As is likely with many of us in this class, this is the
first linguistics class I have ever taken. That being so, I came into this
class knowing virtually nothing of linguistics, and, although I suppose it
should be expected, it has been quite dumbfounding to see exactly how much I
did not know about something that I have been using since before I can
remember. Being introduced to the IPA seemed relatively straightforward; I
already knew how different letters represented different sounds, and it seemed
logical that there would be a "international alphabet," which detailed the sounds of all the world's languages. However, soon afterward
it was revealed that these sounds can be altered by a wide selection of rules,
suprasegmentals, et cetera, and suddenly the illusion of simplicity was gone.
How in the world did I learn all of this at such a young age, without even
being explicitly taught?
I could go on to describe the wonders of a fresh human mind
and cite from a Psych textbook all of the amazing powers it holds that allows
it to learn such a complex communication system, but that is not what I intend
to do. Besides, it appears to me now that the complexities of speech that
baffled me at the beginnings of the quarter are only the tip of the iceberg.
Rather, the true hurdle that must be surmounted to successfully use human
language appears to be its more abstract components.
There are two particular reasons that I say this (however, let
me preface by noting that this not meant as a roadmap of an argument, as I believe
the conclusion is pretty easy to arrive at and already widely held. I am simply
providing a personal anecdote of how it became clear to me). First off, the
activity we performed in class today* (the one in which we applied phonological
rules in different orders) felt like clear-cut computation. Upon completing the
exercise, I immediately began to imagine how easily a computer program could do
exactly what we were doing in the activity. With every discrete unit
categorized and an ordering of rules given, a computer could simply trace
through the words and apply the rules as necessary. This brings me to my second
reason, which is an activity I recently partook in during section for SYMSYS
100, in which we reviewed several manuscripts of failed attempts to pass the
Turing test. The sentences which the computer programs produced were, for the
most part, completely abiding with the rules of English, and, in isolation, perfectly
acceptable. What the computer utterly failed to do was convey any unified sense
of meaning or comprehension through its responses to the interviewer. For example, two questions posed in relation
to one another would be answered in a complete disparate manner, as the program
had no way in which to associate the content matter of the first with the
following question. Inexplicit relations between ideas went completely unaddressed;
how could a computer ever understand the social and cultural implications of
the content of the interviewer's questions? Implications which, of course, are
context-specific and subject to how culture changes over time. Even a simple
connection, such as realizing that the question "Do you prefer red or
white?" does not refer to the colors themselves when following a mention
of wine, is beyond the computer's ability. It then appears, at least from the
perspective of programming, that the true difficulty of language is the complex
web of relations between the ideas which language represent, rather than the
system (language) that is used to communicate these ideas. While it is
difficult to put anything outside the range of technology, I imagine that
surpassing this particular obstacle in artificial language processing will take
a long, long time to complete.
*(10/8)
I definitely agree to your last conclusion, and I believe your observations of our ability to realize and adapt to continuously varying contexts in our language can shed light to an important aspect of language, and moreover intelligence.
ReplyDeleteLanguage is subtle and difficult for computers to understand and imitate because it is a merged collection of several different ‘levels’ of understanding. We can identify spoken language as a series of sounds, but also as a collection of discrete words each formed by grouping such sounds, as a continued expressions of meanings driven from those words, etc. Our intelligence is capable to consider such numerous levels at once. On the other hand, computer and the equipped language are so limited in their quality that a statement in computer language cannot even describe itself (one meta-level). We could do much more: we can use language to describe language, language to describe how to use language to describe language, and so on. Perhaps this is one of the core reasons why computers cannot imitate natural languages so well in spite of their incredible computing abilities.
I completely agree with your conclusion about the extreme complexity to hard-code in context and found your example, "Do you prefer red or white?" to be right on point. As of now, yes, imagining a computer that can understand the "complex web of relations between ideas" and language seems far-fetched. However, I would not count it out.
ReplyDeleteToday we are only used to personal voice recognition assistants like Siri or Cortana, which solely function with voice information. In the near future, I can see these same technologies being paired with computer vision (Computer vision is a field that includes methods for acquiring, processing, analyzing, and understanding images and, in general, high-dimensional data from the real world in order to produce numerical or symbolic information, e.g., in the forms of decisions (from wikipedia)) to further understand context. Imagine your phone can tell you are drinking wine and therefore uses statistical inference to realize you are talking about wine when you ask about red or white. I do not believe that notion is completely crazy. Plus, I would not be surprised if new technology like these begin to show up in the next decade or so considering how fast the industry has been innovating.
While I completely agree with the information stated here about the nature of computers and their lack of ability to communicate in full conversation, I wonder if language is really the idea that we are talking about.
ReplyDeleteAs language seems to have been defined in class it is a way of expressing and communicating ideas, thoughts and other intentions to other people. A deficit in one's ability to communicate via language describes a lack of understanding and ability in that language. For instance though I may not be able to speak Spanish I can still communicate using other methods when I am in Spain and while it might not be the most effective, it will still be a form of communication.
This idea applies to the discussion being had here about computers; there does not seem to be any problem with the computer's ability to use language. As Spencer said the sentences follows the rules of grammar and are clearly communicating ideas. Yet, these ideas do no\t make sense in terms of the conversation that was happening. One might say that the ideas were simply being used in the wrong context. This does not mean that there was a lack of ability to use language though. It was simply a miscommunication or an error in the use of language.
Though this idea might seem trivial I really think it is worth exploring this cognitive issues between using language incorrectly or at the wrong time and simply not being able to use language entirely. This idea can have further implications for SymSys ideas like Artificial Intelligence and also how we communicate with those who have brain disorders that limit their ability to use language clearly, but not their cognitive processes.
I would agree with Nicolas' claim that language processing by computers will get better in the near future. I am currently taking CS 221, and in our first lecture, Percy Liang demo'd something we'd be creating later in the quarter. He had set up a computer with no knowledge of the world, and could ask it questions or input statements. You could ask it, "is Anne's sweater blue?" and it would answer that it didn't know. If you inputed the statement that "Anne's sweater is blue", then it would then agree with you. The computer could recognize new knowledge about the world and internalize it - even when the logic of statements got complicated.
ReplyDeleteWith the massive amounts of information all over the Internet and after witnessing this cool demo, I have a naive confidence that a computer's language processing abilities can only get better - and fast. I agree with Spencer that there is a difference between hard computation and actually understanding the world. As humans, our brain's language processing is based on actual understanding of concepts in the real world. A computer's, however, would be based on probabilistic associations based on cold data. Does this qualify as understanding?
As someone also taking Linguistics for the first time, and having gone through the Turing machine exercise in Symsys 100 I find this post very interesting.
ReplyDeleteIt seems to be that there are many overlying issues with speech recognition by a machine. One of the biggest seems to be a machine's inability to deal with a variety of speakers. This means variations in speaking speeds, pronunciations, etc. These all have to do with technical aspects of understanding a language's differences in its building blocks, or smaller parts. However, this post draws attention to another issue that may not be in the spotlight as much, which is context recognition.
I think that context recognition is a bit more of a difficult problem, though I don't have enough programming experience to justify that being the case for speech recognition programs. In a standard paragraph, the possibilities for which sentences and statements relate to each other are huge, and given that they are whole sentences or ideas, and not just building blocks being said with variation, I think this is the limiting factor for speech recognition. Common use of speech recognition, or even more simply, language recognition (Turing test), typically involve short statements, or at most a few sentences, usually all pertaining to the same idea. Changing up ideas mid statement is a surefire way to get a confused or completely incorrect response from the program being utilized. Unfortunately devising a solution to this doesn't appear to be simple, because of the wide variety of contexts that can be had. Not to be a downer on those working on this, but I think it will require a fairly intelligent artificial intelligence capable of self-learning to solve this issue. That kind of AI doesn't seem like something that will be developed for quite a while.
Artificial Intelligence and speech recognition are such fascinating topics. In theory, the way machine learns to understand language mimics the way an infant gradually acquires the vocabulary, syntax, and contextual information for his native language.
ReplyDeleteThe process of learning a language involves gathering a lot of sensory information from observing speakers and the context in which they use language. To this day, even though most of us are fluent in English, I think in some circumstances we still get stumped in some conversation simply because it's a setting we're used to or some surprise element that went beyond our expectation.
We constantly form expectations about how a conversation goes. In your example of an interview, the interviewer comes in with some expectations about how each question will be answered.
"What are the skills that you bring to this company?"
Expected answer: "I have exceptional planning and organizing skills"
Unexpected answer: "I used to bar tend at the club downtown"
When we are caught by surprise like that, we take a split second to adjust our train of thought and gives this speaker a little extra attention. In the meantime, we think to ourselves, why did this person mention this? Perhaps, this experience has helped him develop certain skills. So we readjust our expectations.
I think it's this process of expectation that drives the logic behind communication. Already Machine Learning algorithms such as Bayesian Classifiers, which describe the way we form and change expectations.
As someone who has learned a second language, I have found first hand that there is a large difference between knowing the words and structure of a language and understanding implicit meanings and subtle variations to which native speakers don’t give a second thought. It takes years to fully learn a language, whether through first- or second-language acquisition. Over the course of those years, countless connections are made determining connotations and implicit meanings of words. These connections allow us to understand and communicate the nuances of our ideas. This seems to be the issue with speech-recognition software: they can have all the grammar, vocabulary, and syntax information programmed into them but they still lack the human capacity for making quick, seemingly unrelated connections and understanding implicit meaning.
ReplyDeleteI'm also in SYMSYS 100 and the complexities of the Turing test have definitely intrigued me as well. After hearing about the Turing test in the class, I looked it up to see if any recent progress had been made and apparently, this past June, there was a program that passed the test. However, as I read more about it, I realized that this program simply found ways to cleverly avoid questions that it wasn't familiar with. So if a question were asked that the computer wasn't programmed to address, then it would come up with a response (usually humorous) that would sound human, but in fact, didn't answer the question at all. Even though this program may have technically passed the test, it still is very far away from imitating a true conversation because it lacks the ability to put the conversation in context and make connections. In order to do this, I believe a program would need to store the conversation in memory and refer to this stored information throughout the rest of the interaction. It definitely would have to be self-learning, as Dylan points out in his comment. Despite these difficulties, I do think that this AI will be here sooner than we probably expect.
ReplyDeleteSimilar to a good chunk of people here, I am also in SymSys100 and I thought it was fascinating to discuss whether passing the Turing test itself is an accurate demonstration of intelligence. Let's think of what happens in a Turing test. You ask a questions and the computer answers. This is very similar to the Chinese Room example by John Searle, which brings to light the difference between semantics and syntax. When a computer can take in a question and answer it as a human would, they have perfected syntax. In many ways, computers are great at syntax but not semantics. Is syntax alone a testament for intelligence? I guess that is something we should consider if we decide to use passing the Turing test as a demonstration of intelligence.
ReplyDelete