to write a letter in the style of Hemmingway. You can learn more on this technology in the references beneath this post.
The applications for speech therapy seem endless, with the most immediate one being generating therapy stimulus items and materials. This Blog post details my experiences which were very promising, and also highlights some of the limitations I observed.
My conversations were generated with the research version of ChatGPT freely available from the OpenAI website. Some of queries were generated on March 11 with ChatGPT release 3.5, and the remaining queries were generated with ChatGPT 4.0 on March 16.
Creating Stories for Use in Therapy
First I experimented with the technology to create a teaching material for a patient with reading and attention challenges. She has difficulty differentiating “sk” from “sh" while reading text aloud; for example she reads the word “shirt” as “skirt.” The “h” and the “k” are sufficiently similar letter forms (look at them closely), that they are difficult for her to distinguish while in an “s” cluster.
Here is the question I posed to ChatGPT:
Can you create three paragraph story using high percentage of words that start with sk and sh
Here is what ChatGPT gave me:
Skylar was a skilled skier, but she never expected to have a skiing accident. As she was speeding down the steep slope, she suddenly hit a patch of ice and lost control. She shrieked as she fell and crashed into a tree, hurting her shoulder and shattering her skis into pieces. She was lucky to have escaped with only minor injuries, but the experience shook her to the core.
Shane, who witnessed the accident, rushed to help Skylar. He had been skiing on the same slope and had noticed the ice patch earlier, but he didn't warn Skylar in time. He felt ashamed of himself for not doing enough to prevent the accident. He skilfully called for the ski patrol and assisted Skylar until they arrived. He also offered to pay for the damage to her skis and apologized for not being more careful.
After the incident, Skylar and Shane became friends. They often went skiing together and improved their skills. They shared a bond of trust and understood the importance of being cautious while skiing. Skylar learned to be more aware of her surroundings, while Shane learned to speak up when he noticed potential dangers. They both agreed that their friendship started on the slopes because of their shared love of skiing and their experiences with skis and snow.
Figure 1 Request for story with target speech sounds
This worked great as a speech therapy material; I can use this story flexibly in speech sessions. For example, I can ask my middle school patient to read the story aloud as a baseline or a probe. She can underline the “sk'' and “sh” clusters in words and then read the story aloud while attending to the letter shapes. Or I can dictate a paragraph from the story to her and ask her to write while attending to the target sounds.
The story is loaded with the target sounds and gives her many opportunities to practice. The topic and the plot is also interesting for my middle schooler. So many of the off-the-shelf materials from publishers like Teachers Pay Teachers or Super Duper are too immature, or on subjects that are not interesting to her.
ChatGPT gives me the power to create materials targeted to her interests. When we get tired of this story, I can generate a new one easily, and pick a new subject such as puppies or basketball.
A day later, I rephrased the initial question and added a new condition: difficulty with reading plus age range (middle school).
Here is the question I posed to ChatGPT:
Please give me a three paragraph story for a middle school child who is having trouble distinguishing SK from SH while reading.
Here is what ChatGPT gave me:
Once upon a time, there was a middle school student named Alex who loved to read. However, whenever he came across words that contained "SK" or "SH" sounds, he struggled to distinguish between them. It was particularly frustrating for him because it would slow down his reading and make it hard to understand what he was reading.
One day, Alex's teacher noticed that he was struggling with these sounds and decided to help him. She gave him a list of words that contained both sounds, such as "sky" and "shy", and asked him to practice saying them out loud. She also encouraged him to listen carefully to the way his mouth moved when he made each sound.
Over time, with practice and the guidance of his teacher, Alex became more confident in distinguishing between "SK" and "SH" sounds. He no longer struggled with these sounds while reading, and his reading comprehension improved. Alex was grateful to his teacher for taking the time to help him and realized that with a little practice and guidance, he could overcome any challenge.
Figure 2: Request for a story to target sounds during a reading task, middle schooler
This result was more problematic. Children who are having difficulty reading letters aren’t necessarily having difficulty producing the sounds those letters represent. It is questionable that improving articulation of a sound will improve ability to read the sound in a word. We know that written words are processed differently in the brain than verbal language.
For reasons based on research on clinical phonetics and reading research, a speech therapist would never put this story in front of a student.
Create Targeted Word Lists
I also experimented with ChatGPT to generate a list of single words containing “sk” and “sh” for my middle schooler.
Here is what I asked ChatGPT:
Please give me 10 words that have SH or SK in the medial position of the word.
Here is what ChatGPT gave me:
Mishap
Risky
Taskbar
Brisket
Whiskey
Lashkar
Mosquito
Deskwork
Muskrat
Showcase
Figure 3: Request for therapy word lists containing SH and SK
This also worked great as a therapy material with the exception of word 7 which I talk about below.
In another ChatGPT session, I made a request for words of increasing complexity for a 14 year old patient with Cerebral Palsy who is learning to control his speech motor system in order to be more intelligible with longer utterances.
Here is what I asked ChatGPT:
Please give me a list of increasingly complex multi-syllable words by developmental level.
Here is what ChatGPT gave me:
Sure, here is a list of increasingly complex multi-syllable words grouped by developmental level:
Early Childhood (ages 2-5):
Animal names: hippopotamus, rhinoceros, giraffe
Color names: turquoise, magenta, chartreuse
Food names: spaghetti, lasagna, avocado
Nature names: butterfly, dragonfly, hummingbird
Toy names: xylophone, trampoline, Play-Doh
Middle Childhood (ages 6-12):
Geographical names: Mississippi, Saskatchewan, Kilimanjaro
Scientific terms: photosynthesis, gravity, metamorphosis
Historical terms: Renaissance, Industrial Revolution, colonization
Math terms: algorithm, fraction, circumference
Artistic terms: impressionism, expressionism, surrealism
Adolescence and Adulthood (ages 13+):
Medical terms: diagnosis, prescription, anesthesia
Legal terms: subpoena, affidavit, jurisdiction
Philosophical terms: existentialism, nihilism, stoicism
Literary terms: allegory, metaphor, protagonist
Sociological terms: stratification, discrimination, globalization
Figure 4: Request for therapy word lists by developmental level
I found this list to be almost useless. All these words are too complex for the age levels listed, especially for a child with a developmental disorder or a language disorder. I wonder what would have happened if I specified “language disorder” or “motor speech difficulty” in the request?
Conclusion: Great Promise and Theoretical Limitations
Speech Therapists are creative and often make their own materials. They are typically early adopters of technology such as the iPad (Fernandez, 2011) . Given my initial experiences with ChatGPT 3.0 and 4.0, I see great promise with the technology, and suspect the uptake by speech therapists will be very quick. but I also see serious theoretical limitations. These limitations can be side-stepped by an experienced SLP. And by the way, OpenAI is transparent about ChatGPT limitations on the Query page and elsewhere on their website.
It is also important to understand that ChatPGT is still experimental, and it constrained by the training dataset.
Here is a list of ChatGPT limitations for SLPs that I discovered (I’m sure I just scratched the surface).
Lacking judgment about developmental norms
I laughed when ChatGPT suggested turquoise, magenta, and chartreuse as color words for 3-5 year old children (see the last request, Figure 4). These colors might be useful for a 12-year old future fashion designer, but not for your average four year old who is still learning primary colors (“red” “green” “blue”). Also what 4-year old can even say the word “chartreuse” much less identify it? Queries based on developmental norms will likely yield problematic results for now.
Lacking a research-based understanding of language disorders and treatment
The second story shown in Figure 3 is instructive. ChatGPT queries based on treatment goals, or nuanced topics like reading or the phonology of the English language are probably not going to yield good results at this point. (I wonder if the training dataset could eventually include research literature or current textbooks.)
Incomplete understanding of how language is represented phonologically
Take a look at Figure 2, the list of words with “sk” clusters. Mosquito (word 7) contains the “sk” sound when you produce it verbally, but isn’t spelled with “sk” - it is spelled with “qu.” Mosquito is an incorrect teaching word for my student, who is confusing “sk” and “sh” while she is reading and writing. Queries based on phonology are probably going to be problematic for now.
Developmentally inappropriate or incorrect semantics
Back to the word list in Figure 4. A xylophone is not an object that most children are exposed to, or have in their homes, nor is it a toy. It is a musical instrument. Also I would not consider “trampoline” to be a toy in the strict sense. The semantic understanding of the algorithm seems limited, and queries based on the semantics of actual experience may end up being “off” or inappropriate somehow.
Try it yourself
ChatGPT is easy and fun, and very addictive. Generating text-based therapy materials is a good application of this technology for speech therapists, as long as they are mindful of the limitations. Give it a try to generate therapy materials while the service is still free. While you are learning and experimenting, I’m sure you will think of other applications of ChatPGT for speech therapy. Please share them in the comments below!
Follow these easy steps:
1. Click this link go to the website
2. Click the box “Try on ChatGPT Plus” at the bottom of the screen.
3. Select “for personal use” and go through the steps to create an account.
4. Type in your questions in the blank box at the bottom of the page, and click the triangular “paper airplane” icon at the right of the box to submit.
Wait for the answer to appear—it might take a few seconds.
Sources
On technology adoption/app development by speech therapists:
Fernandes, B. (2011). Is the iPad revolutionizing speech therapy? From an SLP & APP developer. Leader Live.
Du, Y., & Tekinbas, K. S. (2020). Bridging the gap in mobile interaction design for children with disabilities: perspectives from a pediatric speech language pathologist. International Journal of Child-Computer Interaction, 23, 100152.
Du, Y., Grace, T. D., Jagannath, K., & Salen-Tekinbas, K. (2021). Connected play in virtual worlds: communication and control mechanisms in virtual worlds for children and adolescents. Multimodal Technologies and Interaction, 5(5), 27.
On machine learning: Coursera has a lucid explanation of the three types of machine learning in this blog post
On a type of machine learning where the network seems to “learn on its own”: MIT news has a understandable blog post
On the difference between ChatGPT and search: Shabham Singh of Demand Sage has a great blog post
On large language models: Angie Coi at NVIDIA has a informative explanation in her blog post
On the new ChatGPT-4: New York Times has a pithy, fun-to-read article.
On the benefits of using ChatGPT in Healthcare: New York Medical College has a thoughtful blog post written entirely by ChatGPT that discusses some of the pitfalls (patient privacy, data protection, biased decision making) and benefits (providing health information, serving as a virtual assistant, automating certain tasks).
