#SASSY

This project is going to be very much exploratory programming. I have no idea if this is going to work. It will be fun.

The program will need something to chew on. I think "example data" would be a better term than "test data". Testing is the process of confirming a program coforms to its design. At this stage we have no design.

I have chosen five paragraphs from various design documents. They include a paragraph from a preliminary design, a functional requirement, a use case, an architectural tactic and a component description.

The input to the process is an RDF model that represents the information in the paragraph. This is intended to simulate an extract from what will be in SASSY's knowledge database.

I am currently constucting these models.

#SoftwareArchitecture
#NLG #NaturalLanguageGeneration

#SASSY

There is one classification of the RST reationships that we can use. They can be either structural or semantic. The structural relationships will come from the overall design of the document. The semantic ones come from the data properties - i.e. the predicates of the RDF statements.

In practice this means that the sorted order for the RDF statements within a paragraph is based on the semantics.

This SASSY subproject aims to take lists of RDF statements and sort each list based on the semantic ranks. The RDF statements will then be agregated into sentence sized chunks.

The next step is the internal organisation of the sentences, such as adjectival phrases and subordinate clauses.

An interesting step is to identify the flow of focus and use that to replace nouns (Fred) with pronouns (he) or noun phrases (the boy). For a design document it is especially important not to introduce any ambiguity.

The final step is to create a grammar tree for each sentence and pass it to the NLG program.

#SoftwareArchitecture
#NLG #NaturalLanguageGeneration

#SASSY

The core observation of Rhetorical Structure Theory (RST) is that text can be recursively subdivided into two parts, termed the "nucleus" and the "satellite". The nucleus can be understood by itself, whereas the satellite depends on the nucleus.

RST researchers spent a lot of time trying to classify the relationships between the nucleus and satellite. Not much agreement was reached.

Their approach to text generation was to select a communicative goal which would indicate a general plan. The plan would suggest relationships and the database would then be searched for matching data.

For SASSY the communicative goal is simply "tell us everything you know". The planning then collapses to a sorting process.

While playing around with ways to sort the data it became apparent that the actual set of RST relationships was not the important aspect. What was driving the algorithm was the ranking of the relationships. This is why the researchers could never agree on what relationships should be used - it doesn't matter.

For an RDF knowledge database I think this means that I just need to assign a "semantic rank" to each property. We shall see ...

#SoftwareArchitecture
#NLG #NaturalLanguageGeneration