Older blog entries for mentifex (starting at number 89)

Artificial Intelligence in Russian

Sat.6.OCT.2012 -- Negation of Russian Verbs

In the free, open-source Russian artificial intelligence (RuAi), we need to work on the negation of verbs before we can implement the calling of the VisRecog module from the VerbPhrase module. When we type "Ty ne znayesh menya" for "You do not know me" into the current RuAi, it answers incorrectly "Ya znayu ne tebya" for "I do not know you," and the negational adverb "nye" for "not" is in the wrong place.

After experimentation with diagnostic "alert" messages, we moved the nay-saying code into the same area of VerbPhrase that uses parameters to select a Russian verb-form. Thus we got the RuAi to put the negational adverb before the verb and not after the verb.

Sun.7.OCT.2012 -- Negation of Putative Be-Verbs

Our next task in creating Russian artificial intelligence is to implement the negation of unexpressed, putative be-verbs in Russian. Currently the Dushka AI assumes tentatively the occurrence of a be-verb after any noun or pronoun begins a statement in the nominative. We need to introduce special handling of the negative adverb "NYE" so that the RuAi still waits for a putative be-verb. Although we are tempted to let there be a corrigend default negation of each putative be-verb, we owe it to Occam to let an actual negation determine what happens.



MindForth Programming Journal -- 2012 June 29

1Fri.29.JUN.2012 -- IdeaPlex: Sum of all Ideas

The sum of all ideas in a mind can be thought of as the IdeaPlex. These ideas are expressed in human language and are subject to modification or revision in the course of sensory engagement with the world at large.

The knowledge base (KB) in an AiMind is a subset of the IdeaPlex. Whereas the IdeaPlex is the sum totality of all the engrams of thought stored in the AI, the knowledge base is the distilled body of knowledge which can be expanded by means of inference with machine reasoning or extracted as responses to input-queries.

The job of a human programmer working as an AI mind-tender is to maintain the logical integrity of the machine IdeaPlex and therefore of the AI knowledge base. If the AI Mind is implanted in a humanoid robot, or is merely resident on a computer, it is the work of a roboticist to maintain the pathways of sensory input/output and the mechanisms of the robot motorium. The roboticist is concerned with hardware, and the mind-tender is concerned with the software of the IdeaPlex.

Whether the mind-tender is a software engineer or a hacker hired off the streets, the tender must monitor the current chain of thought in the machine intelligence and adjust the mental parameters of the AI so that all thinking is logical and rational, with no derailments of ideation into nonsense statements or absurdities of fallacy.

Evolution occurs narrowly and controllably in one artilect installation as the mind-tenders iron out bugs in the AI software and introduce algorithmic improvements. AI evolution explodes globally and uncontrollably when survival of the fittest AI Minds leads to a Technological Singularity.


2 Fri.29.JUN.2012 -- Perfecting the IdeaPlex

We may implement our new idea of faultlessizing the IdeaPlex by working on the mechanics of responding to an input-query such as "What do bears eat?" We envision the process as follows. The AI imparts extra activation to the verb "eat" from the query, perhaps first in the InStantiate module, but more definitely in the ReActivate module, which should be calling the SpreadAct module to send activation backwards to subjects and forwards to objects. Meanwhile, if not already, the query-input of the noun "bears" should be re-activating the concept of "bears" with only a normal activation. Ideas stored with the "triple" of "bears eat (whatever)" should then be ready for sentence-generation in response to the query. Neural inhibition should permit the generation of multiple responses, if they are available in the knowledge base.

During response-generation, we expect the subject-noun to use the verblock to lock onto its associated verb, which shall then use nounlock to lock onto the associated object. Thus the sentence is retrieved intact. (It may be necessary to create more "lock" variables for various parts of speech.)

We should perhaps use an input query of "What do kids make?", because MindForth already has the idea that "Kids make robots".


3 Sat.30.JUN.2012 -- Improving the SpreadAct Module

In our tentative coding, we need now to insert diagnostic messages that will announce each step being taken in the receipt and response to an input-query.

We discover some confusion taking place in the SpreadAct module, where "pre @ 0 > IF" serves as the test for performing a transfer of activation backwards to a "pre" concept. However, the "pre" item was replaced at one time with "prepsi", so apparently the backwards activation code is not being operated. We may need to test for a positive "prepsi" instead of a positive "pre".

We go into the local, pre-upload version of the Google Code MindForth "var" (variable) wiki-page and we add a description for "prepsi", since we are just now conducting serious business with the variable. Then in the MindForth SpreadAct module we switch from testing in vain for a positive "pre" value to testing for a positive "prepsi". Immediately our diagnostic messages indicate that, during generation of "KIDS MAKE ROBOTS" as a response, activation is passed backwards from the verb "MAKE" to the subject-noun "KIDS". However, SpreadAct does not seem to go into operation until the response is generated. We may need to have SpreadAct operate during the input of a verb as part of a query, in a chain were ReActivate calls SpreadAct to flush out potential subject-nouns by retro-activating them.


4 Sat.30.JUN.2012 -- Approaching the "seqneed" Problem

As we search back through versions of MindForth AI, we see that the 13 October 2010 MFPJ document describes our decision to stop having ReActivate call SpreadAct. Now we want to reinstate the calls, because we want to send activation backwards from heavily activated verbs to their subjects. Apparently the .psi position of the "seqpsi" has changed from position six to position seven, so we must change the ReActivate code accordingly. We make the change, and we observe that the input of "What do kids make?" causes the .psi line at time-point number 449 to show an increase in activation from 35 to 36 on the #72 KIDS concept. There is such a small increase from SpreadAct because SpreadAct conservatively imparts only one unit of activation backwards to the "prepsi" concept. If we have trouble making the correct subjects be chosen in response to queries, we could increase the backwards SpreadAct spikelet from one to a higher value.

Next we have a very tricky situation. When we ask, "What do kids make?", at first we get the correct answer of "Kids make robots." When we ask the same question again, we erroneously get, "Kids make kids." It used to be that such a problem was due to incorrect activation-levels, with the word "KIDS" being so highly activated that it was chosen erroneously for both subject and direct object. Nowadays we are starting with a subject-node and using "verblock" and "nounlock" to go unerringly from a node to its "seq" concept. However, in this current case we notice that the original input query of "What do kids make?" is being stored in the Psi array with an unwarranted seq-value of "72" for "KIDS" after the #73 "MAKE" verb. Such an erroneous setting seems to be causing the erroneous secondary output of "Kids make kids." It could be that the "moot" system is not working properly. The "moot" flag was supposed to prevent tags from being set during input queries.

In the InStantiate module, the "seqneed" code for verbs is causing the "MAKE" verb to receive an erroneous "seq" of #72 "KIDS". We may be able to modify the "seqneed" system to not install a "seq" at the end of an input.

When we increased the amount of time-points for the "seqneed" system to look backwards from two to eight, the system stopped assigning the spurious "seq" to the #73 verb "MAKE" at t=496 and instead assigned it to the #59 verb "DO" at t=486.


5 Sun.1.JUL.2012 -- Solving the "seqneed" Problem

After our coding session yesterday, we realized that the solution to the "seqneed" problem may lie in constraining the time period during which InStantiate searches backwards for a verb needing a "seq" noun. When we set up the "seqneed" mechanism, we rather naively ordained that the search should try to go all the way back to the "vault" value, relying on a "LEAVE" statement to abandon the loop after finding one verb that could take a "seq".

Now we have used a time-of-seqneed "tsn" variable to limit the backwards searches in the "seqneed" mechanism of the InStantiate module, and the MindForth AI seems to be functioning better than ever. Therefore we shall try to clean up our code by removing diagnostics and upload the latest MindForth AI to the Web.


Artificial Intelligence in Russian

1. Thurs.9.FEB.2012 -- Unspoken Be-Verbs as a Default

The Russian-speaking artificial intelligence Dushka needs a default BeVerb module that will silently assert itself as the automatic carrier of thought until a non-be-verb takes over from the provisional default. In our coding of a Russian mind, we will assume that any noun or pronoun, beginning a thought in the nominative case, is automatically the subject of a putative BeVerb until proven otherwise. In this way, our cognitive software will prepare for a BeVerb and switch automatically when a non-be-verb occurs.

We should work first on the comprehension of putative be-verbs and second on their generation, so that what we learn in comprehending be-verbs may be used in generating thoughts involving a BeVerb. So we type into the AI a Russian sentence to see if the software can understand it.

Human: душка робот

Robot: ДУШКА ЧТО ДУШКА ТАКОЕ

We said "Dushka is a robot" but the AI responded only, "Dushka -- what is Dushka?" We need to implement a default BeVerb in the comprehension of a sentence that lacks a visible BeVerb.

In the InStantiate module, we can trap for the input of a "c==32" space-bar when the "seqneed" is set to "8" for want of an incoming verb. We may then do something outrageous, but normal for Russian. From InStantiate we may provisionally send into AudMem a space-bar character with an "audpsi" of "800" for the verb БЫТЬ ("to be"), so that the AI is ready to record any noun coming in as a predicate nominative in conjunction with the be-verb. Now, if we implement such an outrageous step, it is possible that our AI memory-banks will become replete with quasi-spurious engrams of infinitive be-verbs that typically do not materialize. It could be that the presence of a spurious be-verb engram will not matter, if the cancellation of the default occurs as soon as some actual verb comes in. Then cancelling the spurious default will involve removing or nullifying any associative tags laid down momentarily during the enactment of the default.

2. Fri.10.FEB.2012 -- Instantiating Imaginary Be-Verbs

In the InStantiate module we will now experiment with code to create in auditory memory a pseudo-engram of a non-existent be-verb after the perception of a nominative noun or pronoun. Since the Russian-speaking mind waits for a predicate nominative, it needs at least an imaginary be-verb as the holder of associative links between subject and predicate nominative.

Now inside InStantiate we have assembled the code that creates a be-verb pseudo-engram in the three memory arrays for "Psi" concepts, Russian words and auditory engrams. The Psi node is automatically creating a "pre" tag that links the pseudo-verb back to its subject. We need to implement code that will finish the intermediation of the unspoken Russian BeVerb between its subject and the predicate nominative. The code must also cancel or uninstall the imaginary BeVerb if a real verb occurs instead of the provisionally expected BeVerb.

3. Sat.11.FEB.2012 -- Integration of Default Be-Verbs

We have the AI pretending that a BeVerb comes in after a nominative subject, and now we need to create the "seq" tag from the subject to the default BeVerb. First in the InStantiate module we insert a line of code declaring that the pseudo-be-verb is indeed a verb with respect to its part of speech, so that the following code will try to reach backwards to the subject engram and install a "seq" tag referring to the now not-so-imaginary BeVerb. We run the Dushka AI and we type in, ты робот -- which is Russian for "You are a robot", but without the be-verb. We are puzzled when Dushka answers, Я ЧТО Я ТАКОЕ ("I -- WHAT AM I?") and that's all she wrote. It may indicate that her concept of self has been activated by the input referring to "you", but she does not seem to have understood the input. We check the diagnostic display, and we see that her concept of self now has a "seq" tag referring right back to herself instead of to the default Russian BeVerb. What went wrong? We look at the JavaScript source code again, and we see that it was not enough to set the part-of-speech as a verb. We go ahead and we set the Psi concept-number to be that of the Russian be-verb. Then we run the Russian AI again with the same input and we sit there in shock when the AI announces to us: Я РОБОТ. Dushka has just said to us, "I AM A ROBOT" in Russian. From the diagnostic display we discover that the same changes that made Dushka able to understand the idea, made her able to think the idea.


Artificial Intelligence in Russian

Fri.3.FEB.2012 -- Recognizing Inflections

For the Russian-thinking Dushka AI Mind, we have perhaps stumbled upon a way to avoid the hard-coding of noun paradigms and instead to let the Russian AI learn the inflected endings of Russian nouns from its own experience. For example, right now the Russian artificial intelligence (RuAi) fails to recognize the Psi concept #501 БОГ in the following exchange.

Human: я уважаю бога ("I honor God.")
Robot: ТЫ УВАЖАЕШЬ БОГА ("You honor God.")

Robot: ЧТО БОГА ТАКОЕ ("What is God?")

The diagnostic display reveals that the software has almost recognized the word for God.

559. Б 0 * 1 1 0
560. О 0 * 0 1 0
561. Г 0 * 0 1 501
562. А 0 * 0 0 902
Aha! Suddenly it becomes clear that two things are happening. The Psi concept #501 is indeed being recognized at first, but perhaps the provisional-recognition "prc" variable is not being set, and so AudInput calls NewConcept as if the AI were learning a new word instead of recognizing an old word.

Sat.4.FEB.2012 -- Learning Russian Like a Human Child

Now in a very rough way we have trapped for "zad1" in the AudRecog module so as to recognize a noun (БОГА ) with one character of inflection added onto it. Because the noun was indeed recognized, the InStantiate "seqneed" mechanism tagged the noun in the "ruLexicon" with a "dba" of "4" to indicate a direct-object accusative case. In other words, the Russian AI learned a new noun-form as a human child would learn it, that is, from the speech patterns of another speaker of Russian.


Artificial Intelligence in Russian

Tues.31.JAN.2012 -- Generating and Recognizing Verbs

In our Dushka Russian AI we have the problem that new verb-forms generated on the fly by the VerbGen module are not being recognized and tagged with critical parameters as they settle into auditory memory. However, it looks as though a verb does get recognized if the "audpsi" tags for the verb in auditory memory extend far back enough to cover the stem of the verb. Therefore, instead of devising ways to bypass the operation of ReEntry calling AudMem, calling AudRecog, we should perhaps instead implement a "backfill" of any verb generated in the VerbGen module to let the "audpsi" tags extend back to the last "pho(neme)" of the verb-stem. Then the "provisional recall" mechanism in AudRecog ought to recognize the verb-form generated by the VerbGen module.

We created a "vip" variable to hold the value of "motjuste" when VerbPhrase calls VerbGen and to transfer the known concept-number of the verb, near the end of the stem in VerbGen, into the provisional "prc" variable for AudRecog. In this way, we got the AI internally to recognize and record verb-forms generated internally by the VerbGen module. However, to get the AI to call the correct verb-forms, we had to modify some recent OldConcept code for deciding what "dba" value to store with a lexical item. Now we have a problem with tagging the "dba" of a simple word like МЕНЯ when it comes in.

We can not rely on the form of МЕНЯ to tell us its "dba" because it could be genitive or accusative. We need to extract clues from the incoming sentence in order to assign the proper "dba" during the storage of МЕНЯ.

Wed.1.FEB.2012 -- Tagging Engrams with Parameters

We can perhaps rely on the "seqneed" mechanism of InStantiate to provide the "dba" parameter for a noun or pronoun entering the mind as user input. (Perhaps the "seqneed" variable should change to a "seqseek" variable for greater clarity.) We may be able to strengthen the use of "seqneed" by adding a kind of "pass-over" when a preposition is encountered, so that the software continues to look for a direct-object noun when a preposition-plus-noun combination is detected and skipped.

Where the InStantiate module tests for a "seqneed" of "5" and encounters a satisfying noun or pronoun to become a "seq" for the verb, we make the assumption that the time "t" identifies the temporal location of the noun or pronoun in both the Psi array and the "ruLexicon" array. We insert two lines of code to first "examine" the Russian lexical array and then to substitute a numeric "4" for the "ru4" flag of the "dba" value. Since the noun or pronoun is going to be the "seq" of the verb, that same noun or pronoun warrants a "dba" of "4" as a direct object that should be in the accusative case. However, we may need to make other arrangements if the verb is intransitive and the noun must be in the nominative as a predicate nominative.


Artificial Intelligence in Russian

1. Sun.29.JAN.2012 -- Verbs Without Direct Objects

Today in the Dushka Russian AI we begin to address a problem that occurs also in our English AI Mind. Sometimes a verb does not need an object, but the AI needlessly says "ОШИБКА" for "ERROR" after the verb. We need to make it possible for a verb to be used by itself, without either a direct object or a predicate nominative. One way to achieve this goal might be to use the jux flag in the Psi conceptual array to set a flag indicating that the particular instance of the verb needs no object.

We have previously used the "jux" flag mainly to indicate the negation of a verb. If we also use "jux" with a special number to indicate that no object is required, we may have a problem when we wish to indicate both that a verb is negated and that it does not need an object, as in English if we were to say, "He does not play."

One way to get double duty out of the "jux" flag might be to continue using it for negation by inserting the English or Russian concept-number for "NOT" as the value in the "jux" slot, but to make the same value negative to indicate that the verb shall both be negated and shall lack an object, as in, "He does not resemble...."

During user input, we could have a default "jux" setting of minus-one ("-1") that would almost always get an override as soon as a noun or pronoun comes in to be the direct object or the predicate nominative. If the user enters a sentence like "He swims daily" without a direct object, the "jux" flag would remain at minus-one and the idea would be archived as not needing a direct object.

2. Sun.29.JAN.2012 -- Using Parameters to Find Objects

While we work further on the problem of verbs without objects, we should implement the use of parameters in object-selection. First we have a problem where the AI assigns activation-levels to a three-word input in ascending order: 23 28 26. These levels cause the problem that the AI turns the direct object into a subject, typically with an erroneous sentence as a result.
In RuParser, let us see what happens when we comment out a line of code that pays attention to the "ordo" word-ordervariable. Hmm, we get an even more pronounced separation: 20 25 30.

Here we have a sudden idea: We may need to run incoming pronouns through the AudBuffer and the OutBuffer in order unequivocally to assign "dba" tags to them. When we were using separate "audpsi" concept-numbers to recognize different forms of the same pronoun, the software could pinpoint the case of a form. We no longer want different concept-numbers for the same pronoun, because we want parameters like "dba" and "snu" to be able to retrieve correct forms as needed. Using the OutBuffer might give us back the unmistakeable recognition of pronoun forms, but it might also slow down the AI program.

Before we got the idea about using OutBuffer for incoming pronouns, in the OldConcept module we were having some success in testing for "seqneed" and "pos" to set the "dba" at "4=acc" for incoming direct objects. Then we rather riskily tried setting a default "dba" of one for "1=nom" in the same place, so that other tests could change the "dba" as needed. However, we may obtain greater accuracy if we use the OutBuffer.

3. Mon.30.JAN.2012 -- Removing Engram-Gaps From Verbs

Yesterday in the Russian AI we experimented rather drastically with using the "ordo" counter to cause words of input to receive levels of activation on a descending slope, so that the AI would be inclined to generate a sentence of response starting with the same subject as the input. We discovered that the original JavaScript AI in English was not properly keeping track of the "ordo" values, so we made the simple but drastic change of incrementing "ordo" only within OldConcept and NewConcept, both of which are modules where an incoming word must go through the one or the other.

Today we have sidetracked into correcting a problem in the VerbGen module. After input with a fictitious verb, VerbGen was generating a different form of the made-up verb in response, but calls to ReEntry were inserting blank aud-engrams between the verb-stem and the new inflection in the auditory channel. By using if (pho != "") ReEntry() to conditionalize the call to ReEntry for OutBuffer positions b14, b15 and b16, we made VerbGen stop inserting blank auditory engrams. However, there was still a problem, because the AI was making up a new form of the fictitious verb but not recognizing it or assigning a concept-number to it as part of the ReEntry process.


Artificial Intelligence in Russian

Thurs.26.JAN.2012 -- Insufficient Activation of Subjects

The most glaring problem in the Dushka Russian AI right now is that the AI does not fully activate the subject-pronoun when we type in a short sentence of subject, verb and object. Without a proper subject to provide parameters, the AI fails to select or generate a proper Russian verb-form.

When we type in "люди знают нас" ("People know us"), as an answer we get "ВАМ ЗНАЮТ ТЕБЯ" -- a mishmash of "to you" "they know" "you". In general, the AI seems to be taking the final object entered as input and trying to convert it into the subject for a response.

Thurs.26.JAN.2012 -- Using the "seqneed" Variable

The Russian AI is not setting a Psi "seq" flag when we enter a Russian word as the subject of a following verb. When we inspect the recent 10nov11A.F MindForth code for clues, we discover that in October of 2011 we made major improvements to the method of assigning "seq" tags. We began using the "seqneed" variable as a way of holding off on assigning a "seq" until either the desired verb or the desired noun/pronoun made itself available. However, apparently in the English JavaScript AI we wrote the "seqneed" code only for needing nouns and not yet for needing a verb. No, we did write the code, but it involved avoiding the English auxiliary verb "do", so we accidentally removed the verb-seqneed code from the RuAi. Let us put most of the code back in, and see what happens. Upshot: Once we put the code back into InStantiate, subjects of verbs once again began having a "seq" reference to the verb. The AI even skipped an adverb that we then inserted as a test.


Artificial Intelligence in Russian

Fri.30.DEC.2011 -- Russian AI Bootstrap Words

In the ru111229.html version of the Dushka Russian AI we coded the AudBuffer to load Russian characters during SpeechAct and the OutBuffer to move each Russian word into a right-justified position subject to the changing of inflectional endings based on grammatical number and case for nouns, and number and person for verbs. Next we need to determine which forms of a Russian word are ideal for storage in the RuBoot bootstrap sequence.

It seems clear that for feminine nouns like "ruka" for "hand", storage in the singular nominative should suffice, because other forms may be derived by using the OutBuffer to remove the nominative ending "-a" and to substitute oblique endings of any required length.

For regular Russian verbs in the group containing "dumat'" for "think" and "dyelat'" for "do", it should be enough to store the infinitive form in the RuBoot module, because the OutBuffer can be used to create the various forms of the present tense. If a human user inputs such a verb in a non-infinitive form, such as in "ty cheetayesh" for "you read", the OutBuffer can still manipulate the forms without reference to an infinitive. This new ability is important for the learning of new verbs. Since there is no predicting in which form a user will input a new Russian verb, the OutBuffer technique must serve the purpose of creating the verb-forms and of tagging their engrams with the proper parameters of person and number.

Since JavaScript is not a main language for artificial intelligence in robots, our Dushka Russian AI serves only as a proof-of-concept for how to construct a robot AI Mind in a more suitable language. We use JavaScript now because it can display the Russian and because a Netizen can call the AI into being simply by using Internet Explorer to click on the link of the Душка AI Mind.


Russian AI Mind Journal

These notes record the coding of the Russian AI Mind Dushka
in JavaScript for Microsoft Internet Explorer (MSIE).

Mon.26.DEC.2011 -- Creating the OutBuffer

Today in the Dushka Russian AI we will try to create the OutBuffer function to change the declensional ending of a Russian verb.

Tues.27.DEC.2011 -- Right-justifying ЗНАЮ

Although we have created the OutBuffer module to permit the SpeechAct module to hold a Russian verb right-justified in place for a change of inflectional endings on the fly, we are finding it difficult to obtain an "alert" report of the exact contents of the OutBuffer towards the end of a pass through SpeechAct. Into SpeechAct we put a diagnostic "alert" box, and then it appeared that OutBuffer was being called but no data were being revealed.

By testing for the contents when four characters trigger an IF-clause, we have determined that the OutBuffer does indeed take a word from the PhoBuffer and display the word in a right-justifed position. We were able to toggle from English to Russian typing and input the Russian verb for "I know", which soon showed up in a right-justified location when the WhatBe module asked a question about the Russian word. Now we are ready to design code that will intercept a Russian verb being "spoken" and change its inflectional ending on the fly, a feat which we will consider to be a major advance in our creation of a Russian AI Mind.


JavaScript Russian AI Programming Journal

These notes record the coding of the Russian AI Mind Dushka in JavaScript for Microsoft Internet Explorer (MSIE). The free, open-source Russian AI will grow large enough to demonstrate a proof-of-concept in artificial intelligence, until the intensive computation of thinking and reasoning threatens to slow the MSIE Web browser down to a crawl. To evolve further, the Russian AI Mind must escape to more powerful programming languages on robots or supercomputers.

Thurs.22.DEC.2011 -- Selecting the Bootstrap Vocabulary

We would like soon to implement the diagnostic display so that we may observe the build-up of the innate Russian vocabulary. Therefore we copy the necessary code from the English AI Mind and we troubleshoot until suddenly we observe a diagnostic display. Now the way is clear for us to keep adding Russian words until we have enough innate concepts for the Russian AI to demonstrate thinking in Russian.


80 older entries...

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!