Robotics Reference
In-Depth Information
When a sentence has been syntactically and semantically analysed,
it is still possible that its context within a whole text might have some
effect on how it should be interpreted. A word such as “it” or “he” in one
sentence might relate to an object or person in the previous sentence,
or even earlier. Consider the two adjacent sentences: “Harry ate an ice
cream. He enjoyed it.” A semantic analysis of the second sentence needs
to refer back to the first in order to determine who “he” is, and what
“it” is. And clearly the problem can involve sentences that are far from
adjacent: “Harry ate an ice cream. He enjoyed it. His sister ate one too.
She enjoyed hers as well.” Here the semantic analyser has to track back
from “hers” to “one” to “it” to “an ice cream” in order to understand the
meaning of the fourth sentence. This task of deriving meaning with the
help of contextual information is performed by a software module called
a
discourse analyser
.
The First 50 Years of NLP
The above description of some of the multifarious difficulties facing NLP
researchers should help to explain why this particular sector of AI has
made relatively little progress despite half a century of research. Let us
now trace some of the history of this topic and compare the states-of-
the-art at the beginning and end of that half century of development.
The first work with computers on processing natural language was
focussed on translation.
4
Those first attempts involved automatically
replacing words in the source text by their equivalents in the target lan-
guage and then manually adjusting the word order to conform with rules
of style and grammar, so that the translated text (hopefully) appeared
natural. But that approach was fairly quickly seen to be leading nowhere,
and NLP researchers instead turned their attention to the problem of
understanding, in the realisation that if a computer could actually un-
derstand the meaning of a sentence then it could translate that sentence
into another language, or answer questions about the sentence, or store
away the knowledge contained in the sentence for use at a later date. As
more and more researchers entered the field, fascinated by the problems
of NLP, so more and more problems and complexities within NLP were
uncovered. The understanding of natural language was seen to require
highly complex cognitive abilities involving different types of knowledge,
including the meanings of words (and questions such as “when does
4
See the section “Machine Translation” in Chapter 2.
