(Un)natural Language Processing

I'm planning for some of my posts to get into the technical details of AI, some will just be my thoughts and opinions, and others will try to explain an aspect of AI in teams easy for everyone to understand. This is one of those posts.

Natural language processing (NLP) attempts to understand, interpret, and generate human language. In my observation, it does fine at "generate," okay (I guess) at "interpret," but does not really "understand." NLP is based on statistics. It tries to guess what's next based on what word (or words) most often follow what you've typed. That is, most often in the training data. If the training data is really generic, then you might get one response, but if it's contextualized then you might get a different one.

Here's an example: If I were to write "I had several bats in my garage, which..." the next few words could be "I would bring to baseball practice." or "I had to get removed by a specialist." The difference is context. A model trained on sports writing would make the former assumption. A model trainined on nature and wild animal enounters might make the latter. As a reader, you'd know which the writer meant because you'd already have that context.

Another issue is with pronouns. That is, GenAI is terrible at them! Take these two pairs of sentences:

  • The box did not fit in the suitcase. It was too small.
  • The box did not fit in the suitcase. It was too big.

    The word "it" in the second sentence obviously refers to the suitcase in the first pair and the box in the second pair. NLP algorithms have a REALLY hard time figuring that out, becuase they don't actually understand the meanings behind the words. To know which is "it" in the second sentence, one also needs to know what it means to fit in something and (in the case of the first pair) that the thing into which the other thing is going to fit cannot be too small.

    This is why the output of most GenAI programs seems...unnatural. It's not grammatically incorrect, just unnatural. So, if you want to use GenAI to help with your writing, go ahead! Just make sure that you take the time to edit it so that it's not obvious.

    • Comments

    Post a Comment

    Popular posts from this blog

    Welcome to my Blog!

    Image
    My goal for this blog is to share my thoughts on Artificial Intelligence. I started a Ph.D. program in AI at Oregon State University last fall and have been thinking about starting a new blog ever since. There's just too much on my mind and I need to get it out and organized! My first blog post was almost 19 years ago was about my martial arts training. I may have to sneak some martial arts training posts into this one, becuase the subject is never far from my mind but that won't be the main focus. The title of this blog, Actual Intelligence, and the URL, actualintel.blogspot.com , are a double entendre of a sort since I work at Intel Corporation . I *may* share some thoughts about Intel here, but probably not many. I don't follow the tech news very closely and therefore don't want to risk publishing anything that's not already public. I expect that most of my posts will be about what I'm doing at OSU. I may recycle some old writing and/or blog posts th...
    Read more

    My Love/Hate for GenAI

    I have a love/hate relationship with Generative Artificial Intelligence (GenAI). This type of AI what most people think of when (if?) they think about AI. It's what's using large language models (LLMs) to build chatbots, create fake restaurant receipts, and help students cheat on their assignments. First, let me explain what I love about GenAI. It's great at helping people find information, the proverbial needle in a haystack. I love the idea of building an LLM from a company's HR policies and turning that into an internal-only chatbot that employees can use to get their questions answered. I love the idea of building an LLM based on my own writing so that I can generate content "in my own words" using GenAI. (Whether using it is against academic integrity standards is a different discussion.) I love the idea of artists using GenAI to customize their own work to make it more commercially viable. These are all ways that GenAI can be used to make pe...
    Read more

    The Measure of Life

    Image
    This is an essay I wrote in 1995 for an RIT undergraduate school course titled "Science, Technology, and Values." Being a computer engineering student and a big fan of Star Trek, I am fascinated by the character Lt. Commander Data, who is an android. During several "Star Trek: The Next Generation" episodes, more than I could count, Data expresses desire to become more human. In one particular episode, Data's cooperation is sought in order that he take part in an experiment that would involve taking him apart for study. Some guy shows up on the Enterprise and has transfer orders ready for Commander Data so that he might be transferred to his lab and be shut down for study and experimentation, a process that could cause Data much harm, especially since this guy does not fully understand the work of Dr. Sung, Data's creator. Needless to say, Data does not want to participate and decides to resign so that he may avoid a transfer of duty; at thi...
    Read more