Structural Ambiguity In Context

Jean-Baptiste Berthelin, Limsi, Orsay, France.

L&TC-2005,
Poznan, Poland.

(see also a
French version)

Abstract We designed a tool for exploring the Web and looking for a special kind of structural ambiguity, namely, French sentences where the central component can be either an adjective or a verb. Using this tool, we can actually count how many cases of such sentences are produced by Web users, and which reading for them is generally preferred.

1. Introduction

       Structural ambiguity is most simply represented in English by sentences like

             Flying planes can be dangerous,

where the NP flying planes may have either flying or planes as its head. This paper explores a similar construction, making use of the Web as a background where instances of actual user behaviour can be found.


In French, a more complex structure is classically taken as an example of structural ambiguity, combined with some lexical one. Let us introduce it with the sentence

             Le visiteur illustre le livre,

where the verb can be either illustre (illustrates) or livre (delivers), yielding respectively The visitor illustrates the book and The illustrious visitor delivers it as English equivalents.

      

This kind of sentence can be created by combining three ambiguous components:

yielding two different readings, which can be named Subject+Verb+Object and Subject-with-Adjective+Pronominal-Object+Verb, or, if one puts the main components of the sentence in brackets, [le visiteur] [illustre le livre] as opposed to [le visiteur illustre] [le livre]. As will be shown later on, SVO is normally preferred over SOV, although the latter is perfectly well-formed.

Since one can find approximately 40 Adjective-Verb forms and 200 Verb-Noun forms, the total number of such combinations would be something like eight thousand, if no selectional restrictions existed on the assembly of such words. However, teaching material in NLP contains only half a dozen of them. So the "real" count falls somewhere between 6 and 8,000. But how do we find out?

       To answer this question, we developed a Javascript Web-searching tool that sends a request to a web search engine for each Adjective-Verb/Verb-Noun pair. In most cases, no match is found. The "positive" returns must be carefully interpreted, as several cases of "false positive" can be identified, manually for the time being, but the criteria being extremely simple, automation of this procedure should not be too difficult.


2. Searching the Web for a particular kind of ambiguity

       We defined an algorithm for generating the exhaustive search of sentences with such a structure: essentially, a loop with two parameters Adjective-Verb av and Verb-Noun le+vn, la+vn, etc, producing, for each pair, either a negative response or a hypertext link to the search result page for the entry

             "adjective-verb+le-verb-noun".

By following all these links, we explored the Web selectively, and discovered new examples of this well-known kind of sentence. For each successful search about an adjective-verb, verb-noun pair, we thus found one possible subject, or several. By so doing, we use the Web, not only as a disambiguation corpus, but also as a source of discoveries on language use.


3. Our findings so far



4. About the relevance of our examples


The search engine counted the pages found for each of our "natural" examples. It ranges between 1, for most of them, and over 70, in one case, namely

La matière courbe l'espace

with the meaning of "Matter curves space", a sentence found in a great number of physics tutorials.

For each of these examples, it is meaningful to ask which reading is to be preferred, between Subject+Verb+Object and Subject-with-Adjective+Pronominal-Object+Verb.

In all cases but one, the first reading, SVO, is to be chosen and this can be established, without any explicit POS-taggind, by merely checking other occurences of the ambiguous components, while exploring the document retrieved by the search engine. This gives an empirical foundation to the following conjecture: when producing this kind of sentence, writers, consciously or not, expect readers to process them in this way. At this point, we should consider the question of a "Web bias", namely, of a correlation between "writing for the Web" and "favouring the simpler reading of an ambiguous sentence". This will involve comparisons with other corpora, mainly those available in the FRANTEXT database.



5. About an interesting exception

Contrary to what occurs in all other cases, sentences of the form

Le X aveugle le Y

are meant to be read as "The blind X Ys it", rather than "the X blinds the Y", for instance:

L'amour aveugle le conduit

means "Blind love leads him", not "Love blinds the pipe".

One possible explanation of this finding is that the verb aveugler (to blind) is much less frequent than the adjective aveugle (blind), according to the search engine itself, returning approximately 26,000 hits for the former and 350,000 for the latter. Moreover, the subject of the sentence must contribute to its global reading, as, in our example, we find a traditional underlying cliché, according to which "love is blind". As we said above, this phenomenon is probably not corpus-independent, and should be searched for in other places than the Web. Respective frequencies of homographous verbs and adjectives would be best checked in dedicated databases for the implied language.

6. The particular context of NLP tutorials

Some other examples occur only in metalinguistic settings, e.g. in websites dedicated to wordplay, or in tutorials about Natural Language Processing. The most well-known are

(1) Le pilote ferme la porte (either "The pilot closes the door" or "The firm pilot carries her"

(2) La petite brise la glace (either "The little one breaks the ice" or "The feeble breeze freezes her"

(3) Le boucher sale la tranche (either "The butcher salts the slice" or "The dirty butcher chops it"

(4) La substance trouble leur nuit (either "The substance disturbs their night" or "The cloudy substance harms them")

(5) L'ornithologue lâche la plume (either "The ornithologist lets go off the feather" or "The cowardly ornithologist plucks it"

They are actually retrieved by our script, and we can verify that they do not appear in "natural" contexts. In other words, uttering them is useful for illustrating structural ambiguity, but that does not give them a useful content. This remark is probably valid for many of the about 8,000 combinations that do not belong to the set of "natural" examples. For, as well as in Chomsky's famous "Colorless green ideas sleep furiously" example, readers of the randomly generated sentences have a capacity of making up a context in which they are meaningful. Therefore, selected random sentences, in most cases, would be as valid as "canonical" ones for teaching purposes. To observe that effect, readers are invited to look at our dynamic webpages instituteur.html and teacher.html, respectively for French and english.

7. Conclusion

By exploring the World Wide Web in search of examples of a particular kind of ambiguity, we discovered that in almost all found documents, the sentences of this kind had a preferred reading, namely as "Subject-Verb-Object". One remarkable exception reflects the rarity of the corresponding Verb.

Difficulties in interpreting our results are following.

Using the Web as a corpus is perhaps not as straightforward as it appears at first sight. Does Web material truthfully reflect a natural language's possibilities, or a relevant subset of them? Are search engines not biased? How do we deal with instability?

A second set of problems revolves around the usefulness of our findings. Let us admit that we discovered an interesting property of one kind of ambiguity, namely, that a given reading is generally privileged. How does it affect question-answering systems, or the design of parsers, the strategies for information retrieval, the teaching of French as a foreign language, and so on?

A third question is about the possibility of extending our approach to other ambiguous patterns, including some for the English language. We started a campaign of observations about "Flying planes" sentences, but there are a great many of them, so the process is time-consuming. It would perhaps be suited for a project involving a large number of participants.

We are not yet proposing answers to all these questions, but generally speaking, we consider that Web searching tools provide a convenient experimental setting for tackling questions about language ambiguity, and should therefore be offered to students and researchers in NLP.




References

Berthelin, J. B.: Principles for Linguistic Computation
www.limsi.fr/Individu/jbb/plc.html

Brill, Eric and Raymond J. Mooney, 1997: An Overview of Empirical Natural Language Processing, AI Magazine Volume 18 Number 4.

Chomsky, Noam, 1965: Aspects of the Theory of Syntax, Cambridge, MA : The MIT Press.

Gazdar, Gerald and Chris Mellish, 1989: Natural Language Processing in X., Addison-Wesley.

Graner, Nicolas, 1999: Palindrome, http://graner.net/nicolas/OULIPO/exp-palindrome.html

Hoenisch, Steve, 2004: Identifying and Resolving Ambiguity, http://www.criticism.com/linguistics/types-of-ambiguity.php

Kilgariff, Adam and Gregory Grefenstette, 2003: Introduction to the Special Issue on the Web as Corpus, Computational Linguistics Volume 29, Number 3.

Maher, John and Judith Groves, 1996: Chomsky for beginners, Penguin Books.

Max, Aurélien, 2003: The Syntax Student's Companion, http://clips.imag.fr/geta/aurelien.max/SSC/

Pallier, Christophe, 2004: Comparing corpus-based counts versus web page counts as estimates of lexical frequency, http://pallier.org/ressources/lexiquefreq/Pallier_lexiquefreq.pdf