Building Sentences

 

When a person is putting a sentence together, they have a concept of what they want to say – they don’t start with a word, and add another word until the sentence is done. This is also a very bad way to parse a sentence. Far better if we can work from the top down and the bottom up.

With some words (set, run) having many meanings, we can’t be sure we have the correct meaning until the sentence is completely parsed, and the sentence is in a block of text, so sometimes we have to carry uncertainty forward. Some words have multiple parts of speech (“bar” - noun, verb, preposition), and this uncertainty may also need to be carried forward. Some words have multiple parts of speech (“bar” - noun, verb, preposition), and this uncertainty may also need to be carried forward – analysis at the point is suspended until more information is available.

An Example

If we look at the noun, we have ten definitions, ranging from a “rod” – typically a metal bar, to the legal profession – “the bar”. The different meanings are handled through MULTIPARENT operators, which connect to both the noun and the particular noun and a definition. Some nouns have special properties, such as Clausal Nouns, where the noun stands for the following clause: “the assumption that the time of death was after midnight”, and the MULTIPARENT operator requires an extra connection. Many verbs also require this extra connection, as “it amazed John that everything looked so modern” – “amazed” as a verb can be Transitive, TransClausal, TransInfinitive.  There are over one hundred different verb forms – fortunately you don’t have to worry about them as your Unconscious Mind handles all that. The Semantic AI Machine (SAIM) doesn’t have an Unconscious Mind, and does have to worry about them.

Up to now, we have been assuming that the SAIM can resolve the Part Of Speech and definition on first encounter. Not likely. In some large pieces of text, such as a piece of legislation or specification, the SAIM may need to read on another ten or a hundred pages for more information, so the SAIM may need to leave the word in a partially unresolved state (some of the POS removed, some of the definitions removed).

The SAIM prepares itself by reading the heading, subheadings, any Synopsis or Abstract, any glossary. But some words and wordgroups (“a bridge too far”, “catch up with”) may only have meaning when you understand what the piece of text is about. We are assuming the text is without error, but if the piece of text is large, that is probably not the case. The ability to activate the text – to bring it alive – is a powerful tool for finding errors or choosing between definitions, or helping non-expert people to understand what the text is saying, by seeing what it does. . The text may be introducing a new idea – something can’t be dismissed because it hasn’t been seen before – it needs to be reasoned about.

How does a SAIM compare with other methods, such as Deep Learning or LLMs? It doesn’t. People understand the meanings of words and combinations of words in their native language – their limitation (and it is a serious one) is how much of the text they can keep alive in their heads. A SAIM doesn’t have this limitation. The difficulty of how much humans can keep “live” also affects their ability to collaborate, especially on complex projects with many specialties involved, with some abstruse, like stealth paint being washed off in the rain.

LLMs fill a much-needed role for Search Engines, but their inability to synthesise along with their predilection to “hallucinate” (make up stuff) limits their usefulness to “toy” longterm static applications. The change to Semantic AI should strand these other technologies as the dead ends they are (LLMs for Search Engines are fine).

What about programming? For simple tasks, yes. If a program is going up against a nimble and quick-thinking human who knows your system's weaknesses, not so good. An example in the Defence area is where the system sustains damage, and the SAIM must keep as much of the system working as it can, while it attempts to repair or reconfigure what it can.

What comes after Semantic AI? Probably nothing. We will have a common language between human and machine. There may be a place for machine-to-machine transmission, particularly where a simpler machine can carry out a function without needing a high level of plasticity - it gets a dump of a control network rather than the instructions to build it.

Active Semantics website

Semantic AI for Defence website

Four Pieces Limit website