22.03.2013 Views

Seeing clearly: Frame Semantic, Psycholinguistic, and Cross ...

Seeing clearly: Frame Semantic, Psycholinguistic, and Cross ...

Seeing clearly: Frame Semantic, Psycholinguistic, and Cross ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

CHAPTER 6. FUTURE RESEARCH DIRECTIONS AND CONCLUSIONS 232<br />

words, e.g. Resnik 1993).<br />

From Statistical Pro les to Senses<br />

A statistical pro le of this sort is interesting in its own right from the linguistic<br />

point of view, but for those working on NLP, it is only a means to the end of word sense<br />

disambiguation, which is increasingly recognized as a key problem in building practical NLU<br />

systems (witness the fact that an entire issue of Computational Linguistics was recently<br />

devoted to the topic). The statistical study of <strong>Frame</strong>Net words mentioned above istypical<br />

of the sorts of work now being done, in that it will use the h<strong>and</strong>-tagged examples of senses<br />

of words from <strong>Frame</strong>Net as a training set for an algorithm that will then look through the<br />

whole BNC for sentences containing those words <strong>and</strong> make a guess as to the sense of the<br />

word <strong>and</strong> the roles of its arguments, based on the statistical similarity of the contexts to<br />

those in the training set. (For an overview of current word sense disambiguation techniques,<br />

see Ide & Veronis 1998, <strong>and</strong> Light 1997.)<br />

The real test of the statistical pro les of the senses see developed in the rst step<br />

would therefore be how well they can be used to predict the sense from the syntax <strong>and</strong><br />

semantics of the arguments. St<strong>and</strong>ard evaluation procedures such as cross-validation can<br />

then be used to judge the accuracy of this method in comparison with other word-sense<br />

disambiguation algorithms currently in use.<br />

<strong>Cross</strong>-linguistic Survey<br />

The purposes of a cross-linguistic survey would be (1) to test whether native<br />

speakers of various languages actually use the terms listed in bilingual dictionaries to convey<br />

the various meanings of see <strong>and</strong> (2) to nd out the relative frequencies of the words used<br />

in such cases.For this purpose, we could use a translation task; this is less than ideal from<br />

a psycholinguistic point of view, since we aretestingavery high-level skill (translation),<br />

rather than a more direct meaning to form correspondence. However, it is much simpler to<br />

gather cross-linguistic data in this way than to try to replicate a psychological experiment<br />

in each of the countries, especially when dealing with meanings that are very di cult to<br />

depict, such as those of see.<br />

The survey would include at least 40 English sentences containing see, covering<br />

the range of senses. For each target language, at least 10 subjects would translate all of

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!