With all the grammar-oriented design going around here, I’ve started researching other ‘systems’ for language analysis and construction, especially anything that can be linked to my conceptual model of describing knowledge. One school of thought I came across is SFL:

From the link below:

Systemic-Functional Linguistics (SFL) is a theory of language centred around the notion of language function. While SFL accounts for the syntactic structure of language, it places the function of language as central (what language does, and how it does it), in preference to more structural approaches, which place the elements of language and their combinations as central. SFL starts at social context, and looks at how language both acts upon, and is constrained by, this social context.

http://www.isfla.org/Systemics/definition.html

Thoughts?

For some reason, this post sent me back to structuralist de Saussure’s “Cours de linguistique générale” (available online in English translation at http://www.archive.org/stream/courseingenerall00saus#page/14/mode/1up), which says on page 14:

In separating language from speaking we are at the same time separating: (1) what is social from what is individual; and (2) what is essential from what is accessory and more or less accidental.

Language is not a function of the speaker; it is a product that is passively assimilated by the individual. It never requires premeditation, and reflection enters in only for the purpose of classification, which we shall take up later (pp. 122ff).

Speaking, on the contrary, is an individual act. It is wilful and intellectual.

It’s interesting that the software by Systemic-Functional Linguists is all generative, on the speaker side. It seems to me that my bots try to assimilate language passively, as Saussure says, and to make them more individual and “intelligent” they need to develop the generation side - ask questions, develop hypotheses and test them, etc. (MegaHAL generates language, albeit in a rather random fashion - although sometimes it comes up with surprisingly appropriate and funny responses.) While the Natural Language Generation applications cited in one of the papers from the site linked to are things like generating weather reports from databases, it is interesting that they focus on the speaker side, which according to de Saussure is the part of linguistics allowing the individual the most freedom from social influences…

It’s interesting that the software by Systemic-Functional Linguists is all generative, on the speaker side.

There are real problems parsing with functionally rich grammars. The parsers you can download off the web, and work, generally are fairly minimal in regards to categories per unit. Systemic grammar uses three layers of function structure for each clause, and a lot of features (classes) for each clause as well. This really slows down parsing, and if one wants to use statistical parsing, the size of the training corpus needs to be far higher, and the amount of manual tagging for each sentence is also more extensive.

So yes, parsers using PSG or x-bar syntax are common.

There has been several systemic grammar parsers, but yes, nothing released publicly.
I led a team which developed a wide coverage syntactic parser for English, which is still in use in a medical informatics company in the States. However, to get it working, we threw out most of the “functional” basis of the grammar (used a dependency version of the ‘interpersonal’ layer of the systemics.)

  it is interesting that they focus on the speaker side, which according to de Saussure is the part of linguistics allowing the individual the most freedom from social influences…

I don’t buy that. Firstly, every text is embedded in a social context, and we use our
knowledge of context-language appropriateness to both produce texts/speech which
is functional in its context, and to interpret texts/speech in its context.

In generation, one can “hardwire” the context into your program (fix the range of
linguistic choices which are available to express particular meanings), and equally in parsing one can ignore any notion of social context, just treat all language as text.

But a good text generation system has a configurable context (including speaker model
and addressee model), so that what is selected to be spoken (content selection),
and how that content is expressed, can be programmatically varied.

Equally so, knowledge of the context of an interaction can help resolve
questions of meaning in parsing (word sense disambiguation, reference resolution,
etc.). For instance, in a dialogic situation such as a chatbot, knowledge of what
range of meanings typically appear in a given conversational setting help to interpret
what is said. Systems can match the predicted meanings in context against the set
of words produced by the human to “template fill”. etc. This is more true for virtual
assistants, where the conversation has goals, and thus predictable sub-steps
and sequencing. But not for general Chatbots whose goal is just to chat. But
this is not my field, so I may be wrong.

Mick

This really slows down parsing, and if one wants to use statistical parsing, the size of the training corpus needs to be far higher, and the amount of manual tagging for each sentence is also more extensive.

Interesting. Any idea on numbers, like parse speed or training corpus size?

Jan Bogaerts - Mar 6, 2011:

This really slows down parsing, and if one wants to use statistical parsing, the size of the training corpus needs to be far higher, and the amount of manual tagging for each sentence is also more extensive.

Interesting. Any idea on numbers, like parse speed or training corpus size?

Not off hand, but in an SFG grammar, clauses are assigned function labels from three layers
(an NP in a clause could be Subject/Actor/Theme, the NP after the verb Complement/Goal, etc.)
There are lots of details in the grammar to get these different components of the description to work together.

In terms of features, in the NIGEL grammar, the most comprehensive SFL grammar, a clause would typically have 50-60 features assigned to it, each representing some distinct aspect of grammatical variation (e.g., verbal-process:finite-clause:modal-clause:passive-clause:perfect-aspect:declarative:no-object-fronting:not-progressive etc.

Statistical parsers are good at dealing with a small number of categories per unit,
e.g., 1 (NP, Ven, etc.), up to 3 (function label, part of speech, inflection class).
Vastly increase the number of categories, and the amount of input you need
to train a parser increases.

Mick

@Mick: I just took a look at your other website. Looks like you’ve got lots of info on the subject. I’m making a mental note to go through all the info after the Chatterbox challenge. Looks like a serious read indeed.

Hello, Mick, and welcome to chatbots.org!

Oh, boy! More stuff to read! Right now, I don’t have time to do more than skim through the info, but I fully intend on revisiting this as soon as things slow down for me. Once I do, I’m sure I’ll have some questions/comments to make.

Mick ODonnell - Mar 6, 2011:

every text is embedded in a social context, and we use our knowledge of context-language appropriateness to both produce texts/speech which is functional in its context, and to interpret texts/speech in its context.

This might be true for some or most of us, but is it necessarily true for everyone? For example, take a text in a dead language, say Euclid’s “Elements”. The goal is clearly to express truths based on explicit axioms; the social context is irrelevant to that purpose.

Another example: Turing acknowledged a homosexual relationship to police investigating a burglary of his home. Was the social context (the illegality of homosexuality at the time) a factor to him when he answered the police’s questions truthfully? Should he have lied, or should the social context have adapted to him?

In generation, one can “hardwire” the context into your program (fix the range of linguistic choices which are available to express particular meanings), and equally in parsing one can ignore any notion of social context, just treat all language as text.

But a good text generation system has a configurable context (including speaker model and addressee model), so that what is selected to be spoken (content selection), and how that content is expressed, can be programmatically varied.

Sure, and I can see why translation is a focus of the available SFL software. Changing your speech to fit the social context can be a useful tool, but it is always our (or the software’s) choice whether to use that tool, or to aim for expression of a contextless, timeless truth that transcends fickle, arbitrary human social conventions…

Everyone thanks for the input and insights so far, keep ‘m coming

Where I am right now is that I’m trying to build my own functional linguistic model that will fit the rest of my mind-model and can interact with it without any translation steps. So I’m looking for a linguistic model that can be described in concepts and SFL seems to lean that way. So far I have several concepts that are seen in SFL like action, actor, subject, but I also have ‘assigner’ and ‘value’, and I didn’t find that in SFL so far (but I still have quite some stuff to read in regard to SFL as well). Also, my mind-model incorporates tagging as being driven by a learning system based on experience, so there is no manual tagging involved (aside from the base-concepts that represent instinct and other ‘old-brain’ models).

Mick ODonnell - Mar 6, 2011:

But a good text generation system has a configurable context (including speaker model and addressee model), so that what is selected to be spoken (content selection), and how that content is expressed, can be programmatically varied.

In my mind-model the context is continually shifting and adjusting based on contextual mapping of the conversation. This means that both the user and the AI can steer on focusing or shifting the conversation by intention or by leading.

So far I have several concepts that are seen in SFL like action, actor, subject, but I also have ‘assigner’ and ‘value’, and I didn’t find that in SFL so far (but I still have quite some stuff to read in regard to SFL as well).

Can you explain the functionality of those tags?

Also, my mind-model incorporates tagging as being driven by a learning system based on experience, so there is no manual tagging involved (aside from the base-concepts that represent instinct and other ‘old-brain’ models).

Well, in order to learn, the system is going to need lessons. And that means manual tagging, alas.

C R Hunt - Mar 7, 2011:

So far I have several concepts that are seen in SFL like action, actor, subject, but I also have ‘assigner’ and ‘value’, and I didn’t find that in SFL so far (but I still have quite some stuff to read in regard to SFL as well).

Can you explain the functionality of those tags?

Not yet, as I’m still ironing out this part of the model

C R Hunt - Mar 7, 2011:

Also, my mind-model incorporates tagging as being driven by a learning system based on experience, so there is no manual tagging involved (aside from the base-concepts that represent instinct and other ‘old-brain’ models).

Well, in order to learn, the system is going to need lessons. And that means manual tagging, alas.

Not really. My system uses ‘tags’ to store experiences in relation to ‘concepts’. The system will create and manage those tags by itself as it’s an intrinsic part of my mind-model. The system will learn by experience and do the tagging by itself based on that. So no ‘manual tagging’ is involved.

I’m pretty sure that what I’m doing with ‘tags’ is different from the ‘tagging’ that you have in mind