Can machines understand science?

I am taking a new approach to blogging by dictating my thoughts to the machine.

On Friday I am giving a talk called “Can machines understand science?”.  I will argue that the combination of technology and information now allows us to communicate with machines and for machines to communicate with us.  This does not cover the whole domain of human activity but in limited areas such as formal aspects of chemistry machines can perform as well as many human beings.

John Searle devised a thought experiment of the” Chinese room” where a human who did not understand Chinese was in a room.  The person received Chinese characters which they looked up in a large book of instructions which told them how to react to the characters and how to transmit an appropriate output.  If the book of instructions was sufficiently large the machine might output instructions that appeared to be an intelligent response to the input.  If this could be done such that a human observing the process from outside could not tell whether there was a human or a machine inside, Searle asserts that the machine+human “understands” Chinese.

I believe that we are in the same position with some areas of chemistry.  Machines can carry out tasks in a way that cannot be distinguished from humans doing the same task.  For example I believe that a machine can answer some chemistry questions on exam papers as competently as a human.  These will fall into a number of categories such as

·         regurgitating rote to learning

·          carrying out certain algorithms

·         and looking up rules or data in formal procedures.

For example we have now written a system OPSIN which can translate IUPAC nomenclature into chemical diagrams.  Daniel Lowe has now achieved a very high success rate for organic compounds with over 95% conversion and virtually no errors.  This is clearly considerably better than a first year undergraduate who only has a limited chemical vocabulary.

Of course the machine has to read the examination paper and if this is on printed paper this is a technical problem (although current scanning will probably allow most of the essential material to be captured).  There will be ambiguities especially for short lines and dots and the system has to be able to make a reasonable guess [1].  If the paper was available in ASCII or xml then it would be possible to read it without errors.  Having done that the machine has to understand the language in the question but here we are fortunate in that most exam questions are phrased in very formal language and it is fairly easy to apply language processing techniques to understand them.  As a result I assert that a machine could answer an organic nomenclature question on an exam paper.

This is of course only one small sub domain of chemistry and a fairly unusual one in that it consists of a very large number of fairly well explained rules.  There are other parts of chemistry which also have formal rules such as balancing equations or predicting the outcome of well described reactions such as in elementary organic chemistry. In these I believe that machines can do as well as student chemists.

You will argue that this is a very small part of what chemistry is about but we have been able to take it considerably further.  Our machines can now understand parts of research papers such as the experimental recipes.  This is a valuable process in that the machines can now reach the literature much faster than humans.  For example we can read a 300 page patent in a minute or two and understand much of the chemistry. This has been automated to scale to the weekly output of patent offices.

This blog post was dictated.  It is a very different process from I think the style is rather stilted.  However with practice I expect to be able to dictate my thoughts to the machine at least as fast as I can type them.  Stop to think what the machine has actually done in transcribing my audible noises into meaningful English language sentences.  There is a lot that has to be done by the machine but that emphasises how to considerable the advances have been in the last 30 years.

[1] Rant: and of course PDF isn’t much better.