I think that artificial intelligence is not very close to being able to understand such stories in a genuine way. Therefore, I would like to sneak up on it gradually by dividing the problem into parts which can be attacked separately. Here are some of the components:
1. A formalism capable of expressing the assertions of the sentences free from dependence on the grammar of the English language. A good test for such a formalism would be to produce a program for translating from the formalism into any of several natural languages. More weakly, it should be as easy for a human to translate from the formalism into a natural language as to translate from one known natural language to another. Let's call this formalism an artificial natural language--ANL for short.
The grammar of ANL should be trivial and mathematical in character. There would be an ``English'' version in which English words were used as identifiers, but there would still have to be a glossary that gives the precise meaning of the identifiers. There would also be a German and a Japanese version. The translation from the English version to the German or Japanese version would be a simple substitution for identifiers, and a German or Japanese who had learned the grammar could then translate into his language with the aid of the German or Japanese glossary.
This idea has some resemblance to the idea of ``deep structure,'' but I have some doubts about whether either idea is well enough defined to say.
2. A data structure for expressing the facts (apart from expressing the sentences). In such a data structure, it would be definite which robber pushed Mr. Hug first, and what the robbers said even though it is not stated in the story. Clearly some compromise is necessary here, since the data structure need not be able to express positions and velocities of molecules. Like the PLANNER languages, as Robert Moore has characterized them in his 1976 MIT Master's thesis, the descriptions would contain no disjunctions, and might be a collection of relations with constants as arguments where every relation not asserted (in a certain class) is automatically denied.
Alternatively, the basis of this data structure might be various networks of nodes described by sentences in the predicate calculus. Some of the sentences would assert that certain programs applied to the data structures would answer certain questions. When such sentences existed, reasoning would include the operation of the programs. In this way, we would expect to avoid the extreme prolixity that arises when we attempt to do even simple calculations by pure predicate calculus deduction.
The test of success for the ``data structure'' would be that a human could readily formally deduce the answers to the above questions using a proof checker. Most of the proof-checker would be straightforward, but there is a major problem concerned with when it is possible to ``jump to a conclusion.''
3. I see each of the following problems as a difficult AI problem:
a. A ``parser'' that takes English into ANL.
b. An ``understander'' that constructs the ``facts'' from a text in the ANL.
c. Expression of the ``general information'' about the world that could allow getting the answers to the questions by formal reasoning from the ``facts'' and the ``general information.'' The ``general information'' would also contain non-sentence data structures and procedures, but the sentences would tell what goals can be achieved by running the procedures. In this way, we would get the best of the sentential and procedural representations of knowledge.
d. A ``problem solver'' that could answer the above questions on the basis of the ``facts.'' We imagine the questions to be expressed in the ``fact'' language and expect the answers in the ``fact'' language, i.e. we avoid grammar problems in both understanding the questions and in expressing the answers.