| Author |
 Topic  |
|
Armageddon2036
Curious Member
Italy
86 Posts |
 Posted - Nov 16 2008 : 10:21:32
|
Hi. I download ubcd for Windows when It is updated and I built with It an updated version of a bootable windows XP CD. This way is something bad happens I have a goog probability to repair my PC.
Instead of using a windows xp vmware machine in ram disk, what about
using a vmware machine that loads ulitimate boot cd for windows 3.22 using an iso stored in RAM? This way a PC with 1-2 GB should be enought. The same thing with knoppix. Is a system like that protected against any malicious software or not?
Thinking about that, I would like in Guile3D sdk a tool to build a bootable version of Denise also using ultimate boot CD 3.22. It would be cool an "intelligent" version of a bootable CD that depending on the situation tries to restore a damaged PC. In this case Denise updates could be saved on an external usb key and any procedure successfull could be added to Denise server to enhance her PC repair ability. Everything in order to let her "grow" in any situation.
9) Armageddon2036: "Build a bootable windows CD, please"
Denise: "Searching nlite new version detected 1.4.9.1 ... uninstalling, downloading, updating and installing... done ... searching Ultimate Boot CD for Windows new version detected 3.22 ... uninstalling, downloading, updating and installing... done ... using original Windows XP installation disk ... Windows XP SP2 ... searching ... Windows XP SP3 detected ... downloading ... done ... starting nlite 1.4.9.1 and slipstreaming Windows XP SP3 ... creating iso image ... done ... mounting iso ... starting Ultimate boot CD for Windows ... creating Ultimate Boot CD for Windows ... waiting ... done ... UBCD for windows iso created ... mastering iso image ... done ... Do You want to try It in a vmware machine ?" Armageddon2036 "not now, thank you"
hi. |
Edited by - Armageddon2036 on Nov 16 2008 12:05:52 |
 |
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 16 2008 : 16:29:26
|
Guile some questions for you!
1) I think you have done an amazing work with AIML, but have you addressed its most basic problem? i.e., the fact that the patterns must be said verbatim to match? I mean in your example what if I will ask "I would like to know more about AI" will your system match it with the pattern "What else do you know about *". Your three-layer approach is innovative, but if the system will fail to match the user's input with a pattern in the first layer, the rest will fail as well.
2) Let's assume that Denish will have to tell a story to the user (e.g., a 700 word story). a) does your system allow automating annoation of texts with animation tags? b) at least do you have some kind of interface in your engine to assist the user in this process? If yes does this interface create the AIML files automatically as well?
I envision an interface where the user will have to worry only about the story creation and not about the technical parts (e.g.to annotating texts with animation tags), but is it perhaps to early to ask for this? |
Edited by - giannis on Nov 16 2008 16:37:38 |
|
|
|
Guile3D
Hooked Member
USA
661 Posts |
Posted - Nov 17 2008 : 14:21:21
|
quote:
This looks excellent Guile. The bit where it says How can I help you Guile - did you respond via voice or by text? Also is the camera bit ready? I noticed it on the beta - does it follow your eyes?
regards
Tom
In this example I responded by text. The voice recognition is ready to be used as input for English. Still working on some issues to control echo and noise (Iīm sure your work in this field will produce much better results using voice).
For command and control the engine works with 98% accuracy. For direct dictation using Microsoft engine I have 70 to 90% of accuracy. With Nuance Dragon I can reach more than 90% with previous voice training.
One of my ideas is to build a real-time spell-checker in the input box for voice rec.
I have the motion detection ready. Iīm using this feature to trigger events like start recording surveillance video.
Face recognition is 100% done. Will have to make this module less memory hungry, so it will not interfere with the character motion.
Face detection to follow userīs head position with characterīs eyes and head are almost ready.
|
Guile 3D Studio
http://www.guile3d.com
"Where reality meets Imagination" |
Edited by - Guile3D on Nov 17 2008 14:24:20 |
|
|
|
Guile3D
Hooked Member
USA
661 Posts |
Posted - Nov 17 2008 : 14:50:14
|
quote:
Originally posted by giannis
Guile some questions for you!
1) I think you have done an amazing work with AIML, but have you addressed its most basic problem? i.e., the fact that the patterns must be said verbatim to match? I mean in your example what if I will ask "I would like to know more about AI" will your system match it with the pattern "What else do you know about *". Your three-layer approach is innovative, but if the system will fail to match the user's input with a pattern in the first layer, the rest will fail as well.
2) Let's assume that Denise will have to tell a story to the user (e.g., a 700 word story). a) does your system allow automating annotation of texts with animation tags? b) at least do you have some kind of interface in your engine to assist the user in this process? If yes does this interface create the AIML files automatically as well?
I envision an interface where the user will have to worry only about the story creation and not about the technical parts (e.g.to annotating texts with animation tags), but is it perhaps to early to ask for this?
Gianis, AIML is as good as the work done by itīs creator on typing all variations and possible combinations for an specific question.
Although itīs limitations and patterns matching philosophy, it has some nice features that give us good flexibility to address this problems, like the tags SRAY and THAT.
This tags allow us to code our AIML ready to understand almost all possible variations for a specific question:
For the former example used (What else do you know about *), the bot would also understand is we ask:
- What is AI?
- Tell me more about AI.
- What do you know about AI?
- Do you know something about AI?
- Iīd like to know what AI is.
and even more complicated ones, like:
- AI? What is it?
- I need to make a college work about AI. Can you help me?
Of course we have to code all this variations at least once, so the engine will use it for every different question. This will work because the main subject is referred by a wild character (<STAR> or *), in our case, "Artificial Intelligence".
The tag <THAT> works like a short term memory, trying to link actual subject with past conversation. Itīs very limited, so this is one of the reasons I created the tag <PASCAL>. With this tag every conversation session with all itīs properties and attributes (verbs, predicates, adjuncts etc..) are saved into a SQL database.
All this data is then pos-processed using the third layer Iīve mentioned before. For this real AI layer, Iīm using some concepts from the "Transformational grammar" inspired by Chomsky's Minimalist Program (http://en.wikipedia.org/wiki/Transformational_grammar)
About your last question, the engine will have a build in AIML editor. Itīs something similar of what we find at PandoraBot:
When we donīt like the bot response, or the bot gives a too generic answer (probably one that came from the "*" Ultimate AIML pattern, that is the one triggered when thereīs not a exactly match), weīll have a pop-up AIML editor where we can write a more appropriate answer. This will append code to a "user.aiml" file.
This editor will also have all pascal scripts, character animation, expressions and other functions that this engine will understand.
|
Guile 3D Studio
http://www.guile3d.com
"Where reality meets Imagination" |
Edited by - Guile3D on Nov 17 2008 15:16:37 |
|
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 17 2008 : 17:16:08
|
One last question for everyone....
To my knowledge Guile's engine, is currently the one and only tool on the planet that brings the vision of full multimodal Virtual humans several steps closer. What do you say? All opinions are welcome. I have seen several software packages dealing with one aspect of VH (e.g., 3D animations) but nothing so complete. Can we consider Guile3D as the only company that currently owns such technology? |
|
|
|
Dante
Hooked Member
220 Posts |
Posted - Nov 17 2008 : 17:17:43
|
| I too have seen many projects that would connect a character to a conversational agent, none, however, are so vast, powerful and potentionally having a bigger effect then this. |
The hand that rocks the AI cradle, may well rule the world. |
|
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 17 2008 : 18:17:34
|
Guile another question for you!!!
You are saying:
<<Of course we have to code all this variations at least once, >>
If I understand correctly what you saying, each variation for every single question/phrase will have to be coded seperately correct?Let's put some numbers down: In a hypothetical situation where a researcher has a language corpora of 1000 questions and 10 variations/question he/she will have to code 10,000 patterns correct? Into this, add the animation, system, etc tags that he/she will have to add, the end result is still the same a massive project, impossible to be done in a resonable amount of time.
Even if the interface of your AIML editor is extremely innovative, a task such as the above is massive. I think you should seriously considering implementing a method to handle the repeative patterns. |
|
|
|
Guile3D
Hooked Member
USA
661 Posts |
Posted - Nov 17 2008 : 18:55:19
|
quote:
If I understand correctly what you saying, each variation for every single question/phrase will have to be coded seperately correct?Let's put some numbers down: In a hypothetical situation where a researcher has a language corpora of 1000 questions and 10 variations/question he/she will have to code 10,000 patterns correct? Into this, add the animation, system, etc tags that he/she will have to add, the end result is still the same a massive project, impossible to be done in a resonable amount of time.
Even if the interface of your AIML editor is extremely innovative, a task such as the above is massive. I think you should seriously considering implementing a method to handle the repeative patterns.
Actually youīll just have to code one "template" for the 1000 or more questions, because youīll be working with wild cards, so youīll have something like:
- What is *
- What do you know about *
- Tell me more about *
- Give me more info about *
then, for all this patterns, depending of the subject, youīll have questions and answers like:
- What is a dog
- A dog is a mammal, usually known as menīs best friend.
- What do you know about Computers
and so on.....
|
Guile 3D Studio
http://www.guile3d.com
"Where reality meets Imagination" |
|
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 17 2008 : 19:21:34
|
That's not exactly what I mean! lets say that we have the following question:
What can you tell me about location 1?
The pattern for this would be something like:
<pattern>what can you tell me about * </pattern>
correct?
But what about if the user will ask
Is it possible to tell me more about
I think you need to tell me more about
I am interested more about location 1, what can you tell me about it?
Unless you have changed the logic of AIML, there is no way for the above pattern alone to answer these questions. You would need to code the patterns for these questions seperately, and connect them with the initial template using srai correct? Although SRAI makes the job easier, what I am saying is that it is impossible for the bot master to think all the possible variations of the same question. What I would like to see is something like:
<pattern> what can you tell me about * </pattern>
The above pattern should be able to answer all the questions containing 1 or all of its words. Something very similar to how google works. This way you can minimize the hassle with AIML and keep the number of questions/phrases in minimal levels.
Is this what you have done with your engine? |
|
|
|
Guile3D
Hooked Member
USA
661 Posts |
Posted - Nov 17 2008 : 19:44:00
|
Kind of... the <Pascal> script gives this power to AIML, but my main concern at this moment is not just with the user input, but to give the bot the ability to learn from itīs conversations, extracting useful and meaningful data from it. On the other hand, if the bot doesnīt understand the input, probably the processed data will not have good relevance.
My goal at this point is to develop an AI engine core that is as easy as AIML, but with much more power so it can remember, learn, do actions and really generate answers from sentences build by itself from previous analysis, and not from hard coded inputs.
If I succeed with this initial stage Iīll already be very happy , and them move to next level, like we used to do playing DOOM! (how I missed that times!!! Still remember all the time Iīve spent playing Wolfenstein 3D!)
|
Guile 3D Studio
http://www.guile3d.com
"Where reality meets Imagination" |
Edited by - Guile3D on Nov 17 2008 20:57:40 |
|
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 17 2008 : 22:08:23
|
Guile
More questions for you:
You are saying:
really generate answers from sentences build by itself
1) Have you considered building whole stories? A nice example would be to analyze the meaning of whole paragraphs from Wikipedia and other sources and build stories dynamically from there. For example, I would like to see Denish scanning the Internet for information on the Pizza tower and build a story from the most credible resources automatically. That would be an amazing step in machine storytelling.
2) If you are going to build sentances dynamically, how are you going to animate denish? Do you perhaps have some kind of gesture lexicon to automatically accosiate the meaning of phrases with proper animations?
Again, well done for this amazing work!!!! |
Edited by - giannis on Nov 17 2008 22:11:05 |
|
|
|
Guile3D
Hooked Member
USA
661 Posts |
Posted - Nov 17 2008 : 22:48:03
|
quote:
1) Have you considered building whole stories? A nice example would be to analyze the meaning of whole paragraphs from Wikipedia and other sources and build stories dynamically from there. For example, I would like to see Denise scanning the Internet for information on the Pizza tower and build a story from the most credible resources automatically. That would be an amazing step in machine storytelling.
2) If you are going to build sentences dynamically, how are you going to animate Denise? Do you perhaps have some kind of gesture lexicon to automatically associate the meaning of phrases with proper animations?
A storiesī builder....this would be really great, wouldnīt it? Iīm not sure I have the ability / knowledge to code that, but itīs sure a nice feature to pursue!
Deniseīs animation and expressions will be trigger by key words from a lexicon / synonym dictionary, or from direct userīs AIML coded tags among words.
|
Guile 3D Studio
http://www.guile3d.com
"Where reality meets Imagination" |
Edited by - Guile3D on Nov 17 2008 22:49:28 |
|
|
|
giannis
Intrigued Member
195 Posts |
Posted - Nov 17 2008 : 23:23:59
|
Yes a storybuilder would be an amazing feature for denish and it shouldn't be that difficult. You can construct a score system simila r to google for rating the content resources on the WWW so Denish will know that a piece of information from resource A has greater value than another piece of information from resource B. Using standard NLP techniques you can then summarize each piece, put it in a meaningfull order and present it to the user. You already have an API for WIKIPEDIA so why don't you take it from there? I would love to see Denish in tourist applications presenting interesting stories about a location.
About the animations how do you deal with the problem of multiple animations playing in a single user's phrase? I mean several of your keywords may be present in a sentance and hence the character may trigger several animations at once. Will the animation A finish before playing before animation B begins?
|
|
|
|
tom_moir
Intrigued Member
New Zealand
144 Posts |
Posted - Nov 17 2008 : 23:25:02
|
What I wasn't sure about Guile is how the information is parsed in the first place. For example if I search wikipedia for cats you could get no end of answers - big cats,wild cats,lions etc! So does Denise ask you for further information when Wikipedia returns say 10 options? Or can you pre-empt things a bit by saying I need to know about Lions.
Thanks
Tom
|
|
|
|
tom_moir
Intrigued Member
New Zealand
144 Posts |
|
|
Topic |
|
Universal Virtual Human Interface (part 2)
