What AI and Bots can still learn from Hemingway

All of the current key players in Artificial Intelligence (AI) like Google, Microsoft, Amazon, Facebook, IBM, Apple etc. are also working in the field of natural language processing and understanding. The reason is obvious: all these players have a strong motivation to do so as they provide services and/or products that increasingly use common language as an interface to their billions of clients around the globe. Amazon's Echo device and the virtual assistant ALEXA, Apple's SIRI, Microsoft's Cortana, Google's Now and hundreds of chat bots are examples of this trend.

They all work on the assumption that machine "understanding" of text, messages and human conversations is possible and close by just using the standard techniques and current knowledge and methods of common NLP (natural language processing), Deep Learning and Neural Networks, Bayesian and statistical models, hidden Markov models and some other less known linguistic models.

There is no doubt that AI has made some substantial progress in machine "understanding" of text and spoken language especially in the fields of text-to-speech, speech-to-text and simple one-command-to-action processing (see for example ALEXA and SIRI).

However, in the more demanding areas such as chat bots where free conversations should be possible and standard there is pretty much no progress to be seen since the last 30 years or so. Microsoft's infamous chat bot TAY's release and its dumping within 24 hours is the best example and proof for this. It shows that real natural language understanding is an extremely complex process and will not be resolved within the near future. Why is that so ?

It's because machine language "understanding" requires understanding humans and human culture, not just their language !

A very telling example for this is a short story credited to Ernest Hemingway. The folklore has it that Hemingway once made a bet with his buddies about who could write the shortest story. Hemingway won the bet with the following famous 6 word short story:

"For Sale: Baby shoes, never worn"

Now, we humans understand this story and quickly get the gist of it and even many implied clues. We understand that the author's story is actually a very sad, if not depressing story told in just a few cold words. The author uses the form of a straight forward, unemotional advertisement to tell something rather dramatically different.

We understand that the key message of the writer here is NOT that somebody wants to sell baby shoes. Then it would just be an advertisement nobody would bother reading.

The implied scenario of the author here is that the baby mentioned probably died before it could ever wear the shoes, but this is left un-said on purpose and it is left to the reader to figure out what might have actually happened that somebody makes a sales post like that.

It gets depressing when one starts to think about the story and what might have happened in more detail: The 'unworn" word implies that the baby died. It's unlikely that one sells baby shoes because one doesn't like them anymore, especially if they are really unworn. But why would any human ever want to sell the shoes of their baby that probably just died ?

Wouldn't parents rather want to keep them as a memory ? So, the story even implies that the parents (or parent) needed to sell the shoes probably due to financial issues. If so, these must be serious issues if the sale of baby shoes and the little money usually expected from it would help the financial situation. And why would the parents actually mention that the shoes are unworn ? Doesn't it imply that they want to achieve a better price for the shoes ? How cruel would that be ?

Or is the baby not dead but rather given away for adoption ? Then the sale and the "unworn" part would make more sense and would not be so cruel.

Or is it rather that the baby was born dead and hence could never wear the shoes ? But in this case it must have been a late death during pregnancy or right after birth before kids start wearing shoes as the parent(s) already bought the shoes in advance expecting the baby to wear them. They would probably not have bought shoes right after they found out about the inception.

Or do the parents just wanted to get rid of the shoes and the associated memories as it would hurt them more to see them and remind them of the child's death and they have no financial issues ? But if that were the case, why would they not just dispose them and just throw them away ?

Many of these kind of thoughts come up when we read this story. The interesting thing here is that nothing we think about after reading the story is actually mentioned anywhere in the story. No parents, no death of a child, no adoption, no financial issues, no motivations and intentions of the parents etc !

As a matter of fact, the story is such a genius piece of writing of Hemingway just because it does NOT mention any of these potential background issues but puts the story in such a way that the reader is kind of forced to create the potential background and motivations behind the story after reading it in his own mind !

It's what the author forced upon us by phrasing the short story in these simple 6 words that makes the story to what it is - a literature masterpiece.

What the story does NOT tell is the real story !

It's because we read this story as a piece of literature and not as an advertisement that changes all the meanings of the words and the content of the story for us! If it were just an advertisement for shoes, we would not bother to think about it for a minute - unless we were looking for baby shoes.

Now put yourself in the shoes (pun intended) of an AI robot that is supposed to read and "understand" this short story ! How could it ever figure out anything we just discussed ? How could it ever not read Hemingway's story as an advertisement (assuming it would not yet have been published on the web or anywhere like here) ?

It could only "understand" the story if it had any understanding of human values, goals, intentions and cultures, if it would understand what authors are trying to do with their readers and how people behave in certain situations and how they would cope with a child's death for example.

None of the current popular AI technologies would have a even a chance to figure this out: no Deep Learning net, no Bayesian logic, no NLP can help here - and neither can any combination of them !

The only approach that could partially (!) solve this issue is an "old" AI technique that very few AI researchers are still aware of that uses so called Scripts invented by Schank and Abelson already in the 1970s. It works on the idea of building schemes as blue-prints for common situations as models for background information to understand language. Its also a technology for planning complex tasks.

Speaking Lions

A defender of current AI would maybe argue: this Hemingway story is a constructed and very special and rare case of text and not like usual communication in messages, emails or "normal" human conversations etc.

This could not be further from the truth !

A lot of information and meaning altering effects in every day communications are hidden in the complex background of our cultures and our ways of life. Most often we do not explicitly say what we actually mean and what we want to say.

Just read these sentences and think carefully about what is actually and literally said and what is the real intention and meaning behind them:

• Nice car ! But too expensive.
• I hate this politician.
• I love you.

The famous philosopher Ludwig Wittgenstein once said:

"If a lion would speak to us, we would not understand him".

He was not implying that we would not be able to understand the sounds the lion generated or the words he would utter...but they would just make no sense to us because a speaking lion would not fit into our cultural background at all. The best we could make out of this would probably be that it would be a trick performance or a hallucination or some other reason why we would think the lion was actually talking to us (maybe some sophisticated hologram that looks like a lion etc...).

We would not understand how a lion could talk and have a casual chat with us over a cup of coffee or a nice human bone. A speaking lion would destroy most of our scientific assumptions about lions and humans and completely destroy our current world view.

Today's AI researchers need to understand that machine language "understanding" is still just like lion's talk...

E.Schoneburg

Hong Kong, March 10, 2017