Classifying AAC symbols for ease of use (Displayed in en-GB)

Questions around the classification of symbols becomes even more important when considering a range of languages which have little in commone. In particular if they are not part of the European group of languages.

Issues around orthography and the changes that diacritics can make to text to speech output. Or left to right and right to left placement if one is thinking about moving through a sentence or phrases with a minimal number of selections and distance to cover.

In the past AAC symbols have often been classified their use:

• Aided vs. Unaided: Whether the symbol requires an external device or instrument (aided, like a picture on a board) or is produced by the body (unaided, like a gesture).
• Static vs. Dynamic: Whether the symbol is motionless (static, like a printed word or picture) or changes over time (dynamic, like a gesture or animation).
• Iconic vs. Opaque: How much the symbol visually resembles its referent (iconic symbols are readily guessable, opaque ones are not).
• Set vs. System: Whether the symbol collection has internal generative rules for creating new symbols (system) or is a fixed list (set). 

Pampoulou and Fuller proposed The Multidimensional Quaternary Symbol Continuum (MQSC) 

• Non-linguistic Symbols: These have little to no inherent internal logic and virtually non-existent expansion capabilities. Any new symbols must be provided by the developer (e.g., a simple, fixed set of pictograms).
• Pre-linguistic Symbols: These have a rudimentary internal logic, allowing for some limited expansion or modification (e.g., some symbol sets where simple rules apply).
• Linguistic Symbols: These possess a sophisticated internal logic (generative rules), enabling the user to create virtually unlimited new symbols and express a wide range of ideas and thoughts (e.g., Blissymbols, which have an inherent logic allowing for complex meaning generation). 

But what if we needed to classify symbols by topic, part of speech and other features such as culture, language and core or fringe usage.  Perhaps with the support of AI models this could be done more easily as all previous classifications have mainly been used for research and had to be carried out manually - added to a database for filtering.     Large Language Models and visual recognition  might come into play although clearly abstract symbol concepts and a lack of well labelled training data could be an issue.