Data processing: speech signal processing – linguistics – language – Linguistics – Natural language
Reexamination Certificate
1999-09-23
2002-06-11
Edouard, Patrick N. (Department: 2644)
Data processing: speech signal processing, linguistics, language
Linguistics
Natural language
Reexamination Certificate
active
06405162
ABSTRACT:
FIELD OF THE INVENTION
The invention relates to techniques that semantically disambiguate words using rules, referred to herein as “semantic disambiguation rules” or simply “disambiguation rules”.
BACKGROUND
Segond, F., Aimelet, E., and Jean, C., previously developed semantic dictionary look-up (SDL), a technique that uses dictionary information about subcategorization and collocates to disambiguate word sense. The SDL uses a dictionary, specifically the Oxford University Press-Hachette bilingual French-English, English French dictionary (OUP-H), as a semantically tagged corpus of different languages. SDL selects the most appropriate translation of a word appearing in a given context, and reorders dictionary entries making use of dictionary information.
To extract functional information from input text in order to match against OUP-H information, SDL uses an incremental finite state parser. The parser adds syntactic information in an incremental way, depending on the contextual information available. SDL matches relations extracted by the parser against collocates in the OUP-H, and, if a match is found, SDL reorders the dictionary entry to propose the OUP-H translation that includes the matching collocate rather than the first sense in the OUP-H. In case of information conflict between subcategorization and collocates, SDL gives priority to collocates.
Dini, L., DiTomaso, V., and Segond, F. also previously developed Ginger II, a semantic tagger that performs “all word” unsupervised word sense disambiguation for English. To automatically generate a large, dictionary-specific semantically tagged corpus, Ginger II extracts example phrases found in the text in machine-readable dictionary entries from the HECTOR dictionary described in Atkins, S., “Tools for corpus-aided lexicography: the HECTOR project”,
Acta Linguistica Hungarica
, Budapest, Vol. 41, 1992-93, pp. 5-72. Ginger II attaches to each headword in this text the dictionary sense numbering in which the text was found. This provides the sense label for the headword in that context. Ginger II then builds a database of semantic disambiguation rules from this labeled text by extracting functional relations between words in these corpus sentences.
The rules can be on two layers—a word layer and/or an ambiguity class layer. The rules are extracted directly with a nonstatistical approach, using all functional relations that are found. When an example z is listed under the sense number x of a dictionary entry for the word y, Ginger II creates a rule that, in usages similar to z, the word y has the meaning x. A HECTOR sense number is used to represent the headword in a rule based on an example in the dictionary entry for that sense, while WordNet tags are used for other words in the examples. One type of rule indicates, for a specified ambiguity class, that it disambiguates as a specified one of its members when it has a specified functional relation to a specified word. Another type of rule indicates, for a specified ambiguity class, that it disambiguates as a specified one of its members when it has a specified functional relation to a specified ambiguity class.
Ginger II applies the rules to a new input text to obtain as output a semantically tagged text, giving word layer rules priority over class layer rules. If more than one rule from the same layer matches, the applier uses the notion of tagset distance in order to determine the best matching rule. The metric for computing the distance can be set by the user and can vary across applications.
SUMMARY OF THE INVENTION
The invention addresses problems that arise with the previous techniques of Segond et al. and Dini et al., described above.
The SDL technique of Segond et al. uses information from a dictionary to disambiguate word senses, but depends on finding a very precise match between a relation in input text and a collocate in a dictionary example. After obtaining information about a relation from the parser, SDL accesses a dictionary entry for a word in the relation to determine whether a matching collocate occurs in one of the senses in the entry. If a precise match occurs in one of the senses, SDL selects that sense for the word. SDL will not, however, obtain any information from a collocate that does not precisely match the input text relation. This problem, referred to herein as the “precise match problem”, reduces the ability of SDL to disambiguate words in contexts that are similar but not identical to collocates.
Ginger II of Dini et al. employs disambiguation rules that include ambiguity classes. Ginger II therefore alleviates the precise match problem, because a rule with an ambiguity class may be applicable when contexts do not precisely match. But because Ginger II can produce a very large number of rules from a detailed dictionary, there are often two or more disambiguation rules at the word layer or at the class layer that match a functional relation in an input text. When this occurs, Ginger II computes a tagset distance to determine the best matching rule. In practice, however, the tagset distance technique sometimes fails to select the best rule, and instead selects a rule that produces incorrect disambiguation. This problem is referred to herein as the “incorrect rule problem”, and it is likely to become more serious as more dictionary information is used to produce rules, because more rules will result.
The precise match problem and the incorrect rule problem appear to be in tension: The use of ambiguity classes to alleviate the precise match problem, as in Ginger II, would make the incorrect rule problem worse. On the other hand, limiting the number of possible matches to avoid the incorrect rule problem, as the SDL technique implicitly does, would lead to the precise match problem.
The invention is based on the discovery of techniques that can alleviate both the precise match problem and the incorrect rule problem. The techniques can be used with ambiguity classes, thus alleviating the precise match problem; the techniques also provide flexible ways to select rules, making it possible to alleviate the incorrect rule problem. The techniques, as implemented, employ more dictionary information than Ginger II, thus obtaining more rules, but can nevertheless select rules without difficulty.
The techniques use disambiguation rules derived from different types of information in a corpus, and select one rule rather than another based on the types of corpus information from which the rules are derived. A detailed corpus such as a dictionary typically contains several different types of information, and rules obtained from some types of information are more likely to disambiguate a word correctly than rules obtained from other types. Because of the additional precision implicit in the rule types, the techniques sometimes lead to better rule selection than purely distance-based techniques.
In addition, the rules can include both word-based rules that specify a relation between specified words and class-based rules that specify a relation between a word and a class or between two classes. Where both a word-based rule and a class-based rule match a text, disambiguation can sometimes be improved by selecting the word-based rule rather than the class-based rule. When a text is matched by more than one rule at the same level of specificity, whether word-based or class-based, a rule can be selected based on information type.
Therefore, the invention alleviates the precise match problem by permitting class-based rules and also alleviates the incorrect rule problem by permitting selection of a disambiguation rule based on type of information. Already, the techniques can sometimes obtain better disambiguation results than conventional distance-based selection. Because selection of rules based on information type is more flexible than distance-based selection, the techniques offer the possibility of further improvements in disambiguation results.
The techniques can be implemented in a method that obtains information about a context in which a semantically ambiguous word
Brun Caroline
Segond Frédérique
Edouard Patrick N.
Oliff & Berridg,e PLC
Xerox Corporation
LandOfFree
Type-based selection of rules for semantically... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Type-based selection of rules for semantically..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Type-based selection of rules for semantically... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2944846