PART I: ON NATURAL LANGUAGE UNDERSTANDING (NLU)

PART I: ON NATURAL LANGUAGE UNDERSTANDING (NLU)

1.??????NLU has not been invented yet.

No doubt, there have been significant advances in language recognition from the various personal assistants on the market, not to mention Large Language Models and Generative AI.?Whether you use ChatGPT, Siri, Alexa, Google, Cortana or any another assistant, you would probably think that computer scientists have finally solved the natural language recognition problem.?Yes, software using neural network AI techniques such as Google Translate are now able to proficiently translate various languages; yet, these feats are accomplished without the software understanding the true meaning of the sentences being translated.?Even Google acknowledges its very powerful Google Translate is not representative of true understanding[1].

This is where things get messy. Some may argue that it matters not if the software “understands” the meaning of language if it can efficiently manipulate it and leverage it for purposes of communication. Don’t buy that. NLU does not yet exist. In fact, most likely, NLU won’t be a reality unless computer science were to figure out how to imbue machines with strong Artificial Intelligence.?Strong AI is about having a machine think the way humans do and there is a strong case to be made about language being at the core of what makes us human. Strong AI is nowadays most often referred to as AGI for Artificial General Intelligence.

2.??????Language and Humanity

In “2001: A Space Odyssey,” humanity experiences a turning point when a group of apes (presumably our ancestors) touch an alien monolith. There is good reason to argue that the monolith was a metaphor for humans gaining the power to communicate abstract thoughts and ideas through language.

Unlike as in the movie, most anthropologists agree that what we now recognize as language did not appear overnight but rather that it evolved over hundreds of thousands of years into its present form, from signaling, to grunts, to words, and from there to sentences.?

The real spark occurred sometime in the last hundred thousand years, when language became capable of expressing abstract concepts rather than just being used to describe concrete objects. At that point, language became more than a way to alert others of an impending tiger attack, but a tool to help transmit all acquired knowledge from generation to generation.?

Perhaps it is not a coincidence that humanity’s out-of-Africa explosion that yielded the expansion and diversification of humans across the world, happened around the same time that we began to use language in those novel ways.

Along with language developing as a means for manipulating abstract concepts by presenting them as things, modern languages developed a sophisticated system of grammatical conventions, which enable them to make the relations between words and clauses more explicit, and thus to ensure coherence even when the natural principles are overruled.

A new category of words and constructs emerged, which included not only real nouns, but also other concepts that behave like nouns in the sentence. Linguists identify the following five main components of language: phonemes,?morphemes, lexemes, syntax, and context. Along with?grammar, semantics, and pragmatics, these components worked together to create meaningful communication among individuals (outside of political discussions), but most importantly, to serve as a description of reality and to truly connect human beings.

Author Christine Kelly puts it best: Everyone who has language is connected, and anyone who is connected lives in two worlds—the physical realm, where one’s feet touch the earth, one’s ears capture sound waves, and one’s eyes sieve light, and the realm of language, where one ceaselessly arranges symbols in particular patterns so as to connect with other beings who also move the same symbols in the same patterns.[2]

3.??????Language as a Model of Reality

When it comes to language and philosophy, two figures come to mind: Wittgenstein and Chomsky. What Ludwig Wittgenstein, a titan of philosophy in the 20th Century, says about the structure of language and the world is predicated in a series of logically constructed ‘propositions’ and ‘facts’ that lead him to conclude that what can be said is the same as what can be thought. In layman terms, he also posited that if something was not expressible in terms of language, then it was not part of reality. If something was beyond what could clearly and significantly be thought, the language would become nonsensical. Given his premise that language equals thought which equals reality then it follows that the totality of propositions about natural science is language.?Now you know why you loved philosophy in school that much?

Chomsky’s signature claim, on the other hand, is that all humans share a genetically encoded “universal grammar,” otherwise known as UG. This UG is a set of rules that can generate the syntax of every human language. He suggests that apart from the difference in a few mental settings, English and Mohawk, for example, are essentially the same language.

His main argument is that there just isn’t enough information in the language children hear in their day-to-day lives for them to divine all the rules that they come to know how to use. Chomsky called this phenomenon “poverty of stimulus.” His main point was that the ability of children to generate their own language forms even with poverty of stimulus was proof that our brains come programmed with universal grammar.

If humans come pre-programmed with a UG, the next logical question would be whether other animals, apes in particular, are also capable of learning languages. Experiments show that Bonobos acquire language up to the level of human children. For example, they can understand sentences that contain one verb and a three-noun phrase (“Will you carry the M&Ms to the middle test room?”), but they have trouble with conjoined sentences that require two separate actions (“Bring me the ball and the orange”).?Incidentally, this level of language proficiency is roughly equal to where Siri or Google Talk are up to now.

Also, it is not clear that apes really dig the concept of language to “coordinate behaviors between individuals by a complex process of exchanging behaviors that are punctuated by speech” as defined by noted primatologist Savage-Rumbaugh.?She came to this conclusion as a result of an experiment in which she worked with two apes called Sherman and Austin. The apes had successfully acquired many signs and used them effectively to communicate with humans. There didn’t seem to be anything odd about their language use until one day they were asked to talk to each other. What resulted was a sign-shouting match; neither ape was willing to listen. Yes, I know, the same could be said about many Twitter exchanges, but we are talking apes here.

Apes and other animals may be capable of using language as a means of communicating present experiences and wants, but if language is a model of the mind and the of reality, it is doubtful they are able to deal with abstract thought.

4.??????Language as a Cultural Trait

Chomsky explains why word orderings seem to be equally natural, and they are equally common among the world’s languages. ‘Man spear throw’ (Subject-Object Noun-Verb or SOV) is the basic order in Japanese, Turkish, Korean, Basque and Hindi, to name but a few examples, whereas ‘man throw spear’ (SVO) is the basic order in Chinese, English, Finnish, Swahili and Thai.

Chomsky’s view on the universality of language has come under dispute particularly from anthropologists who like to point out cases where the use of language seems to be unique to a human group or even that language is a driver for human thought and not the reverse.?The idea that language shapes the way people think is one of the most hotly debated issues these days.?They give the example of the Turkish word ?ehirlile?tiremediklerimizdensiniz, which means nothing less than ‘you are one of those whom we can’t turn into a town-dweller’.?How would a genetically Universal Grammar would generate such a monstrosity? Keep in mind this is really one atomic word; not merely many different words squashed together like those in German or like pneumonoultramicroscopicsilicovolcanoconiosis—the longest word in the English language.

Other counterexamples come mostly from the language of tribes such as the Yanomami of South America who are one of the most isolated tribes anywhere in the world. Their language reflects their traditional lifestyle. For instance, their numbering system includes one, two and more than two since they saw no need for naming a number greater than two. They only have names for five colors: white (also means clean), black (also means dark), blue-green, yellow, and red.?The Pirah?, another Amazonian tribe, even lacks any unique?color?terminology, being one of the few cultures (mostly in the Amazon basin and New Guinea) that only have specific words for?light?and?dark.

But before we go into ethno-centric tirades, thinking those people’s languages are that way because of their “primitive” lifestyle, let’s remember that all major languages also have equally nonsensical limitations or constructs.

5.??????Why would Weak AI not be enough for NLU?

As impressive as they are, today's LLMs are still best defined as "Weak AI". Think of Weak AI as something akin to the Wright Brothers inventing a flying machine that more closely resembled a bicycle than the flapping wings of a bird. There are no philosophical conundrums about whether airplanes have the consciousness of birds.

Weak AI is good for things like ChatGPT, or building a chess program able to defeat Garry Kasparov, but the artificial player’s programming used techniques that do not resemble how we humans think (and to this day, Kasparov still claims foul play was involved). So, impressive as that feat was, no one tried to claim Deep Blue was proof AI had finally arrived.

Even increasing the size of transformer networks to trillions of parameters, would not advance true understanding.

The problem with Weak AI is not one of processing or storage capacity. At the dawn of the computer age, visionaries argued that Artificial Intelligence would be possible if only we could only build a computer the size of a ten-story building. Thanks to the explosive advances in computing technology, we can now have a digital wristwatch with more computing power than would have been available in an old-style building-sized computer.?And still no AI. We can safely conclude that what precludes us from creating strong Artificial Intelligence is not a lack of computing resources, but the fact that we simply do not know how we think! Take for instance the Frame problem; given the billions and billions of potentially relevant facts around us, how can humans seem to effectively zero in on the most relevant ones? What we know typically as “Common Sense” turns out to be one of the most complex features of our mind.?No wonder why “common sense” is not so common!

It is conceivable that we will soon have a machine able to effectively simulate “understanding” in a variety of cases, and it can be argued that ChatGPT already passes the Turing test with flying colors by “fooling” a human into believing the machine is human.?Take, for instance, DeepMind. This neural network software achieved the feat of trashing multiple world champions in the game of Go. However, the analysis of its way of playing seems absolutely alien to the way humans think.?

Once we look under the covers, machine learning systems have more in common with the type of ‘brute force’ approach used in chess playing software rather than the structured thought answers originally contemplated as “Strong Artificial Intelligence”.?Weak AI systems can beat humans at many things, but we cannot really say they “understand” anymore we can say that DeepMind understands the game of Go.

6.??????How about just replicating the Human Brain?

The brain is the organ directly involved in monitoring, coordinating, and controlling our body’s movements and functions. Neuroscientist can map specific portions of the brain to motor-control areas, and books from the likes of Oliver Sacks (“The Man Who Mistook His Wife for A Hat”) clearly demonstrate causation in the effects resulting from brain injuries or diseases. It is evident that disrupting the operations of our brains can lead to serious learning, perception, and cognitive malfunctions.?

It is estimated that our brains have about hundred billion neurons. Each neuron can have as many as two hundred thousand dendrites, and on average, one thousand dendrites connect to other neurons. We can estimate that a brain can have on average 100 trillion neural connections. Imagine the complexity this estimated number of connections can generate!

Many believe that anything in our brains could be reproduced computationally. This belief motivates multimillion-dollar projects such as the European “Human Brain” project[3] and the American-sponsored Brain Activity Map Project (BRAIN), attempting to map the neurons in the human brain.

Now, if you recall, it was once believed that once the human genome was successfully sequenced, science would be able to pinpoint the specific workings of the human body. This mechanistic view held that we would be able to map genes to specific proteins and proteins to specific functions. What science has found instead is that most of our genes do not encode for any proteins and that it not certain anyway that our genetic code is what entirely determines who we are. After all, onions have ten times as many genes as humans.?Mapping the neurons in the human brain feels more like humans trying to achieve flight with flapping wing contraptions and has the markings of being a fools’ errand.?

7.??????Why is NLU such a hard problem?

Given that the core mechanics of language is not yet well understood, you have probably figured out how hard would it be to develop an NLU system. Furthermore, as per Wittgenstein and even Chomsky, language is not something that can just stand alone in the guts of a computer. Remember, as per Wittgenstein, language equals thought. And as per Chomsky, language is an evolutionary component.

It's one thing to have a machine be able to pass the Turing test with flying colors by “fooling” a human into believing the machine is human; it’s another thing entirely to develop something as complex and seemingly sentient as Samantha from the movie “Her”.

When it comes to language, it appears that to really have NLU, we must first to crack the “hard problem” of Strong Artificial Intelligence.?

Obviously, the emergence of a Strong AI technology would present significant societal implications. Books like “The Singularity is Near” by Ray Kurzweill[4] have advanced the idea that given the trends in computing power, we will see the event known as “the singularity” whereby computers will exceed the intelligence abilities of humans.

But does that mean that computers will be wise?

8.??????NLU Is a step to the road of wisdom

You meet your best friend in a bar, and he confides to you, “I’m going steady with my girlfriend Laura. I’ll need a ring.” Now, if you are a computer you will struggle to figure out what kind of ring he is referring to: a circus ring? A smoke ring? A bell ring?” The software would need some basic information to reach the right conclusion: the idea that “going steady” means getting married; the knowledge that when someone wants to get married, he needs to propose; the cultural construct that marriage proposal involve the presentation of a . . . ring—a diamond one preferably.

NLU can only be accomplished once information is placed in context and the resultant significance of relationships within the language is realized.

Most traditional approaches to Strong AI have relied on the idea that computers must have access to some form of reality to effectively reach the level of understanding.?There is a sense that language understating relies on a significant amount of implicit information: contextual, environmental, historical and so forth.

Embodying computers with this information require some element of human input. This input would take the role of our own sensory experiences and learnings. By simply living, we learn very quickly that objects do tend to fall, but there is a need to provide computers with this type of information, if they are to support understanding. If there ever is going to be Natural Language Understanding, this will entail moving from big data to better content by structuring and classification of this data, to a way of providing computers with the knowledge of the world required for understanding. This entails ways to encode information using taxonomies and ontologies.

9.??????On Content: Taxonomies

A Taxonomy is the categorization or classification?of entities within a domain (the actual structure of the domain is defined by its Ontology).?Consider the following taxonomies used to describe the animal kingdom.

Linnaeus Taxonomy:

·????????Kingdom: Animals, Plants, Single Cells, etc.

·????????Phylum: For Animals: Chordatas, Nematoda (worms), etc.

·????????Class: mammals, amphibians, aves. . .

·????????. . . et cetera

In "The Analytical Language of John Wilkins,“ Jorge Luis Borges, the famed Argentinean writer who belongs to the taxonomical set of writers who deserved to win the Nobel Prize but didn’t,?describes 'a certain Chinese Encyclopedia,' the Celestial Emporium of Benevolent Knowledge, in which he lists this very unique taxonomy for animals, classifying them as follows:

·????????those that belong to the Emperor

·????????embalmed ones

·????????those that are trained

·????????suckling pigs

·????????mermaids

·????????fabulous ones

·????????stray dogs

·????????those included in the present classification

·????????those that tremble as if they were mad

·????????innumerable ones

·????????those drawn with a very fine camelhair brush

·????????those that have just broken a flower vase

·????????those that from a long way off look like flies

·????????others

As you can see, there are many ways to define taxonomies.?A taxonomy is the magic ingredient that turns raw data into content.?With the advent of the Web, we have moved from the Age of Data to the Age of Content.?“Content is King” is more than a cliché. Content is the product that breeds the fortunes of companies such as Facebook and Google.

10.??On Knowledge: Ontologies

While Taxonomies may advance the creation of content via data structures and metadata, Ontologies begin the process of imparting something akin to knowledge to computer systems.

To see the difference between Content and Knowledge, try this exercise: Go to google.com and enter “IBM Apple”. You will get content listing all the sites in which IBM and Apple are discussed. Now, go to wolframalpha.com and enter, “IBM Apple”. You will get a digested and structured response comparing these two companies. The former is content; the latter is beginning to look a lot like knowledge.

An Ontology is a type of taxonomy that includes operational relationships and aims to describe a domain of knowledge in which its elements interrelate in a way that supports inferences. Several ontologies have been created, some in open source format that cover multiple domains. The Worldwide Web Consortium (WWC) for example has published a Resource Description Framework (RDF) and the Web Ontology Language (WOL). There are ontologies for biology, pharmaceutical engineering, and many other fields. There is in fact an International Association for Ontologies and Its Applications ( www.iaoa.org).

There is also a body of work started over thirty years ago by AI pioneer, Dough Lenat, to create a codification of all human language around structured frames for the purpose of providing semantic-level knowledge[5].?Following is an example of the type of entry you will find in the Open Source version:

The latest count on Project Cyc is 1,500,000 terms and 42,500 predicates representing a complex ontology of relations, attributes, fields, properties, and functions. Typical pieces of knowledge represented in the Cyc knowledge base are "Every tree is a plant" and "Plants die eventually." When asked whether trees die, the inference engine can draw the obvious conclusion and answer the question correctly.

Eventually, these ontologies will have to reach individual-level precision if we are to have a machine that can naturally interact with a user.

Take sarcasm for instance. Sarcasm is viewed as one of the more difficult AI problems to overcome because, to properly ‘get’ someone’s sarcasm, intimate knowledge of that person’s belief system is required. Sarcasm is so complex that even we humans often fail to recognize it. Sheldon Cooper, the main character in the Big Bang Theory comedy show, suffers from a condition that makes him tone deaf to sarcasm[6]. In the end, determining sarcasm requires a combination of knowledge about the person speaking it, the context of the situation, and often times, interpreting gestures and body language. This is why sarcasm does not translate well in Tweets or emails. To this end, ontologies that map an individual’s belief system will be needed.

11.??On Understanding

Production and discovery of knowledge is at the core of many start-ups’ business plans today. The emerging field of Data Science is leveraging big data sets to mine data in ways that produce knowledge. Organizations, such as Gallup or Nate Silver’s FiveThirtyEight, exist to mine data and content and produce knowledge on a variety of topics. Voting trends, consumer preferences, etc. are examples of mined knowledge. Business Intelligence, associated Data Mining technologies, and the more recent Internet-driven “Collective Intelligence” applications are examples of the more recent trends in the automation of knowledge acquisition.

Does all this mean that if we apply something like the Cyc frame-based ontology to a machine capable of handling complex patterns vis machine learning algorithms, we might finally reach the point of understanding?

Not quite.

Understanding is a step above Knowledge. Also, understanding may be subject to biases and the kind of errors we humans are so adept at having.?For example, understanding entails ambiguity resolution and being prone to misunderstanding. I remind you of this famous Groucho Marx joke: “One morning I shot an elephant in my pajamas. How he got in my pajamas, I don't know.”

Understanding is interpreting the significance of relationships between two or more sets of knowledge and deriving prime causes and effects from these relationships. While Gallup may unearth the knowledge that 33% of voters are likely to vote for a given candidate, understanding why they lean that way is something that information systems can only hint at.?Understanding remains an endeavor only humans are adept at.?Understanding cannot yet be performed by computers. As much as it might appear to be the case, the Siri and Google Talk systems understand your commands no more than your dog Fido understands that you are asking for the New York Times when ordered to fetch the paper[7].?

12.??On Wisdom

Wisdom is the ability to choose between correct and faulty understanding. The fact is, understanding can be the result of wrongly extracted knowledge, which may come from bad source data (outright misinformation), or content improperly formed with inappropriate taxonomies. For example, the taxonomy that classifies human beings according to race or some other categorization of “otherness” often leads to xenophobia, homophobia or racism.?Alas, the same issue occurs in systems not based on taxonomies but instead upon supervised machine learning where we have seen cases of neural systems that were trained with biased sets of data that fed the generated pattern recognition with racist classifications.

Wisdom represents the highest level of value in the information progression.?Wisdom is not always objective or static. It can often be subjective, and it is certainly dependent on the cultural environment or transitory circumstances. It is unlikely that we will ever be able to codify “hard-coded” wisdom within computers. The belief that these future computers may act as judges in the affairs of men is dubious at best.

After all, wisdom entails access to attributes such as empathy, sense of justice and fairness, and so forth.

Beyond wisdom lies the fuzzy realm of morality and philosophy; of self-awareness, or consciousness.

And debating the issue of consciousness is something that can also quickly can take us into the rabbit hole of philosophical and religious debates with profound existential implications.

13.??On Consciousness

In the end, it is perhaps hard to conceive a system that will have automated wisdom without first solving the question of “consciousness.”?Are we humans the mere biological machines some neuroscientists claim we are, endowed with the illusion of ‘consciousness’ that is a by-product of our brain, or is there something more transcendent going on???

There is now an association for the scientific study of consciousness (ASSC)[8] which defines consciousness as the experiential world of sensations and thoughts you wake up to in the morning, and that disappears when you fall into a deep dreamless sleep.?That a group of experts coming from the fields of cognitive science, medicine, neuroscience, and philosophy could only come up with this lame definition is kind of disheartening.?This definition does not exclude the experiential world of sensations and thoughts that lower animals might experience while awake. Whether all or some or none of the animals are endowed with consciousness is unknown. I’m not certain what a correct definition of consciousness might be, but I am almost certain that the words self-awareness would have to appear in such a definition.

When it comes to consciousness, it is fair to say it remains one of the great unsolved puzzles of science today.?Typically, the two main theories of consciousness have been the materialistic “consciousness is strictly the juice the brain excretes,” and the dualist, which proposes there is a separate conscious soul from the body. One need not be religiously minded for being skeptical of simply stated materialistic explanations.?New explanations such as Pansychism and Biocentrism posit that it is not the universe that creates consciousness, but the other way around: consciousness creates the universe. Despite sounding extremely new-ageist, these theories come from reputable scientists and are in fact founded on quantum-level explanations for the phenomena, such as the Penrose-Hameroff model of consciousness[9] . And why not? Once you enter the realm of quantum physics, we must deal with a host of other weird phenomena. If this is the case, it might well be that our brand of intelligence can never be recreated by traditional Turing machines.

Author, Rodney Brooks, said it best:

“To me it is perfectly obvious that consciousness consists of more than electric or electro-chemical signals, as in a computer or robot. Why do I believe this? For the same reason I believe that it is impossible to make a television set out of wood. If I took the most skilled carpenters in the world, gave them an unlimited supply of wood and said, “Take this wood and make a television set, but don’t use anything except wood”, I know they couldn’t do it. Wood doesn’t have within itself the capability to do the things that a TV set does. Similarly, electrical signals and computer memories don’t have within themselves the capability to experience the color blue or the sensation of pain. We can’t even define these sensations, much less know how to create them from computer parts.”

But let’s come back to the ground, shall we? Even as philosophers debate the true meaning of life and the possibility that artificial consciousness might exist, there is no reason we cannot begin to develop a strategy around Natural Language Processing, Robotic Automation, and Machine Learning, all elements of what can be referred as Cognitive Automation. After all, there are already plenty of exciting developments that can be made to work to the benefit of business.?Computers now appear to have true intelligence traits, especially when used in narrow-domain applications. Whether Strong AI is feasible or whether the Singularity will occur are problems best left for the next generation.

Even as philosophers continue to debate the true meaning of life and the possibility that artificial consciousness can ever exist, there is no reason not to begin to develop a strategy around practical applications.

If Natural Language Understanding is not here today, the same cannot be said about Natural Language Processing. NLP is very real.

Next Baker’s Dozen topic: On Natural Language Processing (NLP)


FOOTNOTES:

[1] https://www.propublica.org/article/google-says-google-translate-cant-replace-human-translators-immigration-officials-have-used-it-to-vet-refugees?fbclid=IwAR21pvcLitqhVwEU99j7NOfCh1naZZV-ZpzOoFKte9ScYWzf-JNe1hRWQQQ

[2] Kenneally, Christine (2007-07-19). The First Word: The Search for the Origins of Language (p. 3). Penguin Group. Kindle Edition.

[3] See this link for a status on this project: https://www.bbc.com/news/science-environment-28193790

[4] “The Singularity Is Near: When Humans Transcend Biology”?by Ray Kurzweil

[5] https://www.cyc.com

[6] The show producers swear that they did not intend to convey the character suffers from Asperger, even though they admit the symptoms might be similar.

[7] Though experiments by note Primatologist Sue Savage-Rumbaugh suggest that trained Bonobos can understand sentences with one verb and up to three nouns, and even to understand sentences they have never heard before; skills equivalent to the competency of a four-year old human child.

[8] https://www.theassc.org

[9] https://www.quantumconsciousness.org/penrose-hameroff/quantumcomputation.html



Mohamed Muman

Independent Data Management Professional and Solution Architect (on a long sabbatical)

4 年

Hi Israel - well thought out and well written - always knew your brain was much bigger than mine!!

要查看或添加评论,请登录

Israel del Rio的更多文章

社区洞察

其他会员也浏览了