Here is an update on some recent OpenCog work that may be relevant to some of you…

(Email samir@vettalabs.com if you have detailed questions)

The main update is that the “virtual pet QA system” now answers questions regarding many more spatial relationships; for the full list see

http://www.opencog.org/wiki/EmbodimentLanguageComprehension_Questions

(The work here was in creating rules to identify when these relationships hold between objects as perceived and recorded in the LocalSpaceMap. Ideally these rules would be learned of course, but for a short-cut we’ve hand-coded them…. Making these rules is easy in itself, but required various improvements to the virtual-pet perception infrastructure, which were useful improvements anyway…)

Along the way to doing this, the SpaceServer and LocalSpaceMap inside OpenCog were extended to use 3D rather than just 2D like before. What Samir and Fabricio are working on now is:

1) Loading FrameNet into the Atomspace to enable better reference resolution (this will also be useful for PLN work). They are using an XML version of FrameNet so they don’t need to spider the website…. Note that our license for FrameNet covers only research uses; you need to pay to use it for commercial apps…

2) Building a simple OpenGL-based visualizer so as to visualize what the virtual pet is seeing at each point in time (this should be useful for debugging issues with spatial reasoning, etc.). Basically this is a visualizer for the 3D LocalSpaceMap structure inside OpenCog…

During some recent reading, it struck me that a useful framework for thinking about and talking about sentence generation is the MTT or “meaning-text theory” of Igor Mel’cuk, et al Here is one readable reference:

Igor A. Mel’čuk and Alain Polguère, (1987) “A Formal Lexicon in Meaning-Text Theory”, Computational Linguistics, vol. 13, pp. 261-275.

portal.acm.org/citation.cfm?id=48160.48166
www.aclweb.org/anthology/J/J87/J87-3006.pdf

Within the context of that theory, the output of the Stanford parser is strictly at the SSynR or “surface syntactic representation” level, while, as a general rule Relex attempts to generate the DSynR or “Deep syntactic representation” structure. Some of what I’ve been trying to do with opencog is towards the “SemR” structure, as described in that paper.

The more I read about MTT, the more it seems to capture some of what we are trying to do (defacto are doing) with NLP within opencog. In particular, the MTT concept of a “lexical function” (which is not really described in that paper??) could be a particularly strong way of guaranteeing correct syntactic output for segsim, nlgen or NLGen2

– Linas Vepstas

Time that we post a status update!

OpenCog has been a little more quiet than usual over the last couple of months. The developers list is still sporadically active, but some of the main developers are having to spend time on other work related projects meaning less AGI-driven focus (want to change that? donate here). We’re following several options for establishing further funding for the end of 2009 and through 2010, but we’ll see how that goes.

Instead of writing a long summary post, I’ll just give some bullet points:

• Dr. Ben Goertzel spoke on building beneficial AGI at the Singularity Summit last month (video here).
• Cassio Pennachin and Dr. Joel Pitt attended the GSoC Mentor’s Summit at the Googleplex in Mountain View, which led to meeting FOSS developers from around the word. This also allowed them to meet up with Moshe Looks (MOSES and PLOP author) for dinner and discussions around AGI with a foray into Newcomb’s Paradox.
• Dr. Linas Vepstas released RelEx 1.2.1, an affiliate OpenCog project, along with the related project/dependency Link-grammer 4.6.5

.

I’m sure there are other items of note, so to the other contributors reading this, please feel free to comment and I’ll update this post ;-)

I spent the weekend comparing the Stanford parser to RelEx, and learned a lot. RelEx really does deserve to be called a “semantic relation extractor”, and not just a “dependency relation extractor”. It provides a more abstract, more semantic output than the Stanford parser, which sticks very narrowly to the syntactic structure of a sentence.

I wrote up a few paragraphs on the most prominent differences; most of my updates were to the RelEx dependency relations page.

Here are the main bullet points:

• RelEx attempts basic entity extraction, and thus avoids generating nn noun modifier relations for named entities.
• RelEx will collapse the object and complement of a preposition into one. Stanford will do this for some, but not all relationships.
• RelEx will convert passive subjects into objects, and instead indicate passiveness by tagging the verb with a passive tense feature.
• RelEx avoids generating copulas, if at all possible, and instead indicates copular relations as predicative adjectives, or in other ways.
• RelEx extracts semantic variables from questions, with the intent of simplifying question answering. For example, “Where is the ball?” generates _pobj(_%atLocation, _$qVar) _psubj(_%atLocation, ball), which can then pattern-match a plausible answer: _pobj(under, couch).
• RelEx attempts to extract comparison variables.

Its also clear to me that I could split up the relex processing into two stages: one which generates stanford-style syntactic relations, and a second stage that generates the more abstract stuff. This might be a wise move … Since RelEx is already more than 3x faster than the Stanford parser, this could attract new users.

– Linas Vepstas

I’ve recently resumed work on the question-answering chatbot, and am trying to get it to comprehend a broader range of questions and statements. The “big idea” is to create a number of “sentence patterns” that the pattern matcher can recognize and respond to. The reason this is a “big” idea is because I am trying to avoid anything algorothmic or procedural — everything is to be done by specifying OpenCog hypergraphs, and NOT by writing C++ code, or scheme code (or python code…etc). The reason for working entirely with patterns and hypergraphs, rather than with C++ or scheme, is because this puts the “knowledge” of the system into a form that AI routines can manipulate it: learning algos can learn new hypergraphs; statistical algos can gather usage information on which hypergraphs get triggered, and so on. This is all easer said than done: although I’ve eliminated a fair amount of question-answering code previously written in C++, I’ve also had to write some new scheme code. Bummer. :-(

Patten matching is now used through-out all of the OpenCog NLP pipeline, although not in a unified manner. The Link Grammar parser uses patterns (called “disjuncts”) to determine how the words in a sentence can link to one-another, thus “parsing”, or pulling the grammatical structure out of a sentence (this paper provides an excellent overview). The RelEx dependency relation extractor applies patterns on the link-grammar output to extract syntactic relations. For example, the sentence “John threw a rock” becomes

_obj(throw, ball)
_subj(throw, John)

after RelEx gets done with it. And now, there are a dozen patterns inside of OpenCog that can pick out certain kinds of questions and statements from RelEx output, and pattern-match questions to find answers to them.

For example, the new OpenCog patterns convert “The capital of France is Paris” into

capital_of(France, Paris)

and similarly, “What is the capital of France?” into

capital_of(France,what)

Treating “what” as a variable, there is yet another pattern that matches up the form of the question to the form of the answer, thus deducing that “what” must be “Paris”.

Somewhat harder is using patterns to distinguish similar from dis-similar concepts, so that sentences like “John threw a green ball” aren’t used as answers to questions such as “Did John throw a red ball?”: the word “ball” with modifier “green” has to be detected as a different entity than the word “ball” with modifier “red”: these are two different entities (called “semes” in the code). In fact, out of laziness, I’ve punted on this one: the promotion of word-instances to “semes” is done by code, rather than by pattern matching. But soon, I hope, this will change. In the meanwhile, the README file provides a more detailed discussion.

Here are some patterns that work these days:

<me> John threw a green ball.
<me> Fred threw a red ball
<me> Mary threw a blue rock
<me> who threw a ball?
<cogita-bot> Syntax pattern match found: Fred John
<me> who threw a red ball?
<cogita-bot> Syntax pattern match found: Fred

<me> Did Fred throw a ball?
<cogita-bot> Truth query determined “yes”: throw

<me> Did Fred throw a red ball?
<cogita-bot> Truth query determined “yes”: throw

<me> The color of the book is red.
<me> What is the color of the book?
<cogita-bot> Triples abstraction found: red

<me> the cat sat on the mat
<me> what did the cat sit on?
<cogita-bot> Triples abstraction found: mat

And here are some that don’t yet work: “Did Fred throw a green ball?” — gets no reply, because the system can’t find an answer, and doesn’t make the common-sense leap of “can’t find answer-> answer must be no”. Another common-sense problem is illustrated by: “Did Fred throw a round ball?” — the system doesn’t know that balls are round, and simply assumes that a “round ball” is some special kind of “ball”. Oh well. There’s work to be done.

You can try out the chatbot yourself (when its up, and not broken!) on the IRC chat channel #opencog on the freenode.net chat servers.

– Linas Vepstas

“The Singularity Institutes “main purpose” is meant to be to investigate whether a recursively improving intelligence can maintain “friendliness” towards human kind. “

Okay, but my standpoint is: Why does the recursively improving intelligence need to be non-human? It seems counter-intuitive to me to devolve this power to something outside of ourselves – and also a bit like we’re just trying vainly to become a kind of God, creating another type of being.

I think the main reason there is a focus on AI rather than improvement of human intelligence is because it’s so damn hard to do experiments on people’s brains. It’s ethically difficult to justify various experiments, and it only gets harder as things become more regulated (and rightfully so for the most case). I think they’ll definitely be continuing research into this stuff though. For myself, occasionally taking Modafinil enhances my productivity significantly (so long as I maintain focus on what I’m meant to be doing, it’s easy to get enthralled with something that interests me, but isn’t related to my work).

But there’s no exclusion of human intelligence amplification from the singularity concept. If we create smarter humans, then this begets even smarter humans. Again we can’t really predict what those enhanced “humans” would do, because they are a significant step smarter than us.

Human intelligence amplification has a whole raft of other ethical issues associated with it though too. When it becomes more mainstream/available it’s going to be a major political and social issue. What happens when not everyone can afford (or wants) to enhance themselves? Will we develop two classes? One of naturals and another of post-humans? Will employers require certain professions to use performance enhancement (say for example, for brain surgeons performing long surgeries)? It’s also going to raise the question about the ownership of our bodies. There are laws against taking recreational drugs, but for some, LSD helps with certain types of thought and could be seen as a form of an intelligence manipulator (or an amplifier of certain facets of intelligence).

At the moment at least, governments and enforcement agencies seem completely uninterested in actively stopping this, due to the prevalence of various performance enhancement drugs throughout academia and other cognitively demanding professions. Obviously it’s not necessary, but for some the edge or boost it gives them is sufficient to outweigh the risks of off-prescription drug use.

That was my main beef with Kurzweil: He assumed that the intelligence beyond the horizon would be non-human. This, of course, begs the deeper philosophical question: what *is* “human?” But I’m sure you’ve mulled that one over plenty. Indeed, it would be interesting to hear what you have to say on the subject.

It simply strikes me that, the human ego being what it is, we would naturally be trying to improve our own intelligence and not worrying too much about creating AI as a standalone entity. Am I wrong? Is the human ego instead more interested in giving birth to a new species?

There is one other reason there is a focus on AI. Which is related to human ego. Once humans have an advantage over others, such as a significant step up in intelligence, then there’s a good
chance some will use that power over others in a negative way. It’s the old adage, “power corrupts, and absolute power corrupts absolutely” – and it’s my opinion, that this could lead to the same power/intelligence imbalance between an intelligence-amplified human and humans, as that between humans and dogs.

Or more succinctly, humans have plenty of evolutionary baggage that we
don’t necessarily want to amplify!

Yeah I’ve thought about that a little as well. What I wonder (and this really is just pure speculation) is whether hyper-intelligent humans would have more faculty for reason and logic; and, if so, if they would see the benefits in spreading the gift and/or using their gifts in more compassionate or benevolent ways…? Because, surely, that’s evolution as well.

Again, it would seem somehow counter-intuitive to improve intelligence without, for example, attempting to improve our capacity for processing emotions — our ‘emotional intelligence’, as it were.

I tend to think that many of the world’s problems today come down to a focus on what is scientifically possible rather than what is philosophically ‘good’ or ‘true’.

Basic agreement, improving intelligence is really just a catch all for all sort of cognition improvements. Depending on where we focussed neuron regeneration, we could expect to see improvements in different aspects of cognition.

Science and technology will inevitably advance, and while we can regulate technology to some extent, we can’t stop it. Telling people they can’t work on AI would merely push it underground, and one of the reason I work on OpenCog is because it’s an open-source framework. It can be inspected by other experts for flaws to ensure there isn’t some hidden time-bomb sitting in the code, and allows an international approach which might otherwise be over-regulated within a university-specific project (this ignores the actual issue of friendly vs. unfriendly AI, since the latter doesn’t have to be intentionally designed – I may write more about that one day, but SIAI and it’s fellows have already written plenty on that).

To bring it back to AI, what place is there for emotions in an artifically constructed intelligence? Is there a capacity for compassion or fear, if they serve a function for processing input and generating responses? I am guessing that cognitive scientists have thought about this one…

That’s a BIG topic, so all I’ll say is that my opinion is that emotions are just a particular mind state. Sure this mind state is influenced by hormones and neurotransmitters, but so is the rest of the brain’s functioning.

My opinion is that emotions are just extremely strong aspects of a human’s mental world. If we gave an AI an extremely strong desire/goal to make humans happy, and the AI design had some kind of reward based system (such that the AI was trying to maximise these) and achieving these rewards caused other effects in the AI’s mind, such a propensity to use positive phrases to describe the world, then is there any reason to believe that the AI isn’t happy itself?

Disclaimer: I’m no longer in the employ of the SIAI, so my thoughts are not endorsed by them at all.

The link-grammar parser uses labeled links to connect together pairs of words. In order to capture the idea of proper grammatical construction, any given word is only allowed to have very specific links to its right or left: for example, verbs have their subject on the left, and an object on the right. Link-grammar defines hundreds of different link types, and there are typically dozens or even hundreds of ways that these can attach to a word. Each allowed set of links is called a “disjunct”. So, for example:

MVp- Js+

is a disjunct that says “there must be an MVp link from this word, going to the left, and an Js link, going to the right”. This disjunct commonly connects prepositions to a verb on their left (the MV- link) and the object of the preposition on the right (the J+ link).

A good way to think about disjuncts is to imagine them as very fine-grained part-of-speech tags. Thus, when one sees “MVp- Js+” associated to a word, one knows not only that the word is a preposition, but even a bit more: its a preposition that took a singular object. Disjuncts classify words not just into crude part-of-speech categories, but much finer categories: thus verbs are not just as transtivie or intransitive verbs, but mgiht be transitive verbs that take both direct and indirect objects, or participles, etc.

Siva Reddy, a GSOC 2009 summer student, prepared a table of the frequency of occurrence of different disjuncts in a large collection of text. The top six entries are

Ds+ 950275.635843
Xp- 838569.90527
A+ 616522.664867
AN+ 566658.997313
MVp- Js+ 563082.649325
MVp- Jp+ 446487.310222

and these are exactly what one might expect:

• Ds+ connects the determiner “the” to nouns: and of course, “the” is the most frequent word in the English language.
• Xp- connects the period at the end of the sentence to the start of the sentence, so of course its frequently observed.
• A+ connects adjectives to nouns, AN+ connects noun modifiers to nouns.
• As noted above, MV connects verbs to modifying phrases, and J connects prepositions to objects, so that MV- J+ is the disjunct that most prepositions will get. Js connects to a singular object, Jp connects to a plural count or mass noun.

A graph of rank vs. frequency is shown below:

Disjunct rank vs. frequency of occurance

As can be seen, the distribution is more or less Zipfian, with a power-law exponent of 1.5. The fact that the long tail appears to be linear indicates that grammatical construction in the English language appears to be more ore less scale-free: difficult and akward constructions are increasingly rare. The fact that the graph is not purely Zipfian, but instead has a knee for the most common grammatical connections suggests that the most common grammatical constructions are “less common than they should be”: almost as if English speakers are resisting the use of formulaic sentence constructions. So, for example, since adjectives and noun-modifiers appear near the top of the rank, this suggests that English speakers “could have” used more adjectives and noun-modifiers, but didn’t. Quite why this is so is not clear. Perhaps the use of anaphora and references in general helps decrease the need for lots of modifiers.

The open questions are then:

1. Why a power law of 1.5?
2. Why is there a knee?
3. Does this result hold for other languages?

The corpus used here consists of approximately 1 million sentences, obtained by parsing entire Wikipedia articles, Voice of America news stories, and 10 books from Project Gutenberg, including War and Peace, Jane Austen, and some scientific or medical texts.

– Linas Vepstas

Recently Jared Wigmore, a student of Waikato University, New Zealand, created a tool for visualizing PLN as part of a visualisation project.

BIT visualizer

In my opinion, the BIT visualiser shows great promise as a tool for understanding the complexities of BIT expansion. In particular, the cross joins between sub-trees make it much clearer how sharing of sub-trees is occurring. The size of the BITNodes reflect their fitness evaluation in determining which node of the inference tree will be expand next and will inevitably be useful when we get to the stage of tuning the the fitness heuristic.

Being a prototype, there is plenty of scope for continued development, a couple of the many ideas that immediately come to mind are:

• expansion of BITNodes by clicking on them (this would require OpenCog to provide an XML-RPC interface first however), and
• thematic colouring of rules so that it’s easier to distinguish between the subtrees.

This is part of the bigger challenge of general AtomSpace visualisation. How do we convey knowledge about the processes that are going on in a digital mind to humans in a meaningful way?

Hands-on tutorials are planned for the next month or so; we’ve already had a few on PLN, and my turn is coming up, for the opencog NLP pipeline. So I thought I’d wire up a cute demo for the occasion: a rough, crude IRC chatbot, “La Cogita”. It can answer simple questions about straightforward statements. Nothing fancy … it doesn’t do any reasoning at all … but it can work off of the basic syntactic structure of English sentences to find answers. Thus, for example:

<linas> Mary ate a mango
<cogita-bot> Hello linas, parsing ...
<cogita-bot> linas, you made a statement: Mary ate a mango
<linas> what did Mary eat?
<cogita-bot> Hello linas, parsing ...
<cogita-bot> linas, you asked a question: what did Mary eat?
<cogita-bot> The answer to your question is: mango

Its meant to be a demo of the basic NLP pipeline within OpenCog. It takes input text, runs it throught the Link Grammar + RelEx parser, imports the results into the OpenCog atomspace, sucks in a small common-sense database, and waits for questions to be posed. The common-sense database is derived from MIT’s ConceptNet (OpenMind/CommonSense project), and so one can have interactions like the following:

<linas> cogita-bot: what is a saxophone?
<cogita-bot> Hello linas, parsing ...
<cogita-bot> linas, you asked a question: what is a saxophone?
<cogita-bot> No answer was found to your question.
<linas> hmm
<linas> cogita-bot: what is an instrument?
<cogita-bot> Hello linas, parsing ...
<cogita-bot> linas, you asked a question: what is an instrument?
<cogita-bot> The answer to your question is: woodwind r bass harmonica An_Oboe Oboe megaphone saxophone chronometer drum scale ukulele cymbal instrument
<linas> Heh. Complete with assorted linguistic garbage :-)

You get the idea. Don’t ask it anything more complicated than the above examples: it will fail to find any answer. Again, it does no reasoning at all. Its as thick as a brick. You can test-drive it at the #opencog channel on the freenode.net IRC network. Assuming its not down for development.

Next up: wire in NLGen for natural-language output, and start taking baby steps in actual reasoning. Anyway, I’m pretty excited, as this means that a lot of the basic bits&pieces are working, and I can now dive into the deep end, and start working on the hard stuff.

– Linas Vepstas

The decision process for the 2009 GSoC projects has been completed. You can read Ben’s announcement on the opencog-soc Google group.

The accepted projects are:

• Joel Lehman – Extending MOSES to evolve Recurrent Neural Networks
• David Kilgore – Python Interfaces For OpenCog Framework API
• Ruiting Lian – Natural Language Generation using RelEx and the Link Parser
• Rui Liu – Application of Pleasure Algorithm Project
• samir souza – Integration of Language Comprehension with Virtual Agent Control in OpenCog
• siva reddy – Statistical Learning and Refinement of RelEx Graph Transformation Rules
• Jeremy Schlatter – Distributed and Persistent AtomSpace
• Kemal Eren – Neurobiological data analysis in OpenBioMind
• Xiaohui Liu – Improved hBOA by integrating the BBHC and implement the simulated annealing algorithm

More detail is available for each on the GSoC OpenCog home page.

This interesting fact and this analysis, thus this is the point.

Except that made for very short essays that were no where near the word limit.

Part of that was because I was used to maths, physics, and computer science problems. These fields don’t tend to push students to writing supposition about theory, at least not at the University I went to, and it’s usually pretty concrete for the level taught to undergraduates. Then midway through my study I switched tracks into Biology and flailed around for a while. They expected you to write more than a paragraph at a time, and form coherent arguments in exams without the benefit of cut-and-paste! The horror. After while I got over my initial shock, and actually got reasonably good at writing (my recent resurgence of posts on this blog is an attempt to practice those skills which are getting rusty from too much code and analysis of experimental results). Even though I managed to get pretty good, I still didn’t quite understand. I thought it was a waste of time when I could construct the argument for my 2000 word essay in a paragraph.

I got over that though. I saw that one has to also explain their reasoning, because any conclusions about the real world based on science usually contain assumptions. It’s important to not only be explicit about these assumptions, but to also explain why these are the assumptions chosen.

Even more recently, I’ve worked on economic attention allocation (ECAN) in OpenCog, which controls the flow of attention within an artificial mind. Why is this necessary? Because the mind’s resources are finite – based on the computers available memory and processing ability, and we obviously also have a finite capacity for focussed thought too (well, yes, it’s finite because we only have so many neurons, but I’m referring more to the limited number of concepts and relations we can consciously work with at any one time). Much of how the attention allocation works in OpenCog is based on importance diffusion between close concepts, but the concepts also need to be primed with importance in some way. If diffusion was the only process that moved attentional importance, you’d eventually end up with a homogeneous soup where everything was equally important!

The way this happens is somewhat complicated in OpenCog, and not immediately relevant to the point, but one thing to note is that external stimulus excites contextually relevant information. What that means is, if OpenCog sees that a cat walked past… then knowledge about cats (they are furry 4 legged felines) and perhaps locomotion/kinematics is stimulated (if the cat keeps walking in that direction it will fall off that balcony and fall due to gravity… okay, so maybe that’s more about prediction, but general knowledge about walking would be available in order to be used to make that prediction). By stimulating this knowledge, it’s made more important in the systems mind. Beyond the cat example[1] the same goes for reading someone’s ideas. If someone just tells you “the sky is green, purple and sparkly” and leaves it at that, you’d probably just think they were taking some particular effective hallucinogenics… especially if you were inside, at a conference about refrigerators (if it was a metereological conference, you might be willing to concede that they were right, because they are an expert in the field and have a reason for believing as such). But if the the refrigerator conference attendee first prefixed their statement with “Y’know, I was just holidaying in Alaska, and I saw the Northern lights, up there… “, then assuming you vaguely knew about auroras changing the colour of the sky, then you’d be able to believe the refrigeration expert. Or at least believe they weren’t in the habit of attending refrigeration events under the influence of mind-altering substances.

[1] It’s okay, the cat didn’t actually fall off the balcony.

We are happy to announce that the SIAI has been selected again this year to participate in the Google Summer of Code program as a mentoring organization. GSoC is an annual program that awards successful student contributors a 4500 USD summer stipend to work on open source and free software projects for three months. Around one thousand students worldwide participated in GSoC 2008, with eleven students working on OpenCog related projects. Students may apply for GSoC 2009, beginning at the SIAI organization page. The student application period closes on April 3, 2009 at 19:00 UTC.

I’ll be attending Burning man. I’m immensely looking forward to this as this is the first year in several that’s actually been feasible for me to get there from New Zealand. I’ll be with an Australian theme camp called Straya that a friend of mine put me in contact with, and who’ll also be there.

As well as Burning man, I plan to hang out in Washington D.C with Ben to talk about our work on OpenCog. Then I’ll stay in San Francisco for 5-6 weeks (end of Sep till start of Nov) to attend the Singularity Summit followed by the CogDev Workshop (an OpenCog coding jam, details to be finalised, but likely to be just after the Summit).

If you’ll be at any of these events and want to chat, drop me a line

I spent the day trying to formalize a mapping of RelEx output into OpenCog, so that I can apply wordnet to perform word-sense disambiguation. I’m excited, yet despairing on how I’m just barely scratching the surface, and how it just takes so much time to get anything done. I’ll post the resulting spec to the opencog mailing list in a few days.

Now that doesn’t mean a consciousness can’t be revived, the pattern is still there, and if the process can be restarted then I suspect the consciousness would continue as if nothing happen. One moment about to die, the next revived. This is essentially what proponents of cryogenics expect to occur.

Did I just contradict myself, by saying that consciousness can be revived from the pattern, even though I claimed the pattern wasn’t the mind? I don’t believe so. The pattern is the painting, the mind is painter. In humans, the painter is the physiological processes that generate the electrical signals shooting through our body and that update the neuronal structure in our brain.

Want to work on AI/language-processing over the Northern Hemisphere summer? Here are some of the ideas for projects proposed. Applications to Google open on the 24th of March.