Notes for 26-Sep-2001.

Review

Content

Syntax vs. Semantics

Syntax alone does not capture meaning. Here, we see that the same parse tree can deliver two sentences that are very different in meaning. In one sentence the subject ("father"), plays the role of agent; in the other sentence, the subject ("the pie") plays the role of goal. The real agent in the second sentence is unseen: presumably an oven is baking the pie. A similar comparison b/t subject roles can be noted in the sentences "Bob rolled the ball down the hill." and "The ball rolled down the hill."

Ambiguity

Another example: consider the sentence "Time flies like an arrow.", which may be interpreted in five ways:

  1. Time passes in the same way an arrow passes (i.e., time goes quickly).
  2. You clock flies in the same way you clock arrows.
  3. You clock flies in the same way an arrow clocks flies.
  4. You clock flies that (look) like an arrow.
  5. Flies (specifically, time flies) prefer arrows.

(Stolen from Stephen Pinker's The Language Instinct.)

What can we do about ambiguity?

Try to disambiguate by looking at the world.
We may be able to eliminate competing interpretations by looking at the world. This is a powerful technique in practice, but it does not guarantee a solution. And it can only be used if we have some ability to inspect the world.
Avoid it.
For some domains and situations, it may not be necessary to disambiguate: prehaps their is no ambiguity to begin with, or prehaps the syntactic analysis can make reasonable guesses (as we do by assigning precedence levels to resolve shift-reduce conflicts while designing parsers for computer languages).
Use common sense.
Many believe that common sense is what helps human disambiguate; some researchers have built large databases to provide computers with a common sense understanding. (This was not discussed in class, but what's a list without three items in it?)

Winograd's SHRDLU program deals with disambiguation and the blocks problem. Here is a sample dialog with SHRDLU.

Case Grammar

We might choose to abandon phrase-structure grammar and use a case grammar instead. Consider the sentences:

  1. John gave the book to Mary.
  2. Mary took the book from John.

These sentences have essentially the same meaning (ignoring subtler implications of the verbs "gave" and "took"), but their syntax is different. In case grammars, we augment the lexical entries of verbs with templates or subcategorization frames. The template is filled by the words surrounding the verb.

Notice how both sentences will conclude that {agent="John", benificiary="Mary", object="Book"}.

Case grammars could be a basis for A-TRANS (more about this in the next lecture).

We might also think of Computer Vision in terms of case grammars: for instance, a "support" (e.g., chair) could be characterized as the template {seat="1 surface,required", legs="4 lines, required", back="1 surface, required"}.

By Stephen Bennett