The primary problem with natural language processing is the ambiguity of language. There are a number of levels at which ambiguity may occur in natural language (of course a single sentence may include several of these levels). First a sentence or phrase may be ambiguous at a syntactic level. Syntax relates to the structure of the language, the way the words are put together. Some word sequences make valid sentences in a given language, some do not.
However, some sentence structures have more than one correct interpretations. In the first place, these are syntactically ambiguous. Secondly, a sentence may be ambiguous at a lexical level. The lexical level is the word and ambiguity here occurs when a word can have more than one meaning. Thirdly, a sentence may be ambiguous at a referential level.
This is concerned with what the sentence (or a part of it) refers to. Ambiguity occurs when it is not clear what the sentence is referring to or it may legally refer to more than one thing. Fourthly, a sentence can be ambiguous at a semantic level, that is, at the point of the meaning of the sentence.
Sometimes a sentence is ambiguous at this level- it has two different meanings. Indeed this characteristic is exploited in humour, with the use of double extender and innuendo. Finally, a sentence may be ambiguous at a pragmatic level that is at the level of interpretation within its context.
ADVERTISEMENTS:
‘The same word or phrase may have different interpretations depending on the context in which it occurs. To make things even more complicated some sentences involve ambiguity at more than one of these levels.
Consider the following sentences; how many of them are ambiguous and how?
1. I hit the man with the hammer.
2. I went to the bank.
ADVERTISEMENTS:
3. He saw her duck.
4. Ram hit Ravana because he liked Sita.
5. I went to the doctor yesterday.
6. I waited for a long time at the bank.
ADVERTISEMENTS:
7. There is a drought because it has not rained for a long time.
8. Dinosaurs have been extinct for a long time.
In fact all the sentences above have some form of ambiguity.
Let’s look at them more closely:
ADVERTISEMENTS:
a. I hit the man with the hammer. Was the hammer the weapon used or was it in the hand of the victim? This sentence contains syntactic ambiguity: these are two perfectly legitimate ways of interpreting this sentence structure.
b. I went to the bank. Did I visit a financial institution or go to the river bank? This sentence is ambiguous at a lexical level: the word ‘bank’ has two meanings, either of which fits in this sentence.
c. He saw her duck. Did he see her dip down to avoid something or did he see “the web-footed bird” owned by her? This one is ambiguous at a lexical and a semantic level. The word ‘duck’ has two meanings and the sentence can be interpreted in two completely different ways.
d. Ram hit Ravana because he liked Sita. Who (Ram or Ravana) is it that likes Sita? This is an example of referential ambiguity. Whom does the pronoun he refer to, Ram or Ravana? It is not clear from this sentence structure.
ADVERTISEMENTS:
e. I went to the doctor yesterday. When exactly was ‘yesterday’? This demonstrates pragmatic ambiguity. In some situations this may be clear but not in all. Does ‘yesterday’ refer literally to the day proceeding today or does it refer to another yesterday.
f. I waited for a long time at the bank.
g. There is a drought because it has not rained for a long time.
h. Dinosaurs have been extinct for a long time.
The last three sentences can be considered together. What does the phrase ‘for a long time’ mean? In each sentence it clearly refers to a different amount of time. This again is pragmatic ambiguity. We can only interpret the phrase through our understanding of the sentence context.
In addition to these major sources of ambiguity (due to multiple word meanings, syntactic, or unclear antecedents) language is problematic because it is imprecise, incomplete, inaccurate and continually changing. Think about the conversations we have with our mends. The words we use may not always be appropriate enough to express the meaning we intend, we may not always finish a sentence, we may use analogies and comparisons to express ideas.
As humans we are adept at coping with these things, to the extent that we can usually understand each other if we speak the same language, even if words are missed out or misused. We usually have enough knowledge in common to disambiguate the words and interpret them correctly in context. We can also cope quickly with new words. This is borne out by the speed with which slang and street words can be incorporated into everyday usage.
How do we ever understand each other?
Some answers are:
How are these done is explained below:
i. Context:
To resolve ambiguity, understand an idea in context.
Consider for example:
“There were two men blocking my escape. One held a hammer, one had nothing in his hands” I knew that I could not hit both of them I hit the man with the hammer.
ii. Familiarity:
We tend to identify with the solutions which are familiar to us, making it easier to understand the language which deals with those situations.
For example:
Since we are familiar with the waiting in doctor’s office ‘a long time’ can be given stable interpretation.
iii. Expectations:
Through our experience we can come to expect certain things in certain situations.
For example:
Select the ending which is most likely to complete the following narration.
Ram went to restaurant and ordered for chicken. When the waiter brought his order to the table Ram;
(i) Started singing a song.
(ii) Ate the chicken.
While either of the ending is possible, mostly the second would be selected.
All of these present difficult problem for the computer and that’s why natural language processing is difficult to understand.
Early attempt at natural language understanding- SHRDLU
SHRDLU is the natural language processing system developed by Winograd at MIT USA, in the early 1970s. It is used for controlling a robot in a restricted ‘blocks’ domain. The robot’s world consists of a number of blocks of various shapes, sizes and colours, which it can manipulate as instructed or answer questions about. All instructions and questions are given in natural language and even though the robot’s domain is so limited, it still encounters the problems as mentioned below.
Consider for example the following instructions to the robot:
Find a block which is taller than the one you are holding and place it in the box.
How many blocks are on top of the green block?
Put the red pyramid on the block in the box.
Does the shortest thing the tallest pyramid’s supports anything green?
It is obvious that each instruction contains ambiguity of some kind.
The causes of ambiguity are given below:
(i) The first sentence refers to referential ambiguity. It is not clear to which word ‘it’ refers to
(ii) The second one is more tricky and it involves semantic ambiguity. Does ‘on top of mean directly on top of or above (that is it could be on top of a block which is on top of the green block)?
(iii) This is syntactic ambiguity is it the block that is in the box or the red pyramid which is being put into the box?
(iv) This is lexical ambiguity: there are two uses of the word ‘support’?
SHRDLU was successful whatsoever because it could be given complete knowledge about its world and ambiguity could be reduced (it only recognises one meaning of ‘block’ for instance and there is no need for contextual understanding since the context is given). It was therefore of no use as a general natural language processor.
However, it did provide insight into how syntactic and semantic processing of natural language understanding can be achieved.