Computer Translation of Homonymy and Polysemy

 

In this paper, I want to discuss computer translation from the perspective of its treatment to homonymy and polysemy. The term homonymy is used when one form has two or more unrelated meanings. Examples of homonyms are the pair bank (of a river)—bank (financial institution). The term polysemy refers to a word with related meanings. For instance, the word “head” can be used to describe the object on the top of the body, that on top of a glass of beer, or the chief of a department of a company.

 

First, I will explain the importance of the treatment to homonymy and polysemy in translation from Sociosemiotic theories. The study of meaning plays an essential role in the school of Sociosemiotics. Meaning is defined from the angle of the relationship between the sign and something out it, in other words, the attribute of the sign or the symbol. Moreover, the relationship is classified into that between signs and entities, between signs and users, and between signs themselves, giving rise to three kinds of meanings: referential meaning, pragmatic meaning and intralingual meaning. Referential meaning serves as the foundation of translation. Without correct identification of referential meaning, translation will either be built on sand or turn into the castle in the air. The school of Sociosemiotics attaches great importance to the realization of meaning potentials. The correct reference is a must for this process of realization. Almost every noun has its referential meaning, but as a special group of nouns, homonymy and polysemy pose special difficulty in finding out their referential meanings. As explained at the beginning, homonymy and polysemy both have a number of meanings, related or unrelated. Therefore in translation, the translator is faced with the immediate problem of identifying the correct reference within the context. In other words, he has to carry out the job of disambiguation. This is the same in computer translation.

 

Sadly, important as the problem may be, it seems to be overlooked. The testing guideline for machine translation, detailed enough to include a classification of the plural forms of nouns, excludes the test of polysemy and homonymy. It is of great importance for us to examine how computers deal with this problem, no matter how difficult it may prove to be.

 

In order to have a clear idea of the difficulty of computer in handling this problem, we need to first look into the process of how human beings resolve this problem. Human beings can clear away this kind of ambiguity by resorting to his cognitive abilities. Theoretically, the process of disambiguation goes through the following three steps. I will take “bank” as an example to illustrate how the hearer can clarify the ambiguity due to its feature of homonymy. Suppose two friends run into each other near a river. One says, “the bank is beautiful.” How does the hearer interpret the reference of “bank”? First, he resorts to the physical surrounding and the reference of riverbank immediately comes to mind. In this way, the ambiguity is removed. Suppose two friends are in a dormitory chatting. One says, “the bank is beautiful.” How can the hearer interpret this word without any hint from physical surrounding? In this case, he has to trace this word back to his working memory, and then it occurs to his that they met each other on the riverbank the day before. Therefore, he realizes that the speaker must be referring to the place they met. Working memory is the second step if no reference can be drawn from physical surroundings. Suppose, two friends meet on street and one says, “I am on my way to the bank. They sent me the wrong bill again!” In this context, the hearer can no longer rely on the physical surrounding or working memory for disambiguation. Therefore, he has to employ his “knowledge of the world” for clarification, which will tell him that a riverbank can never send someone a bill. In this way, he can figure out that the speaker is talking about the financial institute.

 

The above is an illustration put forward by Sperber and Wilson as an example of cognitive capabilities of human beings. From this example, we can see that one resorts to his physical surroundings, working memory and knowledge of the word to clarify the ambiguity caused by homonymy. Therefore, it seems that computer has to be equipped with the same abilities in order to perform the job. Let’s examine the translation of one famous software “东方快车” in China.

 

Generally speaking, the translation software can to some extent distinguish different references of a word and pick out the correct one based upon the context. The following translation is done by “东方快车2003”.

 

1.    The bird has interesting bills.  鸟有有趣的帐单。

 

2.    The bill was passed. 法案被通过。

 

3.    I was sent the wrong bill by the company. 我被公司送错误的法案。

 

4.    I received my bill yesterday. 昨天我收到帐单。

 

Here the homonymy “bill” is translated within different context. Though the sentences are short, they have provided enough clues for disambiguation. Most people won’t have much difficulty in pointing out the right referential meaning of the word in these sentences. However, as we can see, computers are incapable of clear identification. The first translation does not make any sense, which may be attributed to the incomplete collection of referential meanings of “bill”. Here the word means the jaws of a bird together with their horny covering. Probably this meaning is excluded from the dictionary. The translation of the third sentence is also problematic. Most people will interpret the bill as a kind of account, rather than legal bill. This mistranslation should be attributed to a different reason, since the referential meaning of account is included in the dictionary, as illustrated in the translation of the fourth sentence. Apparently here the computer fails to identify the correct referential meaning. Let’s try another word “head”. The following translation is performed by the same software.

 

1.    Use your head and think! 使用你的头并且想！

 

2.    Two heads are better than one. 2个头是更好与比一个。

 

3.    Heads, I win. 头，我赢。

 

4.    He is at the head of class. 他在班的头。

 

5.    He is the head of school. 他是学校的头。

 

6.    He is the head of the English Department. 他是英语的部门的领导。

 

This group of sentences contains a number of mistranslations as well. Like the word “bill”, the referential meanings of “head” are not complete, evidenced by the translation of the third sentence. The first two sentences are not translated very well, because “head” in both these two sentences refers to something intellectual instead of something physical. What is more interesting is the last two sentences, in which the word should be interpreted as “chief of an organization” both. However, with almost the same structure, the word is translated differently. From this comparison, it is clear that there is still great randomness in choosing the correct referential meaning. This kind of arbitrary choice can be further proved by the following example.

 

I was sent the bill. 我被送帐单。

 

I was sent the wrong bill. 我被送错误的法案。

 

It is ridiculous to change the translation of “bill” simply because an adjective is added. More than that, this adjective makes it more explicit that the bill here refers to account instead of legal bill. Randomness in choice of referential meaning asserts itself clearly. 

 

The reason for this kind of mistranslation is self—evident. Computers are not endowed with cognitive capabilities, at least at present, to identify the correct referential meaning, by means of physical surrounding, working memory or knowledge of the world. What computers can rely on is from sheer input. Therefore, the question facing computer translation is how to make the full use of input. To give the question a tentative answer, I propose a solution from the theory of Relevance.

 

As the latest theory in Pragmatics, the theory of Relevance starts from the field of conversation analysis and has been playing an increasingly important role in various linguistic areas. Its application has even been extended beyond literary works to fields like advertisement. There are two major ideas that can be borrowed from the theory to computer translation. One is the principle of economy and the other is the viewpoint of reconstruction.

 

The theory of Relevance holds the belief that hearers of a conversation always try to make the least effort to comprehend as much as possible, which is called the principle of economy. This principle is very enlightening in the design of translation software. For instance, when people see the word “bill” in their daily life, most people will associate it with all kind of accounts they receive every day. Therefore, when they encounter the word in natural conversation, they always presuppose the referential meaning of word to be accounts, rather than a kind of legal document, unless something else in this context changes their mind, like the phrase “Bill of Rights”. As long as the presuppositional meaning does not deter comprehension, the hearer won’t bother to make any effort to change it. In the same way, every word or phrase should be given a “default meaning” in computer translation. It is particularly important for homonymy and polysemy, since they contain two or even more referential meanings. If every homonymy and polysemy can be given a default meaning, as long as the default meaning does not hinder the translation of the rest of the sentence, there is no need to try another referential meaning. The process is the same as how human beings understand conversations according the theory of Relevance. In this way, more resources can be saved to analyze and translate other parts of the sentence.