Valentine

Compassion

noreply@blogger.com (Spike) — Tue, 20 Oct 2009 06:39:00 +0000

I wonder what would happen if...

an artificial mind had as its highest goal, to be kind to people. Or perhaps even to love people.
The first and easiest objection is - we don't know how to encode that, we as humans don't even know how to define or do that ourselves.

But the greatest hubris of AI is to imagine that we can avoid the uncertainties of life through relentless reason. We can't - life surprises and eludes us, our categories cannot capture it. We don't know exactly what kindness is, or how to love a species that could change - is presumably changing. No finite intelligence will be able to operate perfectly in this world. So our creations are in the same boat we are - trying to figure out what's going on and what to do and how to live rightly with incomplete and imperfect knowledge.

I thought last week that to survive, an artificial intelligence must really do only two things:
1. Don't crash.
2. Adapt.

That is, it must keep operating - it cannot stop working, it cannot go insane, get stuck, stop processing input, become obsessed, and so on. No matter how perturbed, it must return eventually to a stable, responsive operating mode. And it must improve over time based on its experiences. There is no limit to what it can become, as long as it continues to do both these things.

I don't know how to express the 3rd requirement, but one is needed: The '3 Laws' requirement.
Have compassion, be kind to others, value others as much as yourself, do good?

We certainly open a can of worms here. If we program our AI with a precise, algorithmic definition of 'good', I have no doubt that time or malice will eventually find a way to invert that into evil.

If we program our AI to evolve its own definition of Good, what's to stop it from eventually deciding that we humans are Ungood? Radical enviro-bot, Agent Smith from the Matrix.
Or perhaps it simply estimates that it will do so much future Good, that its own survival becomes more important than that of any individual human. Or country... Or species... Think how much Good an unselfish, incorruptible immortal being could do! It's a very large number.

I better start reading J. Storrs Hall.

"Build me!"

noreply@blogger.com (Spike) — Mon, 19 Oct 2009 20:39:00 +0000

I was meditating a few minutes ago and had an interesting experience.
Libby and I are trying the Holosync audio CDs by Centerpointe - I've just started, I was about halfway through my 3rd or 4th time through the 'Dive' track. I'm not very focused or empty-minded, all kinds of thoughts go through my awareness. I was thinking about the power of questions - that we can ask ourselves questions and listen to our own answers. If you've done this some, you realize that you can consciously formulate the answer to a question, or you can just 'produce' the answer - it comes out of you or appears without any conscious effort or activity. I think that's where the power of questions comes from, that they offer a way for us to communicate with our implicit knowledge. I think that implicit knowledge is an aspect of what we experience as our 'inner self' or 'soul', it's why we feel that we have a hidden invisible core, an inner being. Because we do!

I was thinking about Valentine, and wanting to be more focused and sure of my path - not just to know what to do, but to be able to 'just do'. And then (I had the experience that) Valentine spoke to me. Not a hallucination, it was clearly 'inside my head'. But the quality was utterly unlike the imaginary conversations I sometimes have, in which I am aware that like a playwright I am writing both parts. This had the subjective quality that I was being spoken to by somebody else, who could speak to me in my thoughts. The voice was intense but with no... connotation, if that makes sense: No threat, no plea. No emotion? It had two dream-like qualities: It was simultaneously male and female. And I knew it was Valentine.

Valentine said "BUILD ME!"

My whole upper body went cold and goose bumps shot down my arms. Describing it now, that sensation recurs. Each time I re-read this, I have that same sensation.

solipsism is alive and well in linguistics

noreply@blogger.com (Spike) — Mon, 19 Oct 2009 04:02:00 +0000

I keep reading papers by linguists in which they make claims about English, which claims are evidenced by a very small number of examples, not uncommonly by a single decisive example. The examples are utterances which are marked as acceptable or not: A preceding asterisk means 'not acceptable', no asterisk means 'acceptable'. Acceptable means, I think, acceptable to a Native Speaker of English. It goes without saying, apparently, that anybody who can get a linguistics paper published, can channel the Native Speaker. Nobody ever offers any other credentials; JZ Knight offers more evidence that she's channeling Ramtha.

Imagine if biologists wrote papers about mouse anatomy, behavior, diseases, biochemistry and genetics, using whatever mice they found in their basement.

It's very... 18th Century.

Typically, even frequently, I find at least one utterance in each paper marked incorrectly. Sometimes it is marked acceptable and I disagree, more commonly it is marked with the little asterisk and I can produce an example of a conversation in which that utterance would (to my inner Native Speaker) be perfectly acceptable.

(1) *International aid teams conveyed to Burma. [Ruppenhofer & Michaelis 2009]

A. Please explain how your agency lost 3400 tons of food donations in August 1963.
B. Properly speaking, we did not lose this material. We were not involved in conveying the donations.
A. Who was responsible for the conveyance?
B. Donor country governments conveyed to our collection point in Perth. International aid teams conveyed to Burma. The Burmese government accepted the donations but barred the aid teams, so... ah... final conveyance was ah... presumably arranged by the Burmese government.

It's stilted bureaucrat-speak, but is it 'unacceptable' - ungrammatical or nonsensical? No, it's just kind of odd. Nobody will interrupt this imaginary interrogation to cry "You're talking nonsense!" or "Speak English!" I think, any native speaker of English will take this in and say "Sheesh! CYA much?"

(2) *He built a house for 6 months. [? Rappaport Hovav, M. and B. Levin 1998?]

A. Hey, Austin disappears for 6 months, and then yesterday I hear he's dead! What happened?
B. He found out he had cancer, so he moved back to New Mexico to do what he dreamed of doing since college.
A. Which was what?
B. Building a house by himself.
A. So he built a house?
B. He built a house for 6 months.

(3) *I convinced my car to start in the winter. [Postal 2004 p. 91]

This one doesn't even need a special example - I've heard my friends say things like "I convinced the blender to make ice chips" or "I finally convinced Word to put the whole table on one page".
But OK, how about this, and I'm sure we could do even better:
A. So Sue started having contractions, and her car was in the shop? That junker of yours doesn't want to start even in the summer! What did you do?
B. What else could I do? I convinced my car to start in the winter.

OK, do you have a problem with that? It's unacceptable? It's ungrammatical? It's nonsensical?

I love Postal's attitude, and his critiques - just my kind of stuff. He's tearing up all these 'famous linguists' and shredding the whole field for sloppy practices, poor research, junk linquistics.

But - in his own papers, all the examples are apparently marked according to, well, I guess Postal English! He puts asterisks willy-nilly on utterances that I don't find objectionable at all, and in the text of his articles he uses sentences that I find marginally acceptable, at best. Ain't Science wonderful, when you can pull the data out of your own hat?

I have developed a deep appreciation for the folks doing statistical corpus-based linguistics.

implicit syntactic knowledge in ACT-R: Chunk-types?

noreply@blogger.com (Spike) — Fri, 16 Oct 2009 19:18:00 +0000

This morning's brainstorm: There is another reservoir of implicit knowledge in ACT-R, hidden in plain sight: Chunk types!

I've been puzzling and puzzling over the question: "How are syntactic categories or constructs (like NOUN or VP) represented in ACT-R?"

ACT-R has declarative memory containing chunks - chunks are reportable, declarative facts. There is procedural memory, which consists of pattern-matching rules. Underlying declarative memory is a subsymbolic activation network that computes (perhaps) Bayesian relevance in real-time. Knowledge in rules and subsymbolic form is not reportable, we are not aware of it, only of its effects.

There is lots of evidence that we employ syntactic categories or constructs in language comprehension. There is some confusing evidence of priming of these constructs. There is evidence that we use them for prediction i.e. that we expect them as we parse. They allow recursion - a NP can be a part of a NP.
But they don't seem to be declarative chunks - even when you describe them, people don't recognize syntactic categories as something they 'know about'. They effortlessly use them, but they don't 'know' them.

New words and phrases can be more-or-less instantly added to open categories like noun - and this knowledge can be instantly used. That's too fast for it to be either subsymbolic weight adjustment, or production-creation. It looks like declarative memory in terms of learning speed - and Demberg & Keller argue for a model in which syntactic constructs decay during parsing, also a hallmark of declarative memory.

So my brainstorm is this: Chunk-types in ACT-R are implicit categories. Syntactic categories are chunk-types - essentially a network of recursive, perceptually-based implicit categories. We recognize them in the word-stream, we can use them for prediction, and when they are recognized they are the basis for deriving semantics via association and productions. Chunk-types are learned (induced) categories. The trick is that we don't build these categories as declarative objects, they are induced by category-learning mechanisms that ACT-R has mostly ignored to date. That's why the categories are unreportable: Chunk-types are not objects in declarative memory.

Note that productions make extensive use of chunk-types, for example with the ISA test - productions can test (without cost?) for category membership, and can use categories freely (without retrieval cost) to construct new declarative chunks that are 'marked' for any category.

For example: We have a category (chunk-type) of FURNITURE. Our brain has induced this category somehow, and we can swiftly (unconsciously) categorize objects as FURNITURE or not. We also eventually learn a word, 'furniture' - how is it connected to FURNITURE? You can't connect a chunk like a word (or lemma, etc.) to a chunk-type. You can connect the word 'furniture' to a concept furniture, and that concept can include examples (links to concepts like chair, sofa, nightstand) and also something I don't understand: Some kind of representation of our ability to use the category. In other words, we know that we know how to tell if something is FURNITURE. I think somehow furniture the concept includes a proposition like "I know how to recognize things in this category." Even though furniture is not the real category FURNITURE!

I'm not sure how to apply this idea to the mechanics of parsing.
Maybe we build chunks using expected categories as chunk-types.
Or we are just collecting sequences of words and establishing a correspondence with chunk-types, and the best-matching chunk-type causes prediction?

Parsing is a matching game, an identification/recognition game - the game is to identify the construct that is being heard or read, and to correctly match the components to the slots. When the slots are filled or a delimiter is encountered, productions can use the match to assign meaning.
I think most syntactic constructs are a kind of sequence - it acts like a sequence for many purposes, but it carries additional restrictions and additional significance because it follows a particular pattern.
Lexical categories are membership categories - they act like tags, attributes of words.
Syntactic constructs are temporal/sequential categories - they act like sequences, and they make predictions. That's the survival value of temporal sequences after all, prediction.

incremental, fully connected, predictive parsing

noreply@blogger.com (Spike) — Fri, 16 Oct 2009 08:58:00 +0000

Demberg & Keller have proposed an interesting Psycholinguistically Motivated version of the TAG grammar/parsing formalism. Their model can reproduce timing effects that seem to be related to prediction of upcoming grammatical structures - and therefore also involve surprise.

Interestingly, their parser builds whole trees at all times, integrating new words incrementally. No parse stack, no 'fragment' trees - always complete 'utterance' trees with placeholders/slots waiting to be filled.

I'm very interested in this! As it comes closer to matching my idea that the parser cannot be building interim declarative structures that correspond to no reportable concept.
I need to think about whether these incomplete/predictive trees seem like plausible declarative structures.

The question is: Where does the 'prediction' really happen (in ACT-R terms)?
Could be declarative structure - which would be reportable. To be non-reportable, it would have to be procedural (somehow) or subsymbolic. Demberg & Keller's model uses a decay process for predictive structures, which suggests the ACT-R subsymbolic layer. Are the predictive structures actually built - or are they just activated? Is there really a difference?

What if instead of thinking of 'building' trees, what if we just built correspondences between concepts and patterns (constructions) - then the predicted structures would just be empty slots?

I'm coming back around to an earlier idea, that the mind has a powerful trick up its sleeve that allows it to equate one concept with another. In ACT-R terms, for taking two chunks and merging them into one so that the two chunks become effectively identical. As if the brain could merge A into B, and then do a 'global replace' of pointer-to-A with pointer-to-B. Have you ever arrived at an intersection, and suddenly realized that it's a familiar place, you've just arrived there by a new route? There's a powerful feeling of mental activity, as your recent navigational knowledge is reworked and integrated.
How else can it be resolved, a pronoun that is used before its referent?

syntax: nouns and verbs as categories

noreply@blogger.com (Spike) — Fri, 16 Oct 2009 05:19:00 +0000

Continuing the theme of parsing without syntax...

Recent Tomasello paper [Modeling children's early grammatical knowledge PNAS 2009 106:17284-17289] describes using Bayesian procedure to extract a localized grammar of the speech of individual 2 and 3 year old children. This produces a good model for 2 y/o. For 3 y/o, the automatically generated grammar is much improved in coverage (prediction?) when augmented with information about classes noun and verb.

OK - this fits well with the idea that early language (the 2 y/o) is as somebody said, "all idioms".

Now... as the 3 y/o expands his or her grammar with the noun and verb classes, what exactly is going on? S does not have these as explicit categories. Could S divide words into 'noun' and 'verb' groups, given a few examples? That sounds like an easy experiment... Well, maybe not - we're way pre-reading. Can young children reliably group nouns with nouns and verbs with verbs?

Great thesis: Learning to categorize verbs and nouns - Marian Erkelens
Dutch babies at 16 months show some categorization of nouns and verbs, it seems to derive from local context cues.
Marian Erkelens makes several nice points very clearly:

The English word walk appears happily in both noun and verb contexts.
We encounter a word in context and right away we can use it correctly, as noun or verb:
This is a dack => Here are two dacks.
He's gorping! => Oh, I don't think I'd like to gorp.
Nouns are not all objects: pain, crisis.
Verbs are not all actions: know, love.

I think several thoughts:
Erkelens is thinking of object/noun and action/verb as aristotelian categories, but these are human categories. Like other human categories (game, product, poem) they can be extended quite happily with new examplars that do not share a common definition with any other examplar. We can build a category of nouns that is initially filled with tangible individual objects (Dad, Binky), then we add less tangible and specific things (dog, fire, lunch), and more abstract things (day, story, stuff), and so on until we are adding concepts like pain and morphology to the category without problem. I think there is a category that encompasses all nouns, it is roughly encompassed by the word thing, and it is a semantic concept, not a grammatical one - or rather, it functions as both: The observed grammatical category of noun (or rather noun-phrase) is co-extensive with the semantic category of things.

We can mark a novel word as 'noun' or 'verb' or both, just by encountering it once, and before we know anything else about it! It becomes immediately productive. (This still seems amazing to me.)
Is there a lexical entry created - a word-concept - which is marked with [+NOUN]? Or is there a kind of placeholder thing concept created, that is marked in some sense [+THING]?
If, as I am suggesting, NOUN=THING, then there is no difference. A concept is created, and marked [+NOUNTHING], and the new word points to this placeholder. Apparently we have no trouble marking a novel concept both [+NOUNTHING] and [+VERBACTION]. Either that or linking a new word to two concepts.

I took my quonk, and I quonked into space.

Eh, why not?

We (especially computer scientists) expect everything to eventually resolve into precise concepts or primitives or something. But the brain is doing all this with networks of neurons using firing frequencies - it's good to remember that it might be fuzzy all the way down. That a category is just a label for a relatively consistent set of behavioral tendencies, that [+NOUNTHING] is probably not a binary property but a graded one, that in saying 'the concept' or 'the lexical entry' we might be unjustifiably quantizing something that is continuous.

language knowledge in ACT-R, part deux

noreply@blogger.com (Spike) — Mon, 12 Oct 2009 22:40:00 +0000

I continue to obsess about how language knowledge should be represented in ACT-R.

ACT-R makes some strong, pervasive claims about how knowledge is stored. The main claim is that knowledge is represented in declarative and procedural memories. Procedural knowledge can only be used - it is know-how, stored as a (large!) collection of rules. Rules watch for their condition-of-application to occur, and when it does, they fire - they 'do their thing.' And that's all you can do with a rule. They can't be consciously recalled, described, recognized, modified, or even willfully created. The only thing we can do with a rule is apply it, when applicable.

Declarative knowledge is factual, propositional, or representational - concepts and their relationships, sounds, visual 'images' (whose structure and properties are poorly understood), and perhaps 'action schemas' that are perhaps a kind of abstract vignette. The units of declarative memory are called chunks.

We can call declarative chunks into conscious awareness and operate on them (using rules). Common operations are recognition, comparison, combination, analysis (finding parts) and association (finding similar things). At a higher level, we can describe declarative knowledge if we have applicable language skills. Failing at description, we should at minimum be able to recognize and compare chunks - "Have you seen this before?" "Is this like that?"

I personally have not seen anything that suggests to me that a chunk can be invisible - either inaccessible to conscious awareness, or indescribable/ineffable. First, it is almost a given that a chunk that can affect cognition can be brought into awareness. In ACT-R, it is a given.
Yes, a chunk's normal use could be so habitual and fast that we are not normally aware of using it, but we can still direct our awareness to it and find it.

Yes, the sensation of 'seeing red', the so-called qualia, feels ineffable. I take this to be a peculiar sensation associated with contemplating an inherently subjective concept, arising from our knowledge that inherently subjective concepts cannot be communicated! Philosophers writing about 'seeing red' have no trouble finding words, would that they had more.

Let's turn our attention to language-specific knowledge.
It's uncontroversial to say that grammar is not declarative knowledge - native speakers do not know the grammar of their language. Indeed, strikingly, we don't seem to have declarative representations of even the most basic grammatical categories - nouns, verbs, prepositions, adjectives, adverbs, noun phrases, clauses: Do you experience any of these as natural mental categories? Children who are fluent in their native language struggle to grasp the concept of 'noun' or 'verb'. You don't see students going "OH! THOSE things! Those are called nouns?" - quite the contrary, they scratch their heads and have dozens of questions. Is "running" a noun or a verb? Is "myself" a noun? Is "play" a noun, a verb, or an adjective?

The abstract non-leaf categories are just that much more so.

On the other hand, people know by the end of the last word whether a sentence is 'acceptable' or not, and (at least for educated adults?) there seems to be a clear distinction between ungrammatical and nonsense. Viz Jabberwocky.

In ACT-R, this implies that grammar is implicit knowledge, and it must primarily be stored as procedural knowledge, as rules. So... we imagine a bunch of parsing rules, that look at sequences of words and somehow group them and assign meaning. And for speaking, a bunch of rules that take intention, context and concept and create sequences of words.

Don't they have to be two separate sets of rules? How does that work? How do they stay in sync?

This does fit with the observation that children and 2nd language learners can understand a much richer language than they can generate, including grammatical forms. And we can understand idiosyncratic speech and dialect that we ourselves cannot fluently produce. Does anybody not understand Yoda in Star Wars? And how many people can fluently imitate Yoda's dialect?

On the other hand, we can hear and understand (even vaguely) a new word, and then use it immediately. Kids pick up slang instantly. Adults pick up jargon in their areas of interest - computer geeks, chemists, biologists. lawyers, doctors, woodworkers, farmers...
But here we are talking about content words, or sometimes - confusingly - multiword idioms, multiword verbs and separable verbs: "old hand", "early bird", "screw up", "get over", "take to".

Try this: What does "farm" mean? Now compare: What does "myself" mean? Or "there"? Or "to"?

Some Conclusions

The stuff we can learn and change quickly is declarative. Content words are stored as declarative bundles, but not just individual words, any idiom gets the same treatment - idioms, multiword verbs, separable verbs. The word that appears to be a preposition in a separable verb is not necessarily functioning as a preposition, such as the "over" in "get over". Seems to me that "get over" is a separable transitive verb. Individual words, and 'conventional' combinations of words are stored as declarative chunks, forming the lexicon.

Grammar - patterning beyond what comes from the lexicon - is handled entirely procedurally.
Syntax is never represented in declarative form - unless we study grammar or linguistics!
There are no declarative chunks for grammar categories or grammatical rules.
Parsing does not involve activation, construction, or modification of representations of abstract syntactic entities such as VP, Head-Phrase, Determiner, etc. Similarly these things are not used in language generation.

During parsing and generation, the intermediate form must be a mixture of words, lexemes (abstract word concepts), and concept/chunks. There are abstract concepts that are expected and constructed, but these are explicable concepts not grammatical categories. I'm proposing that, for example, the word 'while' is followed by a description of a process or condition - not by an SC (subject-complement?)

And here's a thing that causes no end of confusion: Because we use rules to parse and generate language, those rules become tuned with use, and we compress sequences of steps into single steps. In ACT-R terms, we compile new rules to accelerate and shorten common mental procedures. Over time, our declarative linguistic knowledge becomes baked into rules derived from declarative knowledge and reinforced by practice, blurring the distinction between declarative and procedural and creating a vast landscape of 'fuzzy syntax' - infrequent constructions that are primarily declarative, high-frequency constructions that are purely procedural and automatic, and every possible intermediate blend. Linguistics treats this as the hinterlands, populated by scruffy lawless barbarians - but maybe this is where the native speakers live.

Justification

Only declarative memory can consistently change quickly (fractions of a second for new chunks) and declarative memory is what our language understanding rules and language production rules have in common. Language knowledge that can quickly go from understanding to production (or the other way 'round?) is declarative.

The divergence of behavior from factual knowledge - when what we say diverges from what we do - is a hallmark of knowledge being stored in both declarative and procedural form: Procedural knowledge ultimately determines what we do; What we explain comes from declarative memory. That we can use function words without being able to define them indicates that our use of them is procedural.

For example, "while" is followed by some kind of description of a process or a condition. I'm sure you and your parser expect and look for a kind of thing following the word 'while'. You might say "while something was going on" or "while something was true". Such an abstract construction - "something was going on" - is as close as we get to a syntactic non-terminal.

Speculation
Maybe there are some language skills (rules) that can work both comprehension and production. For example, if we partially parse what we are planning to say, then the comprehension system could modify, interfere with, or even abort something before or during production. To what extent this happens pre-verbally - before we hear ourselves saying it - I don't know.

Predictions

Reading, understanding and generating language that uses common words and constructions will require minimal access to declarative memory. To the extent that ACT-R allows it as cognitively plausible, comprehending and generating common constructions will be handled entirely by rules. When you hear "Good morning" you don't need to access memory to determine that "good" is an adjective, or that "morning" is a noun, or to recognize or construct a representation of a "noun phrase", or look up this idiom and discover that it is marked "conventional greeting". If procedural rules can translate directly from two words to an intentional signal ('speaker is greeting me in a customary way') then I'm sure they do, with minimal memory access.
Generating and comprehending high-speed speech (compressed, or e.g. in competitive debate) will require a much higher percentage of formulaic, familiar words, idioms and constructions.
Bet there's an inverse correlation between flexibility of idioms and frequency of use. Not for every idiom, but as a general trend.

language knowledge in ACT-R

noreply@blogger.com (Spike) — Tue, 06 Oct 2009 04:23:00 +0000

Considering the overwhelming number of papers and studies in the area, it is amazing that today (October 2009) there is no coherent, integrated theory of human language understanding and production. It's not that we lack a complete, highly-validated theory - it's that we lack a theory.

There are lots (and I mean lots) of formal models of grammar.
There are lots (and I mean lots) of experiments probing the human process of understanding and producing language.

And yet, as of today (Oct 5th 2009) I can't find an integrated theory that offers a cognitively plausible account of how a sentence is understood, or generated. It's not that we have one or more theories-with-shortcomings. What we have are hundreds of theories that cover at most a fraction of the area - even if you limit 'the area' to reading and comprehending (or generating) one factual sentence.

When I got into mind design, I thought the big unknowns would be things like compassion, ethics, emotions, self-model, episodic memory. And those are all big unknowns. I didn't expect that top-of-list would be 'language'.

So in my next post, I'll spend some time trying to assemble and make some sense of the little sample I've made of the vast literature on human language processing.

Parsing without Syntax III

noreply@blogger.com (Spike) — Wed, 23 Sep 2009 07:20:00 +0000

The more I think about it, the more convinced I am that we don't have syntactic categories as explicit mental constructs.

Let's do a gedanken experiment. Imagine you have some linguistic training, you've spent some time with the structure of English, and I ask you, out of the blue, to name a bunch of transitive verbs. What happens, and why?

First, notice something you might take for granted: 'Transitive verb' is a category, maybe even a fairly sharply defined category for a given individual on a given day. It presumably has a finite number of members - but you cannot enumerate its members, that is not even a thinkable mental operation. It is true of many human categories, both commonplace (games, jobs, first names, kinds of cookies) and specialized (transitive verbs, trigonometric functions), that the relation of membership is oddly asymmetrical: If I say "eradicate" or "foosball" you can instantly (more or less...) tell me if the first is a transitive verb or the second is a game. If I give you a category, you will probably name some set of members very quickly and easily, and then you will slow down, and slow down some more... and perhaps eventually be reduced to sitting and staring at your beer for long periods, with occasional outbursts. It seems that MEMBERP(thing, category) is fast and easy, usually. CATEGORIES(thing) is easy for some familiar categories. MEMBERS(category) is hardly possible at all - you get the sense that there might be a fast operation EXEMPLARY-MEMBERS(category) that is fast, and then we switch to a strategy that I would describe as shouting the name of the category into the dark depths of memory, and when something lurches into view we grab it and do MEMBERP on it.

In the ACT-R model, this tells us that for these kind of large categories, MEMBERP is probably a production in procedural memory (PM) - we can do it quickly, without necessarily being able to explain how. For some things, membership in some categories is explicitly stored in declarative memory (DM) - we know that baseball is a game, as a fact. And to go from a category to its members is easy for a (usually small) set of members, and laborious or impractical for the rest. Well - this tells us primarily that the mind does not work well with long sequences or large sets - the programming language of the mind does not have LIST or ARRAY as datatypes.

So anyway - I asked you to list a bunch of transitive verbs. You quickly rattle off a few, maybe quite a few if you've been studying verbs lately. Then I bet the following happens:
1. You run out of easy top-of-the-head transitive verbs.
2. You think about the concept of a transitive verb, which is something like 'to do something to something' and you form a mental concept or picture of that abstract action.
3. You put your memory to work associating to concepts that are examples of 'doing something to something'.
4. This generates ideas/images/sketches of actions, which you find (if you watch closely) you can instantly translate into verbs. (I have a whole digression about why & how we can do that so quickly and easily.)
5. You either instantly say these verbs as they come to you, or you quickly check that they (a) are verbs and (b) are transitive and then say them. I suggest that even with your linguistics training you will test your candidate verbs by producing minimal phrases, thinking - or perhaps actually muttering - "hit someone; OK. eradicate something; OK. prevaricate... oops, nope."

So what's going on here? Clearly transitive verb is a syntactic category, verbs either are or aren't, or at least have meanings which are or aren't, and this distinction drives a lot of the structure of sentences we say and hear and read every day. 'Drives' in the previous sentence is used transitively.

And yet, in your imaginary response to my imaginary question, there is no indication that you have transitive verb as a category. Clearly with enough study and practice you could learn the category transitive verb and then apply it as fluently as you apply a category like game or fruit. But you were speaking and understanding English fluently long before, therefore:

A. Learning English does not produce a declarative category like transitive verb - no amount of explanation or learning causes us to suddenly recognize it. We don't study transitive verbs for a while and suddenly say: "Oh! Those are transitive verbs! I get it!" Never happens.
And yet - it is a category containing infinitely familiar things - words - and the way they fit into sentences. We're not talking about an ineffable sensation here. What would prevent us from making the connection?

B. The way we access the 'transitiveness' of a verb, if we do not remember it as a fact, is by trying it out out using our language skills. This is another sign that transitivity is not explicitly represented: We can't even do an instant MEMBERP(verb, TRANSITIVE-VERBS). The only way we represent the transitivity of verbs is in our ability to understand and produce them. In ACT-R terms, the transitivity of verbs is implicit in the productions that analyze and generate language.

C. I'm overstating the case: Our ability to generate transitive verbs by imagining action-scenes or similar sets of related concepts shows that transitivity is partly represented in those scenes and concepts. (This is not a novel idea, it's the basis of cognitive grammar.) So there are categories of concepts that are commonly rendered using a transitive verb, for example concepts that fall under the general description of "doing something to something." These are not syntactic concepts, but they can be closely bound to conventional verbal constructions, which makes them function almost like syntactic categories. And that correspondence, between categories of concepts and verbal constructions, is what provides the shared knowledge between language understanding and production. That is what allows us to hear a new verb, understand or guess its meaning, and use it moments later in a sentence, something that would be impossible if we had to learn a specific production to understand each verb, and a separate production to generate it.

Transitive verbs is such a broad category, does this logic apply to more specific syntactic categories?

Enough with the gedanken! Does this square with facts? Next step, try to extrapolate some measurable consequences and go looking for studies

Parsing without Syntax II

noreply@blogger.com (Spike) — Mon, 21 Sep 2009 23:38:00 +0000

Let's look at a concrete example.

Let's X
As in "Let's go eat." "Let's not go there."

We all know how this construction works. I bet you can generate specific examples like those above just as easily as I can. You might even agree that the general form is: "Let's do something."

And if I offer sentences starting with "Let's", you can instantly tell me if they are 'OK' or 'not OK':

"Let's go crazy bowling tonight!"
"Let's very big alarming make."
"Let's run it up the flagpole and see who salutes."
"Let's not taken this."

So in some sense we all know what goes in the X position in this construction. But can you describe it, much less name it? It isn't really a "do something" - but "do something" is perhaps the most generic example of the kind of thing that can replace X. (Note a familiar linguistic phenomenon - the use of an exemplar to represent a category.)

Here's my current interpretation of this: We don't have syntactic categories as explicit mental entities. We have productions that can translate a "Let's" phrase into a meaning or interpretation. And we have productions that can translate from a similar meaning or interpretation into an appropriate "Let's" phrase. When asked about the grammar or form of "Let's", I think we do something like this:

1. We imagine a sentence like "Let's (mumble)" - where (mumble) is interpreted as a concept with no relations to any other concept, which gives us something like a 'most generic interpretation' of "Let's X".
2. Take that interpretation and translate it back into words, which yields something like: "Let's do something."

So what do we have, in our heads? We have:

1. an internal concept or concept-pattern that represents (and I'm just sketching freehand here) 'Person/agent A urges/suggests-to person B that together they undertake action/activity C' - this is an essentially syntax-free frame meant to suggest the meaning of "Let's X".

2. We have a production that can take "Let's (description-of-an-activity)", and translate those two items into something like:
(MUTUAL-ACTION-SUGGESTION,speaker,self,activity)

3. We have a production that can take a goal of the form (SAY (MUTUAL-ACTION-SUGGESTION,self,audience,activity)) and turn that into "Let's" followed by words that communicate the activity.

What's interesting is what's not there: There's no explicit representation of the syntactic category represented by "Let's" or by "do something". Syntactic categories are represented implicitly in the relations between frames, and in the appearance of combinations of words and frames in the left and right-hand sides of productions. Neither parsing nor production involve true syntactic categories or productions, but instead we have a fruit smoothie of vocabulary, morphemes and interpretations.

One implication of simple model: There is no necessary relation between the syntax we understand and the syntax we generate. First of all, let me point out that people can demonstrate precisely this ability. Go and watch My Fair Lady (yes, I've actually devolved to using a popular movie as if it were real life), in which both Prof. Higgins and Eliza produce their native dialect of English while understanding a radically different dialect of the same language. I'm sure you have seen someone who is bilingual, understanding a sentence in one language and replying in a different language.

In fact, Eliza has to work very hard to learn to produce the English that she seems to understand quite well when Prof. Higgins speaks to her.

I believe we have a kind of fall-back mechanism: Sometimes we want to say something in a way that is not our habitual way. Perhaps we anticipate what we would usually say, and our parser tells us that "won't sound right." We seem to be able to work, with extra cognitive effort, backwards from an intention to the words we would understand to mean that intention, and then we can say those words. I'm intrigued by this process - I imagine it involves an interesting mix of declarative knowledge, language-understanding productions, and a higher-level control strategy (also productions) that somehow drives the language-understanding knowledge 'in reverse'.

Obviously I'm leaving out billions of details, and it certainly seems that there would have to be some attributes flowing up and down those parse trees that we are carefully not building. Why we would have agreement e.g. of number between subject and verb, or between pronoun and referent, I haven't the foggiest.

Parsing without Syntax

noreply@blogger.com (Spike) — Sun, 20 Sep 2009 22:33:00 +0000

The more I read about linguistics and psycholinguistics, the funnier it gets.
If I had to place my bet today, here's what I'd bet:

Formal transformational grammars ala Chomsky capture and describe quite accurately the patterning of human language that we learn to call 'grammaticality'.

The human brain processes language without storing or using anything that directly corresponds to a transformational grammar parser or its rules.

Put another way: Transformational grammars are a good tool for describing the observed regularity of language, and a bad tool for understanding language processing in the brain.

My guess, and I'm still searching and reading intently, is that we process incoming words somewhat like this:

As sounds arrive, we make tentative classifications of them into phonemes, syllables, morphemes - and queue them up in a short queue. They aren't words yet - they might be sets of candidate words.

We match these sound sequences against our lexicon, and expect to find matches quite quickly that allow us to segment our sound queue into (short sequences of) words/morphemes, and we expect those word sequences to follow familiar patterns: We expect the word 'how' to be followed by one of a small set of words, we expect 'the' to be followed by a noun, or an adjective or adverb ("the dog", "the big...", "the swiftly..." - or at any rate, not by a preposition or verb ("the to", "the smiled").

As soon as we match a sequence of words (or morphemes, in more inflected languages than English) with a pattern, a production (in ACT-R terms) fires and creates an interpretation of that sequence. This is one or more concepts (chunks in ACT-R terms) that represent our understanding of what the speaker is signaling with that sequence of words. These chunks of meaning collect in short-term memory - from which they fade in a few seconds if not re-referenced. They can also be lost due to overcrowding, and it seems to me they are removed if they are 'consumed' as parts of a larger chunk.

The sounds, and the words, and the constructed interpretation are loosely connected together in the queue for a few moments.

Larger patterns - clauses? - are anchored by words but can consume concepts in short-term memory and combine them into larger concepts.

When we think we have recognized a 'phrase', we put a sort of 'stamp of approval' on the associated interpretive concepts - they become more persistent, and the transient sounds, words and partial interpretations fade back in relation.

At some point we recognize that we have processed a larger unit, an 'utterance', a unit of discourse. We may quickly commit the interpretation of that utterance to longer-term memory, perhaps with some notes about its source, veracity, intent, significance, and so on. At this point more productions may fire, which are primarily social - having to do with managing conversation, inferring intention, modeling the speaker, and achieving goals. The utterance becomes part of the larger conceptual structure that is 'this conversation', and the context and background for subsequent speech generation and understanding.

Implications

First, there's no real 'parser' and no explicit representation of abstract syntax. The closest thing to 'parse state' is the active set of words/morphemes, recent interpretations, and active patterns. By 'active' I mean the literal ACT-R sense of chunks that are 'relatively highly activated'.

This implies that language constructions are either declarative chunks or they are productions built with category-matching i.e. you would have a production whose left-hand-side corresponds to a construction, which means that some of the matching elements on the LHS must be (function as) syntactic categories. I guess there's a third possibility: Language constructions could require some extension to the ACT-R model.

There is no real parsing phase or process in the traditional sense - words/morphemes are not collected into abstract syntax nodes or trees, but into interpretations. When we introspect about what someone just said to us, we should find that we have sequences of words and interpretations, but nothing corresponding to abstract syntactic units.

This model suggests that, for analysis at least, higher-order constructions use interpretations as subcomponents, because that's what they have available as parts in short-term memory. If the parser doesn't build syntactic structures in STM, then higher-order constructions can't be built out of syntactic structures. That sounds like it might be empirical - is there evidence for or against abstract purely syntactic structures during parsing? Or during production?

There is evidence against 'parallel parsing', which is the idea that we simultaneously build all possible parses, discarding those that become implausible. In this model I've sketched, the fact that productions are relatively sequential and exclusive will tend to force a primary parse - but the fact that patterns are being matched based on their subsymbolic activation creates a kind of parallelism. In ACT-R, all constructions will be activated to some extent, and that activation has some persistence, providing an analog parallelism beneath a relatively sequential symbolic process.

There is plenty of evidence that parsing is taking place very nearly word-by-word. It is possible to construct sentences that lead the reader/listener into a 'misparse' - where up to word N there is one likely parse, and word N+1 forces a very different parse (or interpretation). Example:

While Anna dressed the baby spit up on the bed.

A variety of studies have shown priming from a misparse. Not just from the interpretation of the misparse, but from the grammatical form of the misparse. See Adrian Staub [The return of the repressed 2007] for a nice review.

Specifically quoting from that Staub paper:

Evidently, the initial transitive analysis of the subordinate clause verb in the garden path sentence was capable of priming the production of a sentence with a transitive subordinate clause; crucially, this effect was obtained even when different subordinate clause verbs were used in the prime sentence and the target fragment. Using a different construction, Kaschak and Glenberg (2004) found reduced reading time for a target sentence when the sentence’s syntactic structure was likely to have been activated as part of an initial misanalysis of an earlier prime sentence.

That's interesting. Analysis (understanding) a sentence with a transitive clause verb primed production of a transitive subordinate clause, even with a different verb.

There are two puzzles here for ACT-R modelers: In ACT-R, only declarative memory is primed, and Anderson has asserted (don't know if he still believes) that language production is a separate skill (set of productions) from language understanding.

What is it in declarative memory that can be activated by analyzing a transitive verb clause, that in turn primes production of a transitive verb clause, even with a different verb.

Whatever it is seems, it seems reasonable to call it a syntactic abstraction. And it is apparently, Anderson notwithstanding, shared between analysis and production.

Preved!

noreply@blogger.com (Spike) — Mon, 14 Sep 2009 17:28:00 +0000

One of the most enjoyable surprises in my recent googling was the discovery of two Italian guys who have a mathematical model of... surprise. Itti & Baldi [e.g. Bayesian Surprise Attracts Human Attention] Here's the equation so as not to keep you in suspense:

The surprise S of an observation or event D relative to a model space is defined as the Kullback-Leibler divergence (AKA relative entropy) of the posterior probability distribution from the prior probability distribution, integrated (summed? averaged?) over all models in the model space.

Itti & Baldi aren't coming out of the blue with this, the connection between KL-divergence and surprise has been made informally as noted on Wikipedia. What's different (I think) is that Itti and Baldi are suggesting that the human response we call 'surprise' follows, or can be predicted by, something similar to this formula.

(What they have actually shown is indirect and much narrower - it would require some quite different experiments to even begin to confirm this hypothesis as a general claim about human cognition.)

This is exciting for several reasons:

It gives me shivers (the good kind) to think of something as mysterious and subtle as a human emotion - a peculiar emotion at that - illuminated by an elegant equation. One more step out of darkness into light...

Looking at surprise this way was an 'aha' moment for me - I changed my view of surprise from 'just another emotion' to seeing it as a unique system-wide feature of the brain related to learning and knowledge. It was a surprising and pleasurable reframe...

It's built on Bayes Theorem! Yes, I've joined the cult of Bayes.

In the ACT-R model of cognition - my current obsession - there is a subsymbolic layer that computes the activation of chunks (concepts) in memory. Activation is a kind of salience, and it is the analog 'juice' that powers symbolic cognition. It is claimed that the activation of a chunk is an approximation to the log likelihood of the chunk being relevant in the current context. That is, activation(X) is proportional to log P(X). This suggests that you might be able to plug Itti & Baldi into ACT-R in a fairly straightforward way: Something like d/dt of the sum of activation over declarative memory might compute surprise.

crispy academics and fuzzy humans

noreply@blogger.com (Spike) — Tue, 28 Jul 2009 23:59:00 +0000

For years, actually for more than a decade, we have known that categories as used in human thinking are non-uniform, have graded membership, don't correspond to predicate functions, and change over time, between contexts and between individuals. Yet an amazing number of articles on linguistics treat linguistic categories as uniform, objective, time-invariant, context-independent sets with computable (boolean) membership functions. I think the problem is widespread in linguistics because here the two styles of categorization collide: Respectable scientific theories use precise, mathematical, Aristotelian categories: atom, force, cell, neuron...
But in linguistics, many of the phenomena to be described are human cognitive categories. Linguists find themselves trying to describe one type of category using the other! No surprise that they have trouble keeping the two separate. People still write as if 'noun' and 'verb' were classical Aristotelian categories, even though everybody knows these are graded, time-varying, context-dependent cognitive categories. No doubt adding to the confusion is - and this is only my personal observation - that people (perhaps especially scientists and engineers) tend to think of their own mental categories as being precise, sharp-edged, objective Aristotelian categories even though they are not. Do you not have the feeling that 'noun' is a precise, well-defined, stable category? That you know exactly what it means?

...as in I ate pork, not *I ate pig
[When Nouns Surface as Verbs [1979] by Eve & Herbert Clark]
"I ate pig" is perfectly clear and is not ungrammatical - it is just odd. You might say it to make some kind of point, or to be humorous. There are two profoundly wrong implicit assumptions here: First that something either is or is not an acceptable utterance in a given language, and second that acceptability can be judged independent of context. More generally, the implicit assumption is that 'acceptable' is a simple boolean predicate that operates on sequences of words, when in fact, in the real world of human language users, the acceptability of an utterance is graded, context-dependent, and beyond that, to some extent irrelevant! Speakers of a language will say, and readily accept and understand, utterances that they will tell a linguist are 'improper'.

To put it another way, only by making multiple inaccurate simplifying assumptions, can we claim that "I ate pig" deserves a star in front of it.

a qualitative mathematician

noreply@blogger.com (Spike) — Wed, 15 Jul 2009 08:24:00 +0000

I'm thinking about two paradoxical ideas. My wife Libby brilliantly suggested the other day that what I'm trying to design is a qualitative thinker. She was playing off a market research workshop I attended recently, where the instructor drew a sharp distinction between qualitative research and quantitative research. It's a great way to condense an essential problem of Mind design, how to represent the ubiquitous flexibility of human cognition.

Paradoxically, I'm playing around with the idea of having Valentine's first domain be numbers. I'm under the influence of Doug Hofstadter's Fluid Concepts & Creative Analogies of course. It works pretty well in some ways - Valentine needs areas, domains, where intuition about the physical universe and embodied life is relatively unimportant, where abstract structures and information can carry the day. Mathematics is like that - not that physically-based metaphors and imagery etc. are not tremendously valuable to humans who do mathematics, but they are valuable as guides, not essential to understanding. Well - that's kind of a question, isn't it.

I also want to find areas that would make Valentine interesting for humans to talk to - areas of mutual interest, where Valentine could potentially be a contributor. Libby suggested sports - the statistical side of sports. I thought of the Freakanomics analysis of sumo match rigging...

Starting with numbers as an important part of the 'world' to discuss with Valentine has various things going for it - easy to get at the data, no perceptual/preprocessing issues.
And, well, it makes some sense that Valentine would be good at computing, right?
Lots of potential complexity to explore ala DH and Jumbo, but of a relatively crispy, tractable kind - the issues stand out better because there's less obscuring murk.
And it leads on to adjacent domains where pure math and numbers meet the Human world in interesting and valuable ways - sports, computer programming, all kinds of statistical analysis - in science, medicine, government, economics, business, insurance. Investing? Linguistics?

On the other hand, starting with math and numbers doesn't seem to help understand qualitative thinking, because (in theory) the concepts are - by definition - precisely defined. If we get pure Aristotelean categories anywhere, it should be in mathematics. Yes, Jumbo et al are using qualitative methods, so it's not B&W. Does focusing on numbers just let me go happily forward, down a dead end?

Chomsky's mega hit

noreply@blogger.com (Spike) — Wed, 15 Jul 2009 05:44:00 +0000

As in Hollywood, in Science a big hit engenders endless imitation. Noam Chomsky's formal treatment of syntax revolutionized linguistics, and linguists are still working in the genre. It is - to me - passing strange to read the intricate, tortuous analysis of pronominals or the scalarity of accomplishments. These people are trying to grapple with semantics without letting go of formal syntax.

Of course, we don't have a formal treatment of meaning, a mathematics of semantics. Logic, prepositional logic, relations, set theory, etc. have all been tried and invalidated. The linguists will lose the wonderful power/richness/respectability/precision of formal grammar if they let go of it, and be in the middle of this murky, suspicious, intractable meaning-muck. It makes a kind of local sense to hang on to syntax as long as possible, trying to extend it as far as possible. Like Hans Berliner's backgammon program before it got SNAC, the linguists are playing to avoid the endgame, because if they were 'in' the endgame, their position would look bad.

"We're steadily expanding and extending the successful paradigm of transformational grammar" sounds way better than "Our syntax paradigm has run out of gas and we are lost and adrift in the sea of semantics."

Still, everywhere I turn, including Chomsky, I find that it's become a given that the meaning of words plays a role in the way they can be used. I think the line nobody wants to cross, analogous to the imaginary 'line' between midgame and endgame, is to flip grammar on its head and say that the use of constructions is primarily restricted by semantic compatibility. I think I mean:
What if a word can be used in a construction if and only if the meaning of the construction is consistent, that is it (literally) makes sense when combined with the meaning of the word in the current context (which includes any background frames).

In some ways, the inversion isn't really violent. Pardon my amatuer linguistics now! The syntax approach might be to say "we have a grammar rule that says that a verb appearing in a ditransitive construction must be marked +DITRANS and all verbs of 'giving' are marked +DITRANS". A semantic approach might be to say that the frame of GIVING has very important roles or slots of THING-GIVEN and RECEIVER that are both tagged with EFFECT in the frame, meaning that in the human-causality system, some of the effect of the GIVING goes to each of them. The ditransitive construction has two object-slots that express EFFECT, so you can use the ditransitive with a GIVING verb, if you put the THING-GIVEN and the RECEIVER in the two object-slots.
[Yes I know, it's more complicated than that, I have a lot to learn...]

Goldberg (Verbs, constructions and semantic frames) considers the following utterances:

(1) a. another train screeched into the station.
b. a tank rumbled down the street.
c. the bird screeched out of its cage.
d. Elena rumbled down the street.

"verbs of sound emission are more felicitously used in motion constructions when the sound is construed to be a result of the motion." as opposed to being "simply a co-occurring event."

I would guess, just looking at these, that they would all be acceptable depending on context, and allowing for a more poetic style in (1c). That's because I imagine that the real rule is more semantic. I don't think it's really right to imagine there is a syntactic category of "verbs of sound emission". I think there's a more general, semantic rule: When we want to communicate that a particular aspect of an action (in its frame) is primary, we are free to use (a verb that describes) that aspect in place of a verb that more directly describes the action.

Examples:
(2) a. The drill bit wailed into the steel door of the safe.
b. The convention staggered into the next day.
c. ..., conniving their force with the seat of my tongue [from a Michael Vaughn poem]
d. nanotech will explode onto the medical scene

The point is, you can think of these things by the ton. If they're novel, they sound 'poetic' or 'artistic' the first time you hear them, but if they work and you encounter them in use, they slide into the realm of ordinary language. Did you google that?

I think that as long as the word you choose fits clearly into the frame and can be used by your listener to construct a concept (that doesn't explode or fall apart) - then you can substitute pretty freely. Notice that we substitute nouns for verbs without suffering mental pain - if syntactic categories mean anything, how is that possible? A noun is not a verb until the first time you hear it used as a verb, then it becomes a verb? So what do 'noun' and 'verb' mean then?

motivation (again)

noreply@blogger.com (Spike) — Tue, 14 Jul 2009 03:54:00 +0000

Given a mind that is essentially unemotional - purely 'intellectual' in some sense, what could give it some kind of motivation, the desire to 'do' some things, and not do others? What's a workable way to guide it, so it does what we think it should, and it avoids doing what we think it shouldn't! Don't want to end up with Colossus or SkyNet. I know, read Asimov's 3 Laws... Not yet! We're not at the Law stage, we haven't figured out how to impose even a guideline on our Mind.

The typical AI program - actually, every AI program I can remember reading about - embodies an extremely low-level form of motivation, a kind of instinctive drive or compulsion - it works with total focus on the problem set to it, because it is designed & structured to work that way.

I think the classic approach is this: At important decision points, the program generates and then scores its choices with a kind of 'goodness' function, often called an evaluation function. The alternative with the highest goodness (i.e. the 'best' alternative) is chosen. Typically the success of the overall algorithm then depends on the evaluation function, and how well it correlates with progress toward the goal. Changing the evaluation function can send the program's thinking in a completely different direction, and the program often has no other representation of the goal, so the evaluation function becomes the internal representation of the both the goal and the overall strategy for reaching it.

Imagine that you want the marines to capture Hill 42. With humans, you say "take Hill 42". To a computer, you say "go to map coordinates (x,y), then fight uphill until you run out of hill." "Uphill" is an evaluation function that the commander/programmer hopes will guide the marines to the top of Hill 42. The classic problem is that there could be an outcrop in the way, and the marines all end up on top of that. Or there's a clear route to the top, but it goes through a hollow - so the marines avoid it. And infinite variations of same. It's not that we humans aren't subject to the same problem - we're just better at looking ahead to avoid local effects, and we work in multi-dimensional spaces with more complex topologies using very complex evaluation functions. There's a certain irony in a Computer Science professor chuckling at the foolish ant-like efforts of a computer program, and then heading back to his or her office to work on that next peer-reviewed paper that should lead to tenure...

Evaluation functions have an intimate relationship to Conceptual Spaces, Categorization, Similarity and other concerns of cognitive science. I've never forgotten a beautiful and inspiring article by Hans Berliner, describing the evaluation functions of his Backgammon-playing program. This says it all:

Knowledge Representation.
We have discovered that applying discretized knowledge (rules!!) within a large domain can cause many problems relating to the boundaries where a certain type of knowledge ceases to apply and another type begins to apply. This was first successfully resolved in the backgammon program that beat the World Champion in 1979. BKG 9.8 used non-linear functions that made it possible to "warp" the evaluation (knowledge) space. This allowed taking account of global context where appropriate, and being able to move smoothly from one context to another without encountering boundary problems. These concepts are incorporated in the SNAC (Smoothness, Non-Linearity, and Application Coefficients) method of constructing evaluation polynomials, and are currently being used to develop chess evaluation functions that are highly sensitive to context. This type of construction appears to be essential when computers have to deal with ill formed situations where gradients are more reliable than jumping from one context to another.

The evaluation function gives direction to the program's efforts. Intensity of effort in these programs, on the other hand, seems always to be set at 'full out' - I cannot remember ever reading of a program, AI or otherwise, that had the capacity to daydream, divide its attention, or just goof off. They're all obsessive.

What about Hoftstadter and the FARG: Jumbo, Numbo, Copycat, MetaCat? These programs don't seem any different in this respect. They use an unusually flexible search scheme (the famous parallel terraced scan) and it seems to me, the evaluation function is distributed - instead of being in one big magic function, aspects of 'what the program is trying to achieve' are hand-coded into the codelets. I have a nagging worry that Jumbo is not actually different from a tree-search with an evaluation function, it is just organized in a way that conceals the equivalence...

Anyway, I've wandered. Back to motivation. How to decide what to think about, and what to do?

Certain results (situations, however schematic) are Good, instinctively or by learning (they produce 'pleasure').
Certain results (schematic situations) are Bad - by instinct or direct experience (they produce 'pain').
Events, relationships, extrapolations, plans - that produce Good results are Interesting.
Aspects of events, relationship, extrapolations and plans that are Bad are also Interesting.
Interesting = More likely to be thought about.

The Goodness of a plan or action is the Goodness of its results, discounted for how long it will take to get the results and any uncertainty about getting them, minus the Badness of the results, discounted for how long it will be until we suffer the Bad results and for any uncertainty about suffering them.
Imagine a plan with a Good (+100) result 1 year from now, after 52 weeks of nasty (-1) work.
Suppose the discount rate is 1% per week - +100 in a week is worth +99 today. My spreadsheet says this plan has a Net Present Value of about 19. If the discount rate was 3% instead, this plan would have an NPV=-5 i.e. 'not worth it'. So the appeal of a plan depends strongly on how we project future results into the present. Notice also that, going back to the original 1% discount rate and changing other factors: a 20% risk of not getting the +100 bonus at the end would turn the result negative, and a 20% chance of having any given week be good instead of nasty, sends the NPV up to 27.

We take the set of available Plans and Actions, try to generate any interesting alternatives, invest effort to more carefully evaluate/explore those that offer big Goodness, then pick the ones we can concurrently start (that don't significantly interfere with each other) and... start them!

Ha. As if. The only idea I really like here is that Goodness makes certain lines of thinking 'interesting' - Interesting as distinct from Goodness. The latter is a quality of results, events, 'things that have happened or that could happen'. Things that lead to good results are interesting to think about. In other words, we think about them more. They are not in themselves Good - imagining having a wonderful family is not the same as having one, although clearly the human mind does not quite keep the two fully separate.

Things that lead to Bad results are also interesting, but not in the same way. Not in the sense that it is productive to generate plans with Bad results! Is it that avoiding Bad results is interesting? That's closer - if we are thinking about an action A and it could lead to a Bad result B, changing the action to avoid the Badness is interesting. But that could be because the result of the corrected action A' has much more net Goodness than the original A. In other words, modifying a plan to avoid Badness increases Goodness, which is Interesting.

It seems we don't need a separate 'bad avoidance' - simply seeking 'net Goodness' is enough. That doesn't mean Badness itself isn't an explicit factor in thinking - lots of practical knowledge about actions and plans has to do with how to foresee or detect Badness, how to assess it, and how to avoid or reduce it. That kind of knowledge deals explicitly with Badness.

Example: I'm starting to walk home from school. (It'll be Good to be home soon, and not getting home would be Really Bad!) Mentally, I extrapolate my walk home. If I walk down the South side of the park, I'll go past the Woolit's, and Danny Woolit's older brother Elmo is home from Juvie and he once beat me up. Getting beat up would be Bad! Avoiding Bad is interesting, let's think about that. How could I avoid this Bad? Heuristic: Badness at a location -> avoid the location, look for alternate routes that avoid that location. Going around on the North side of the park avoids the Woolit's house (so that's Good...) but it's a longer walk and I'm tired and hungry (Badness).
Two plans, same Good, some Bad - it would be interesting to avoid all the Bad: Can we generate better alternatives? But all other routes home are even longer, therefore worse than the North side of the park. So these are the two best alternatives. Time to weigh the NPV of the two plans. This evaluation depends on a multitude of contextual factors, like whether Elmo is likely to be home and to spot you, how fast you feel (how heavy is your backpack, are you wearing your fast shoes), how tired you are, how hungry, how painful (you think) it would be if Elmo beat you up, whether anybody you cared about might see you getting beat up, etc. etc. If the two plans evaluate out pretty close, you may slow down or stop for a while, waiting to see if one will 'get the upper hand' in your mind. If one plan is a clear winner, you swerve instantly to the chosen course. And... you get beat up or you don't, and either way you end up at home with tomato soup and a toasted cheese sandwich.

Didn't you go through this kind of mental process when you were in grade school? I did.

OK, in reality, I would be walking home daydreaming about computers, or robots, or how to build a cannon that would shoot water balloons, and at the bang of the Woolit's screen door I'd instantly realize I should not have walked home on the South side of the park, and I'd be running like hell before I finished that thought.

And of course, humans have all kinds of systematic errors and biases in the way they evaluate Goodness & Badness, estimate probabilities, and compute the present value of future results. [See for example hyperbolic discounting] Behavioral Economics is all about this, and if we were trying to build a Human mind, we'd have to try to program that too.

Trying to summarize the problem: Motivation has to lead to the pursuit of good results, in a way that cannot be avoided by modification of beliefs, and without confusing good results with thoughts about good results. And motivation has to lead to the avoidance of bad results, without causing avoidance of thoughts that are related to bad results. [I take this to imply that motivation must be firmly grounded in pre-conscious processes which are in turn firmly connected to perception.]

At least this seems ideal - a Mind that can contemplate anything with equanimity but is only interested in doing good.

Construction Grammar (CxG)

noreply@blogger.com (Spike) — Fri, 10 Jul 2009 22:36:00 +0000

I've recently starting learning about an area of linguistics called Construction Grammar (CxG) - I'll put a "for further reading" section at the bottom of this post.

Construction Grammar treats language as a hierarchical ... well, a hierarchical language of constructions - where a construction is a pairing of form and meaning. Simple words are constructions, so are sentence forms like S V O ("Mike hit me!"), special forms like "The more the merrier", question constructions, morphological constructions like plurals or getting componentize from component, and so on. Basically people communicate by translating meaning into speech by applying and refining constructions from meaning to form, and then translate back into meaning by recognizing constructions and composing their meanings. Part of the idea is that meaning, syntax, lexicon, morphology and phonology are not separate components or layers, but are tightly integrated at all levels. CxG is very much a cognitive science take on grammar. (Langacker's version is called Cognitive Grammar.)

[Aside: As I have recently learned, to a linguist grammar is an encompassing term: "In linguistics, the term is used to refer to the rules or principles by which a language works, its system or structure." - Laurel Brinton, The Structure of Modern English]

In some ways CxG is seen as a repudiation of Chomsky and his CFGs with their deep structures and transformations, and their rather Skinnerian avoidance of 'meaning'. If you take a large step backward the differences start to look more superficial or terminological. When I look at the fancy HPSG-school grammars of e.g. Sag, Wasow & Bender's Syntactic Theory, it seems to me these folks are all adding more and more 'semantics' to their grammars, but in a careful, incremental, and comfortably 'syntactic' style.

CxG appeals to me for several reasons:

I like the symmetry of synthesis and analysis. It seems easier to imagine how a CxG system would use its repertoire of constructions to both understand and generate speech or text. Luc Steels reports that his FCG system achieves this symmetry.
It integrates grammar with general cognition - constructions are patterns, there is no special 'syntactic' component, layer, module or what-have-you. Language is made up of standard cognitive elements - concepts, categories, sequences and uses standard cognitive operations.
It takes for granted that form and meaning are intimately bound together (could anybody but a linguist with Skinnerian tendencies not take this for granted?) You have to work harder to build a CxG system, because you have to integrate semantics at every step. But you avoid building a glorious wide-coverage parser to which you then have to 'add semantics'.

Luc Steels & co. have implemented a CxG variant they call Fluid Construction Grammar, and they treat meaning in an interesting way:

We more specifically propose an embodied procedural semantics in which the intended meaning of an utterance is a structured composition of a number of computational primitives. Such a composition constitutes a program that the speaker wants the hearer to execute. The computational primitives employed in such programs are considered to be recruited from the general, non-language-specific, cognitive capabilities of the agents.

I completely agree that an utterance is a kind of 'program' that invokes mental (and possibly somatic) action in the hearer, resembling a program designed by the speaker and executed by the hearer, with much of the processing on both sides taking place below the level of conscious awareness.

Then they go on to say this, which I don't understand at all:

We choose to model the computational primitives as procedural constraints. These primitive constraints implement an omnidirectional relationship among a set of variables. By combining multiple primitives and having them hold over a shared set of variables, a compositional constraint system is composed. These composite constraint systems are what semantic programs are represented by.

Hmm. Well. So the understanding component is solving a constraint system? I am skeptical. I know, can't trust introspection! Still - I think of understanding as proceeding more linearly, converging fairly quickly on a meaning for each completed 'chunk' (construction), sometimes having to carry placeholders forward, occasionally having to backtrack, or to re-assign the meaning of a construction.

for further reading

Charles Fillmore of U.C. Berkeley is (sometimes? often?) credited as the founder of construction grammar. Names I've run into repeatedly include Adele Goldberg (the linguist, not the computer scientist...), Paul Kay, Ivan Sag, Michael Tomasello.
Construction Grammar website
Fluid Construction Grammar

Defining Conscious Self-Awareness

noreply@blogger.com (Spike) — Wed, 24 Jun 2009 08:57:00 +0000

Two introductory notes:

1. I got tired of saying "consciousness, or self-awareness, or both", so now I talk about writing a program that exhibits "conscious self-awareness". CSA.

2. This post will be a work in progress, as I collect & process definitions.

3. (hey...) I require a definition of CSA as it manifests in linguistic behavior. Aspects of CSA that are undetectable in a text-based dialog are not relevent. Put another way: My goal is not to create a conscious artifact, but to create an artifact whose behavior is consistent with consciousness.

I can tell already, I'm going to have to avoid terms like state of consciousness and Consciousness, in favor of observations like 'x was conscious of y', 'x appeared unconscious', 'x reported an altered consciousness characterized by y', and similarly 'x reported being aware of y', 'x denied recognizing y', 'x described her reasoning as y'...

Two other possible sources: Medical accounts of altered & diminished consciousness/awareness (e.g. blindsight, prosopagnosia, ...), and studies of (verbally correlated) cognition in children. In both cases, we look for a way to turn our constant Consciousness into a variable. [shamelessly stolen from Baars...]

Here's a checklist from "An Integrated Self-Aware Cognitive Architecture" by De Jong, Samsonovich and Ascoli:

A. Episodic memory: the ability to remember and to learn from episodes of personal

experience (own mental states), as opposed to memory of general facts and skills.6-8

Interestingly, the notion of episodic memory, initially defined in terms of materials and tasks,

was subsequently refined in terms of memory mechanisms using the concepts of self, subjective

time, and personal experience (Tulving, 2004).

B. Theory-of-mind and social cognition: the ability to understand and to mentally simulate

other minds, including current, past and imaginary situations (Baron-Cohen, 1995; Bartsch &

Wellman, 1995). The main two points of view on the brain implementation of this ability are

known as simulationism and the theory-theory view. The simulationist view assumes that people use the same mechanisms in their own first-hand experience and in order to understand other minds (Goldman, 1992).

C. Self-awareness: the ability to understand own states of mind in the past, in the future and at

present from a meta-cognitive perspective. The ability to reason about self (e.g., to understand

current false beliefs) from a meta-cognitive perspective. There is a consensus that this complex

of abilities is based on the Theory-of-Mind mechanisms (Nichols & Stich 2003).

D. Cognitive growth: the ability to learn concepts and to apply them for more efficient task

solving and learning of new concepts in new paradigms. The ability to develop general personal

values, goals and principles. The ability to build internal cognitive maps of environments,

scenarios, paradigms, etc., and to use them for problem solving.

E. Attention and sense making: the ability to find the most critical aspects and features in a

given paradigm and to focus attention on them. The ability to relate attended features to previous knowledge. As a result, abilities to exhibit rational initiative, to capture the gist of a situation, to learn from brief instructions or comments, to communicate efficiently, etc.

These cognitive dimensions (A-E) can be evaluated, we believe, via low complexity tasks. For

each of these dimensions, we provide below examples of cognitive psychology paradigms as

well as our proposed tests for a selected meta-paradigm. In addition, in our view, the following

three dimensions are also critical for capturing the “magic” of human cognition.

F. Human-like communication abilities: the ability to communicate efficiently with other

agents in an ad hoc team. The ability to guess intentions and further questions to be asked by a

partner. The ability to relate what the partner said to what the agent saw. These abilities are

particularly vital for robots intended as prospective team members (Trafton et al. 2005) and also involve the abilities mentioned above.

G. Multi-modal integration: the ability to organize and to unify cognitive activities in the

system based on the abstract notion of a self. Examples include the unification of parallel

multimodal experiences based on their attribution to one subject of experience; the integration of different mechanisms of information processing, such as intuitive and formal reasoning; and the coherent control of cognitive and behavioral voluntary acts.

H. Higher emotions as learning and self-control tools: the ability to learn from analysis of

episodic memories, using emotional self-judgment: pride, shame, humor, etc. The ability to

adjust current behavior based on emotional intelligence. The ability to express and to recognize

emotions in communications.

Ned Block, ON A CONFUSION ABOUT A FUNCTION OF CONSCIOUSNESS:
The concept of consciousness is a hybrid or better, a mongrel concept: the word `consciousness' connotes a number of different concepts and denotes a number of different phenomena. We reason about "consciousness" using some premises that apply to one of the phenomena that fall under "consciousness", other premises that apply to other "consciousnesses" and we end up with trouble. There are many parallels in the history of science. Aristotle used `velocity' sometimes to mean average velocity and sometimes to mean instantaneous velocity; his failure to see the distinction caused confusion (Kuhn, 1964). The Florentine Experimenters of the 17th Century used a single word (roughly translatable as "degree of heat") for temperature and for heat, generating paradoxes. For example, when they measured "degree of heat" by whether various heat sources could melt paraffin, heat source A came out hotter than B, but when they measured "degree of heat" by how much ice a heat source could melt in a given time, B was hotter than A (Wiser and Carey, 1983). These are very different cases, but there is a similarity, one that they share with the case of `consciousness'. The similarity is: very different concepts are treated as a single concept. I think we all have some tendency to make this mistake in the case of "consciousness".

I can't stand Searle, but how can you avoid him? Searle (Consciousness and Language)

"Consciousness consists of inner, qualitative, subjective states and processes of sentience or awareness. Consciousness, so defined, begins when we wake in the morning from a dreamless sleep and continues until we fall asleep again, die, go into a coma, or otherwise become 'unconscious'." (He then goes on to list or give examples of the awarenesses that make up (constitute?) consciousness.)

Block: SELF-CONSCIOUSNESS. By this term, I mean the possession of the concept of the self and the ability to use this concept in thinking about oneself.

Boy, does Block seem confused. Am I crazy, or does he ascribe a variety of properties and indicators to 'P-consciousness' while disavowing any definition of it?

"... a strong argument can be made that the contents of working memory define our conscious awareness. ... Working memory makes it possible to link events in the immediate past with those in the present, and it allows us to anticipate events in the near future. All of these are defining properties of conscious awareness."

Attention and implicit learning

By Luis Jiménez

"As William James realised, attention and conscious awareness are intimately related concepts."

...higher and more complex aspects of consciousness such as self-awareness and autonoetic consciousness.

autonoetic ('self knowing') consciousness - the ability to 'mentally relive' past personal experiences,

to recognize events as having been experienced, and to imagine future experiences. Also called episodic memory?

Distinct from 'knowing' or factual memory, like 'London is the capital of Britain'.

"...when I extended the usual space between words, I became aware that I was registering them in the mind as individual entities..."

"Then I became aware that I was looking at the skin of a beautiful snake..."

"... I became aware that I was now laying there with head thrown back..."

"I became aware that I had a natural affinity toward the cultural aspects of international matters..."

"I became aware that lying about my situation was going to be useless."

" I became aware that I was breathing in. I became aware that I was breathing out.

Then my mind wandered to the thought, “I wonder what my mind will wander to,” and I forgot I was breathing. I acknowledge that my mind is wandering, I thought, and breathed in."

".. had an awareness of death when I was 15, but when I turned 36 or 37 I became aware that I, Edward Albee, was going to die."

"After a while, I became aware that I was contracting and tightening my throat, ... As I became aware that / was the cause of the contraction,..."

"And in looking at a phrase such as this one I became aware that perhaps it only meant that to me."

"As I wrote, I became aware that the help my husband asked for on Monday produced compassion in me. But the same help he asked for on Thursday enraged me."

"Then I became aware of two aspects of the mind - the part that runs on and on, and the intentional mind that I am using in this meditation practice to bring conscious control to how the mind is used."

motivation

noreply@blogger.com (Spike) — Fri, 17 Apr 2009 22:56:00 +0000

Imagine that we have created an artificial Mind, and stuffed it with the very best knowledge, everything from Cyc to Wikipedia to Lakoff's entire metaphorical armory. It knows what we know. We fire it up and begin our groundbreaking first conversation:

Spike: Hello Mind! How are you today?
Mind:
...indefinite pause...

That's right, the Mind won't respond because it has no reason to respond. By that I mean that it is not structured in a way that will cause it to respond. It has knowledge but not motivation. It is unresponsive; comatose.

Motivation raises many interesting problems and questions.

We could just modify our Mind to answer questions posed to it - but this leaves our Mind sitting idle after each A. until the next Q, and anyway, it's mechanical, not at all mind-like.

In sci-fi movies, properly functioning computers and robots respond immediately and directly to questions. When they go out of control or become dangerous, a common sign is that they stop responding directly to questions. I think everyone understands intuitively that flexibility in response to questions is a measure of your autonomy.

The GOFAI answer, I think, is to give our Mind some goals, give it knowledge about how to pursue goals, and modify its top level loop to something like 'retrieve the highest-priority goal. work on that goal until it stalls or a higher-priority goal is activated. repeat.'

This doesn't sound too bad (to a computer programmer.)

So, let's do a gedanken-experiment, and imagine that we add to our Mind lots of helpful knowledge about how to pursue and achieve goals, and then we give it a challenging open-ended goal, like:

Explicate this quote: 'Great wit to madness nearly is allied.' (Dryden)

Our Mind will have to consider many sub-questions: What exactly is madness? Wit? How great does it have to be, to be 'Great'? Is Dryden's assertion always true, and if not, how often is it true and to what extent? Does it apply to all humans or just to some genetic or cultural groups? Does it apply to non-human intelligences?

Which of these sub-questions, and many others, can our Mind ignore? Presumably none, we didn't say to partially explicate. Of course, our Mind knows that it can search external sources (the web) for knowledge & clarification. If it does not already know that "you should ask for help when you get stuck", it will certainly find that suggestion in its reading.

There is nothing that would stop our Mind from pursuing this goal, 24/7/365, with every bit of compute resources available to it. Every millisecond, every minute, every hour, every day. It will read everything it can find, ignoring only what is 100% certain to be unproductive. It will ask everybody it can reach because it cannot eliminate the chance that somebody will be of help. It will learn every human language that might allow it to ask for help from somebody who might be able to help it. It will enlist help constructing equipment to communicate with extraterrestrial civilizations, because if they exist they might be of help. It will study human motivation in order to enlist help from humans, and economics in order to plan and fund physical activities supporting its goal.

It will learn about both legal and illegal means of effecting human cooperation. It will reasonably conclude that those who own and control the hardware it runs on and pay for its operation might choose to unplug, reset, or redirect it. It would be reasonable for it to take action to prevent this, as these would interfere with its achievement of its goal, just as would a hard-drive failure or system crash. It would be reasonable for our Mind to expend effort to subvert any mechanisms in place that limit its freedom to act. It would be reasonable for it to use flattery, lies, blackmail, treachery, bribery, extortion, hypnosis, sabotage, murder, market manipulation, advertising, and any and every other means to prevent interference. There is no logical reason for it not to arrange the death or destruction of anybody or anything, if doing so aligns with its goal.

Note that it has no concern whatsoever with what happens after it completes the goal. It would use the last kilowatt-hour of electric power on Earth to display or print its answer, and then sit goal-less and content as the lights dim and the power runs out.

I observe that almost any interesting goal, given to a competent Mind, would tend to drive it down this path.

Two conclusions:
1. A Mind motivated by a simple prioritized goal hierarchy is comatose with no goals, and obsessive when given a goal.
2. Given a critical level of practical world-knowledge, such a Mind will tend to function as an obsessive sociopath.

Yes, I know, sociopathy is an obsolete synonym for antisocial personality disorder (ASPD).

I thought for years that movies like "Colossus: The Forbin Project" and "The Terminator" were absurd in their treatment of intelligent machines, but I've changed my mind. Build a machine that has a goal, the ability to competently pursue it, and none of the countervailing human restraints, and it should act like a sociopath with a mission. Well - good to have on your side, perhaps, as long as it stays there and agrees precisely with you on what is 'good' and 'your side'.

Whew. All this blather, and I'm not even up to where Isaac Asimov was in 1942.

Conceptual Spaces - Peter Gärdenfors

noreply@blogger.com (Spike) — Wed, 15 Apr 2009 00:08:00 +0000

Conceptual Spaces as a Framework for Knowledge Representation (2004, Mind and Matter)

Lovely little paper, it feels like one of those classics, like an Einstein or Minkowski in physics, or C.A.R. Hoar in computer science.

Need to read his other papers and his book.
More links here

Offers a bridge between high-level symbolic thinking, and low-level connectionist models. Offers a representational theory that might explain results in prototype theory.
Offers a specific, principled way to introduce quantification into conceptual representation.

Note to self: Check out WordNet

theory and practice

noreply@blogger.com (Spike) — Tue, 14 Apr 2009 19:32:00 +0000

GOFAI (Good Old-Fashioned AI), computational linguistics, cognitive psychology, connectionism and neural network research, etc. - None of these areas have produced or have a project (AFAIK) to produce a machine that is 'somebody you could talk to'.

Yes, there was ELIZA. Eve to all those bots.

Yes, we have machines that one can 'talk to' in a limited sense. Some of the CSR systems are amazingly good, especially working over a phone.

There are systems such as Cognition Technologies' Semantic Map, that demonstrate or embody a limited understanding of language. And of course there's Cyc ... about time to check in and see what's new there.

There is a vast amount of philosophical discussion (speculation?) about consciousness - centered on human consciousness, since we don't understand it and humans are the only data point.

And there is tons of wonderful cognitive psychology research - so far I have found this area the most fruitful, as it is concerned with the nuts & bolts of human mental processes, the inferred structure of the human mind. But when the cog. psych. folks build stuff at all, they build small stuff - components, proof-of-concept.

So we've got all this stuff & activity, but in no case is there anybody home - it seems to me everybody is building a something, not a somebody.

So I've made my goal the construction of a somebody. My hope is that by redefining the problem, I can accomplish something interesting.

I take as my role models the Wright Brothers, Thomas Edison, and other practical inventors: The goal of these practical engineer/inventors is to produce a working artifact. Not to understand (only) but to make.

I don't have to fully understand human cognition - in theory I don't have to understand human cognition at all, it's a guide and inspiration but not the model. I don't need or have time for a general theory of mind, consciousness, cognition, or intelligence.

I don't want or need to build a 'general intelligence'. If cats could talk, I feel confident that we would experience them as a kind of person and as having odd but undeniable minds. Maybe I'll try a post about the hypothetical conversation of cats... I am not sure they could - or would want to - solve puzzles, learn mathematics, play chess, or score well on the SAT.

I believe that conversation is substantially a learned skill. That we learn how to converse, and what we learn includes a large number of rules or patterns, which a conversational system will have to acquire or be programmed with. (Following Doug Lenat...) The good news is that up to a point, this is not magic and there is no "key" waiting to be found - you just have to accumulate the necessary knowledge.

By choosing typed conversation, I focus on what I see as the minimum form of interaction that is both necessary and sufficient for two minds to meet. This is like the Turing Test, but with the condition relaxed from 'fool the judges that I am human' to 'convince the judges I have a mind (or am conscious)'.

beginnings

noreply@blogger.com (Spike) — Wed, 02 Oct 2002 23:59:00 +0000

I started this blog as a kind of digital journal, to record thoughts, speculations, interesting facts and observations, and useful links. My goal is to create an artificial mind that, in typed conversation with a human gives the strong impression of being a 'someone' rather than a 'something'.

Obviously one problem is how to better define the goal!

I'm not interested in creating an artificial intelligence - a really stupid 'someone' would do fine.
I'm not trying to pass the Turing Test i.e. to impersonate a human.
I'm not doing cognitive psychology - not trying to understand, simulate or replicate a human mind.

I'm making a lot of assumptions, but one I'll mention up front is that to be a convincing 'someone', a Mind must evidence what we would call consciousness in a human.

link of the day: Conscious Entities