Introduction to Natural Language Understanding
One of the earliest efforts in AI was to develop systems that could understand natural language; e.g. the Eliza system
(incorporating structured and opportunistic approaches)
Well-respected researchers stated that machines would be able to understand natural language within 15 years
Sections 6.3, 7.1,7.2, Chapter 15 (omitting 15.3,15.4)
These efforts (obviously!) did not meet with success
Natural Language Understanding
Example
The difficulties of dealing with context and breadth were grossly underestimated
The British Medical Journal did a study in 2014 translating 10 medical phrases into 26 different languages
Recently, NLU systems have been more successful – Google Translate and similar systems are useful for simple sentences in many languages
‘There were some serious errors. For instance, “Your child is fitting [note, more UK oriented verb for having a siezure]” translated in Swahili to “Your child is dead.” In Polish “Your husband has the opportunity to donate his organs” translated to “Your husband can donate his tools.” In Marathi “Your husband had a cardiac arrest” translated to “Your husband had an imprisonment of heart.” “Your wife needs to be ventilated” in Bengali translated to “Your wife wind movement needed.’
They tend to do more poorly as they require knowledge of the context of the conversation, as they tend not to have this knowledge
• Operating on a surface level (statistical) model of language
3 © 2020 J.E. Anderson
http://www.bmj.com/content/349/bmj.g7392
4 © 2020 J.E. Anderson
Natural Language Understanding
2 © 2020 J.E. Anderson
Understanding vs. Generation
NLU vs. Speech Understanding
Natural language text generation is a simpler task – you can control the complexity of the sentence and keep it simple
We tend to think of speech and language as being very similar, just because speaking and listening are natural things to us
e.g. generate an explanation in English of the rules that were used to solve a problem in an expert system
But actually understanding speech as opposed to reading text is a much more significant problem
Many shells incorporate elements of text that can then be tacked together to provide longer explanations
Really we’re adding perception to the reasoning problems already inherent in NLU
• Well suited to backward reasoning where the path to the goal forms a structure for explaining why (stepping back to facts)
We’ve seen how difficult perception can be in many robotics examples!
Speech Understanding
Tonal Differences
We’ve also seen that context plays a huge role in perception (e.g. knowing we’re in a hallway makes a robot look for a second wall where it’s seen one already)
The way something is said can have a huge affect on its meaning in most languages
Similarly, much of understanding speech in a noisy environment is about using context to fill in gaps where what is heard is uncertain
e.g. the sentence “I never said she stole my money” conveys 7 different meanings depending on which word is stressed
We all have the experience of saying “what?” to somebody and then before they even answer, realize what they were talking about
Acting classes often use exercises where participants are given only neutral dialog to utter, and must convey all necessary information using tone/gesture
56 © 2020 J.E. Anderson © 2020 J.E. Anderson
e.g. Sarcasm – difficult to follow in print
78 © 2020 J.E. Anderson © 2020 J.E. Anderson
Natural Language Understanding
The goal of an NLU system is to understand a collection of words
The words are presented in written form (i.e. machine readable: each word is a separate unit, punctuation is included). There is no doubt what each word is.
The system produces a symbolic description of the meaning of the words
Understanding is association with known (symbolic) concepts – that’s usually what we mean when WE say we understand something
9 © 2020 J.E. Anderson