Purpose: Discussion of Strong AI Minds thinking in English, German or Russian. Privacy policy: Third parties advertising here may place and read cookies on your browser; and may use web beacons to collect information as a result of ads displayed here.

Thursday, August 30, 2012

aug23ruai

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

1. Thurs.23.AUG.2012 -- Diagnosing Selection of Subjects

As we troubleshoot the Dushka Russian AI in JavaScript for Microsoft Internet Explorer (MSIE), probably the first point of departure must be inserting a diagnostic "alert" message to let us know how the NounPhrase module is selecting the subject for a sentence of thought in Russian. No matter how a subject is chosen, we want the verblock mechanism to force the retrieval of a particular verb from the so-called IdeaPlex.

Our first major problem after some human input is that NounPhrase selects as most activated a potential subject of "tebya" with a carried-over, spurious "verblock" that does not even lead to a verb, but rather to "tebya" itself. We have probably solved this problem already in the English JSAI.

By searching downwards for "tqv" (the source of "verblock") simultaneously in the Russian JSAI and the English JSAI, we discover that in the English JSAI on 15aug2012 we inserted into InStantiate a line of code to prevent spurious carry-overs of the "tqv" value when "seq" is at zero. Now we insert the same code into the Russian JSAI. Running the AI, we do not get an improvement. Then at the start of WhoBe we also put a zeroing out of "tqv" taken from the English JSAI. Still there is no improvement. In the Russian AI, we then intoduce into WhatBe the same tqv-zeroing as was done in the English JSAI. Again there is no improvement.

2. Fri.24.AUG.2012 -- Affecting Activation of Subjects

It may be necessary to implement code that will switch from an oblique case of an activated concept and find nominative nodes to serve as the subjects of incipient thoughts.

If a direct object is left activated at the end of a sentence, all the nodes of that concept should receive a blanket activation through OldConcept or NounAct. Then NounPhrase may choose nominative nodes as candidates for the subject of a sentence. (Maybe we should make nominative nodes receive a higher activation.) So the process of having a residually activated concept switch from being a direct object in an old thought to being a subject in a new thought should work by whatever mechanism puts a blanket activation on all the nodes of a concept.

3. Tues.28.AUG.2012 -- Finding "verblock" Verb-forms

We need to put in some diagnostic messages and see what residual activation occurs for a direct object.

Today in the VerbPhrase module we are building up some code which, in the presence of a positive verblock, will still go to the "verblock" time-point in the Ru-array but will not automatically accept the verb-form originally deposited there, typically during human input. Instead, the new code conducts a search of the ruLexicon to find a verb-form with the correct number and person. Initially we forgot to search for the concept-number, so we accidentally got the correct ending but on the wrong verb.

4. Wed.29.AUG.2012 -- VerbGen Returns Inveniend Verb-stem

Yesterday we made some major progress in getting the RuAi to search for correct Russian verb forms, but the new code was not yet perfect, so today we need to make improvements. However, we should probably save and archive yesterday's version so that we can recover from any unforeseen errors.

Now there is a problem because the new, integrated search-code is finding the correct archival verb-form, if it is available, but the verb is appearing in duplicate. Apparently the rest of the VerbPhrase code is finding a "vphraud" recall-vector all over again. We should be able to thwart that phenomenon.

As we start to prepare some documentation of the AudBuffer, OutBuffer and VerbGen modules, we notice that our Russian AI code needs to make use of pertinent variables such as the "gencon" status flag and the "audbase" recall-vector to identify the verb whose inflectional ending must be changed. As soon as we use "audbase" in our code, the Russian AI stops switching to a different verb and at least outputs the stem of the verb that we are trying to change. Since we have also set the "gencon" flag, VerbPhrase calls VerbGen but does not make its normal main call to SpeechAct, so we do not get an extra verb-form as output.

5. Thurs.30.AUG.2012 -- VerbGen Needs "dba" Parameter

Yesterday VerbGen was returning only the stem of an inveniend verb and not the inflected personal ending. However, delivering the stem was a major improvement in the Russian AI functionality. Today we found that we needed only to set the "dba" parameter properly before calling VerbGen, and the Russian AI was able to provide a correct form of the required verb.