<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1077">
  <Title>A PROPOSAL FOR LEXICAL DISAMBIGUATION</Title>
  <Section position="3" start_page="397" end_page="397" type="metho">
    <SectionTitle>
THE PROPOSED SYSTEM
</SectionTitle>
    <Paragraph position="0"> It is assumed that a grammatical text is to be processed, and that the processor is expected to use the textual context to determine the appropriate sense of each successive content word. Then, in brief outline, the present proposal envisions a processor that will perform three operations: (1) Take a content word from the text and look it up in the lexical database; if a single sense is found, the problem is solved. If more than one sense is found, continue.</Paragraph>
    <Paragraph position="1"> (2) Determine the syntactic category of each sense. If a single category is involved, go to operation three. If more than one syntactic category is found, use a &amp;quot;parts&amp;quot; program to determine the appropriate eatagory. If the word has only one sense as a member of that category, the problem is solved. If the word has more than one sense in the appropriate syntactic category, continue.</Paragraph>
    <Paragraph position="2"> (3) Determine which sense of the polysemous word is appropriate to the text. If the word is a noun, determine which sense can serve as an argument of the verb, or can be modified by an accompanying adjective. If the word is verb or adjective, determine which sense can be combined with an accompanying noun phrase.</Paragraph>
    <Paragraph position="3"> The final operation is the critical step, of course, but before describing how it might be implemented, a simplified example will help to make the central idea clear. Suppose the processor encounters the sentence, the baby is in the pen, and tries to assign the appropriate sense to the noun pen. It would first generalize the given context (e.g., with respect to number and tense), then find words that are semantically related to the various senses of pen and substitute them into the generalized context. It would then undertake a comparison of:  In order to decide that one of these is acceptable and the others are unlikely, the processor might search an extensive corpus for strings of the form &amp;quot;(a/the baby is/was)/(the babies are/were) in the X,&amp;quot; where X is one of the closely related words listed above. If the playpen/playroom~nursery expressions significantly outnumber the others, the conventionally correct choice can be made. In other words, the processor will interrogate a corpus much the way a linguist might ask a native informant: &amp;quot;Can you say this in your language?&amp;quot; That is the basic strategy. Words related in meaning to the different senses of the polysemous word will be retrieved; new expressions will be derived by substituting these related words into the generalized context of the polysemous word; a large textual corpus will then be searched for these derived expressions; that sense will be chosen that corresponds to the derived expression that is found most often in the corpus. (Alternatively, all contexts of the semantically related words could be collected and their similarity to the target context could be estimated.) We assume that the similarity of this strategy to the theory of spreading activation (Quilliam 1968, 1969) is obvious. Of course, in order even to approach the best possible implementation, a variety of possibilities will have to be explored. For example, how much context should be preserved? Too short, and it will not discriminate between different senses; too long and no instances will be found in the corpus. Should the grammatical integriw of the contexts be preserved? Or, again, how large a corpus will be required? Too small, and no instances will be found; too large and the system will be unacceptably large or the response unacceptably slow. Fortunately, most of the polysemous words occur relatively frequently in everyday usage, so a corpus of several million words should be adequate. Or, still again, how closely related should the semantically related words be? Can superordinate terms be substituted? How far can the contexts be generalized? Experience should quickly guide the choice of sensible answers.</Paragraph>
    <Paragraph position="4"> As described so far, the proessor begins with WordNet in order to find semantically related words that can be searched for in a corpus. Obviously, it could all be done in the reverse order. That is to say, the processor could begin by searching the corpus for the given generalized context. In the above example, it might search for &amp;quot;(a/the baby is/was)/(the babies are/were) in the Y,&amp;quot; where Y is any word at all. Then, given the set of Y words, WordNet could be used to estimate the semantic distance from these words to the alternative senses of the polysemous word. A similarity metric could easily be constructed by simply counting the number of pointers between terms. That sense would be chosen that was closest in meaning to the other words that were found to occur in the same context.</Paragraph>
    <Paragraph position="5"> Whether WordNet is used to provide related words or to measure semandc similarity, a major component of the present proposal is the search of a large textual corpus. Since the corpus would not need to be continually updated, it should be practical to develop an inverted index, i.e., to divide the corpus into sentence items that can be keyed by the content words in WordNet, then to compute hash codes and write inverted files (Lesk, 1978). In 'this way, a small file of relevant sentences could be rapidly assembled for more careful examination, so the whole process could be condueted on-line. Even if response times were satisfactorily short, however, one feels that once a particular context has been used to disambiguate a polysemous word, it should never have to be done again. That thought opens up possibilities for enlarging WordNet that we will not speculate about at the present time.</Paragraph>
  </Section>
  <Section position="4" start_page="397" end_page="398" type="metho">
    <SectionTitle>
SOME OBVIOUS APPLICATIONS
</SectionTitle>
    <Paragraph position="0"> Several practical applications could result from a refiable lexical disambiguation device. The fact that people see concepts where computers see strings of characters is a major obstacle to human-machine interaction.</Paragraph>
    <Paragraph position="1"> Consider this situation. A young student who is reading an assignment encounters an unfamifiar word. When a dictionary is consulted it turns out that the word has several senses. The student reconsiders the original context, testing each definitional gloss in turn, and eventually chooses a best fit. It is a slow pro- null cess and a serious interruption of the student's task of understanding the text. Now compare this alternative. A computer is preumfing a reading assignment to the same studant when an unfamiliar word appears. The student points to the word and the computer, which is able solve the polysemy problem, presents to the student only the meaning that is appropriate in the given context--as if a responsive teacher were sitting at the student's side. The desired information is presented rapidly and the real task of understanding is not interrupted.</Paragraph>
    <Paragraph position="2"> Or think of having a lexical disambiguator in your word processing system. As you write, it could flag for you every word in your text that it could not disambiguate on the basis of the context you have provided. It might even suggest alternative wordings. null The application to mechanical translation is also obvious. A polysemous word in the source language must be disambiguated before an appropriate word in the target language can be selected. The feasibility of multilingual WordNets has not been explored.</Paragraph>
    <Paragraph position="3"> Finally, consider the importance of disambiguation for information retrieval systems. If, say, you were a radar engineer looking for articles about antennas and you were to ask an information retrieval system for every article it had with antenna in the title or abstract., you might receive unwanted articles about insects and erustaceans---the so-called problem of false drops. So you revise your descriptor to, say, metal antenna and try again. Now you have eliminated the animals, but you have also eliminated articles about metal antennas that did not bother to include the word metal in the title or abstract--the so-called problem of misses.</Paragraph>
    <Paragraph position="4"> False drops and misses are the Scylla and Charybdis of information relrieval; anything that reduces one tends to increase the other. But note that a lexical disambiguator could increase the probability of selecting only those titles and abstracts in which the desired sense was appropriate; the efficiency of information retrieval would be significantly increased.</Paragraph>
    <Paragraph position="5"> In short, a variety of practical advances could be implemented if it were possible to solve the problem of lexical ambiguity in some tidy and reliable way. The problem lies at the heart of the process of turning word forms into word meanings. But the very reason lexical disambiguation is important is also the reason that it is difficult.</Paragraph>
  </Section>
class="xml-element"></Paper>