A solution for the problem of interactive disambiguation

• Authors: Hervé Blanchon
• Venue: COLING '92 Proceedings of the 14th conference on Computational linguistics - Volume 4
• Year: 1992

After the experiences of dialogue based MT systems with ITS [9],N-Tran [6] and KBMT-89 [5], the LIDIA project aims at theconstruction of a mock-up of a personal MT system for a monolingualuser. One major aspect of the LIDIA project is thus, the study of adialogue of standardization and disambiguation between the systemand the user so as to produce a high quality translation. Thisdialogue satisfies two properties: its questions are explicit, sono linguistic knowledge is required; its questions are monolingual,so no foreign language knowledge is needed. Here, we focus on onepart of the disambiguation process: the disambiguation of thestructure produced by the analyser.The structure produced by our analyser is called MMC(Multisolution, Multilevel and Concrete). Multisolution means thatthe analyser produces every analysis fitting with the syntagmatic,syntactic and logico-semantic model of the grammar (an example isshown fig. 1). Multilevel means that the same structure consists ofthree levels of linguistic interpretation, namely the level ofsyntactic and syntagmatic classes, the level of syntactic functionsand the level of logic and semantic relations. Finally, thestructure is said to be concrete because the original utterance canbe found back by a simple left-to-right reading of thestructure.We have taken into account three kinds of differences betweenthe solutions produced for one sentence, and each kind ofdifference is associated with the name of an ambiguity. We havedefined ambiguities of syntactic classes (cf fig. 2), ambiguitiesof geometry (cf fig. 3) and ambiguities of syntactic, logic andsemantic decoration (cf fig. 4). We have also defined threeprinciples (§ III. 1) to order the questions if there ismore than one to be asked. The first principle is: first of all,find out the right segmentation into simple sentences. The secondprinciple is: for each common predicate in the MMC structure, findout the right subject, objects and adjuncts. The last principle is:for each simple sentence, find the right structure.With those principles we are able to define a strategy (cf fig.5). We have also isolated some patterns in the three classes ofambiguity. The class of ambiguities of syntactic classes needs norefinement (§ III. 3.1). On the other hand we create fourpatterns of ambiguity of geometry (§ III. 3.2) called:verbal coordination, argument structure of the verb, non verbalcoordination, subordination; and three patterns of ambiguity ofsyntactic, logic and semantic decoration (§ III. 3.3)called: logico-semantic labelling, argument order of directtransitive verbs, syntactic labelling.Here is an example with the interpretations for each pattern wehave chosen:Problem of class. Le pilote ferme la porte:The firm pilot carries her. The pilot shuts the door.Problem of verbal coordination. II regarde laphoto et la classe: He looks at the photograph and the class. Helooks at the photograph and files it.Problem of the argument structure of theverb. II parle de l'école de cuisine: He talksabout the cooking school. He talks from the cooking school. Hetalks from the school about cooking.Problem of non-verbal coordination. II prenddes crayons et des cahiers noirs: He takes pencils and blacknotebooks. He takes black pencils and black notebooks.Problem of subordination. L'écolede cuisine lyonnaise est fermée: The lyonnaise cookingschool is closed. The school of lyonnaise cooking is closed.Problem of logico-semantic labelling. Pierrefait porter des chocolats à Lucie: Pierre lets Luciecarry chocolates. Pierre gets chocolates to be delivered toLucie.Problem of argument order of direct transitiveverbs. Quel auteur cite ce conférencier: Whichauthor this lecturer is quoting? Which lecturer this author isquoting?Problem of syntactic labelling. II parle dela tour Eiffel: He is talking about the Eiffel Tower. He is talkingfrom the Eiffel Tower.For each pattern we have defined a method to produce theappropriate dialogue (§ III. 3). These methods use twokinds of processing: projection and paraphrase. To buildparaphrases we use basically three operators: an operator ofsemantic replacement of occurrence, an operator of permutation ofgroups of occurrences and an operator of distribution ofoccurrences. The examples (§ IV) give an idea.In conclusion we can say that our method is quite simple butfixed once and for all. We are going to study two points in thenear future. The first one is to reduce the number of analysis andthus, by getting information from the user, reduce the time tospend on the disambiguation. The second is to try to build toolswhich will allow the linguist, designer of the linguistic part ofthe LIDIA system, to define its own methods of disambiguation.