Nathan Hu - Sep 2, 2010:

Hi John, I believe we are talking about different problems. When I consider chatbot, i’m considering a true AI system, not a natural language application. A commands based natural language interface may easy to practise, but it can never lead to a true AI system.

Interesting.  Having complete NLP ability I think, is certainly a measure of intelligence.  Ability to learn via interactive NLP discussion, in my opinion would qualify as not only A.I., but A.G.I.

I agree. Actually, the induction ability is for learning.

Victor Shulist - Sep 2, 2010:

Nathan Hu - Sep 2, 2010:

Hi John, I believe we are talking about different problems. When I consider chatbot, i’m considering a true AI system, not a natural language application. A commands based natural language interface may easy to practise, but it can never lead to a true AI system.

Interesting.  Having complete NLP ability I think, is certainly a measure of intelligence.  Ability to learn via interactive NLP discussion, in my opinion would qualify as not only A.I., but A.G.I.

John,

Or shall we design a lite version of English called “Concise English” and force everybody to use it with computer?

I like this…however I have a slightly different take.

Give the written medium (books, magazines, blogs, forums, online articles, etc.) we find that there are a lot of ways and a lot of words used to express very similar and basic meanings.  The actual meaning…which may be described by concise English usage…is buried and wrapped inside lots of English.

For example, how many different ways have you read or written this ‘Concise Meaning’ (in lieu of concise English) in a forum, blog, or article ?

  I do not like .......

After we strip away the vocabulary usage, poor grammar, poor spellings, emotion, over/under emphasis…we end up with the core meaning that ‘someone’ does not like ‘someone/something’.  So, we cannot count on a “Concise English” but must seek to derive the “Concise Meaning” regardless of how it may be expressed. 

Great discussion!

Regards,
Chuck

Chuck,

This is why , with CLUES, it does not respond to the user input string, but instead to the meaning.

user input is taken, parse trees are generated.

Then, what I call ‘concept specifications’ are applied.  There is a “one to many” mapping between a specification of a concept and many parse trees.

The parse trees may have different structures (syntax) and even have different words, but that one-to-many map gets converted to a concept name.

Then, there is a reactor script, one for each concept name.

The reactor’s job is to evaluate the concept of what you said with concepts already stated by you and the bot , combining as many statements in the current conversation as possible, along with concepts recorded from statements in its knowledge base.

If we have more than one reactor that produces a response, the reactor that employed the most KB and conversation history facts, wins, since it is deemed more relevant.  If all reactors used the same number of KB and conversation statements, then the engine simply picks randomly among them.

“If-then” rules, rather than being hard coded in the bot, are taken from parse trees of complex sentences…. this is all future functionality (well the ‘concept specs’ are working now), but N.L.R. (naturual language rules) are on the road map

Nathan, you’re right, I mean a NLP software, not a true AI system, when I talk about chatbot. This is a good example of misunderstanding and clarification between human adults.  But “Concise English” is more than a command set, which is just a practical start point in my mind. Chuck got the idea, “Concise English” is the NLP representation of the basic and core meanings of NL, which can be represented in other forms as well, so it does lead to true AI.

Chuck, I agree that “Concise Meaning” is a good term in consolidating all the different ways to express the same meaning. My term of “Concise English” means “speaking and writing English in a concise way”, which is more practical in manipulating a machine or a database, but for human yes we should use “Concise Meaning” since they’re good at making things complex and subtle by using “concise” words.

Victor, good design on list and groups, and thanks for splitting the thread. My initial posting was only about LP tricks, the title of “Challenging issues in NLP” might be too big. But since the thread has this title I’d like to discuss another issue I found very challenging: the representation of meanings of a word. Here’s a good example,

Are you one of those people who are so busy getting ready for a trip that they forget some of the most important things? Then suddenly somebody says, “What about our ...?”  (from a German non-fiction by Monika Wegler, translated by Rita and Robert Kimber)

The first sentence doesn’t request any information from you though it comes as a question, while the second sentence with an embedded question is describing a consequence by an imagined yet very real event. Now how to associate this kind of expression to its true meaning? FYI I myself have no idea. :-(

Hi Victor, It seems we are on the similar way. The difference is that I use a semantic sub-network instead of your concept name.

Currenttly, I focus on automated discovery of if-then rules from text which I called induction. I can also acquire the if-then rules from some simple sentence like “a person is a student if the person is a child.”.  You may handle more complex sentances. I believe you are leading in grammar (syntax) functionality.

Victor Shulist - Sep 3, 2010:

Chuck,

This is why , with CLUES, it does not respond to the user input string, but instead to the meaning.

user input is taken, parse trees are generated.

Then, what I call ‘concept specifications’ are applied.  There is a “one to many” mapping between a specification of a concept and many parse trees.

The parse trees may have different structures (syntax) and even have different words, but that one-to-many map gets converted to a concept name.

Then, there is a reactor script, one for each concept name.

The reactor’s job is to evaluate the concept of what you said with concepts already stated by you and the bot , combining as many statements in the current conversation as possible, along with concepts recorded from statements in its knowledge base.

If we have more than one reactor that produces a response, the reactor that employed the most KB and conversation history facts, wins, since it is deemed more relevant.  If all reactors used the same number of KB and conversation statements, then the engine simply picks randomly among them.

“If-then” rules, rather than being hard coded in the bot, are taken from parse trees of complex sentences…. this is all future functionality (well the ‘concept specs’ are working now), but N.L.R. (naturual language rules) are on the road map

Nathan

Yes, From the very beginning, I have based the design of my bot engine 100% on grammar

Well, the “foundation” is on grammar anyway, then concepts are on top of that.

From the questions you are asking, it leads me to believe you must be very well on your way with your design.

Do you have a thread for your bot ?  If not, you should create one, I would be interested in seeing some sample conversations, even if simple.  Or are you at that point yet?  I"m not quite there myself, I’m thinking well before the new year, I will have some samples up and perhaps a video.

I will show some sample later. I did not complete the expression moudle yet.
There are tools to parse text into trees. But I did not found any existing tools to express trees to text.
I may build a simple one by myself later.

John Li - Sep 1, 2010:

Some tricks used in LP, as I remembered remotely, are something like “which is bigger, Belgium or apple?” and “which is nearer, the Queen or the Queens street?” and “If we shake hands then what is in your right hand?”

These are good and reasonable tricks. Yet in my own design and development of chatbot I found Quotation is very difficult, with or without quotation marks.

This is a small point, but it matters:

the questions I think you mean:

“which is bigger, Belgium or an apple?”

That “an” IS important.  The way your sentence reads (”“which is bigger, Belgium or apple?”—I’m thinking what apple ... the big apple (nick name for new york), or what”.  “an” tells me you mean any apple in general.

also, you do not need the word “the” in..

“which is nearer, the Queen or the Queens street?”

should be just

“which is nearer, the Queen or Queens street?”

Now

“If we shake hands then what is in your right hand?”

That isn’t as much of an NLP problem as it is a knowledge problem.

The bot will simply needs HUGE amounts of knowledge to answer any random question like that. 

it will also have to know , or probably have a “simulation” or model of the physics of people shaking hands… VERY GOOD example !!

The sentence should really be

“If we shake hands, what is in your right hand?”

you don’t need the “then”

The example are good because the bot will have to be forgiving of the grammar—I will use these examples when teaching my bot

Victor Shulist - Feb 8, 2011:

...also, you do not need the word “the” in..

“which is nearer, the Queen or the Queens street?”

should be just

“which is nearer, the Queen or Queens street?”

That, my dear Sir, depends on whether you’re referring to Queens Street, over in town, or “THE Queen’s street, near the castle. Depending on the context , either could be perfectly valid.

Ahhhh.. you see !!  You see!!  so that “the” DOES matter!  “the” and “a”  (definite and indefinite articles) matter   Changes the semantics for sure.