Syntax, Semantics, and Pragmatics in Concert: An Incremental, Multilevel Approach in Reconstructing Task-Oriented Dialogues
Proceedings of the Eighth International Joint Conference on Artificial Intelligence. Natural Language. Pages 721-722.
© 1983 International Joint Conferences on Artificial Intelligence
Abstract
This paper gives an overview of a model for the reconstruction of task-oriented dialogues based on an interactive, multilevel parsing formalism. It is applied to route description dialogues. It will be shown, how the pragmatic aspects of such dialogues are taken into account on different levels of processing. The approach described is based on an extension of the concept of cascaded ATNs. Furthermore this approach uses knowledge sources {KSs) for every participant in the dialogue in which knowledge about the world and a partner model is build up during the analysis of a dialogue. These KSs are supplied to the parsing process, as well. In this paper special importance is laid on the description of the interaction and cooperation of the different processing components of this formalism.
- R. J. Bob row/B. Webber, "Knowledge represen- tation for syntactic/semantic processing." Proc. AAAI, Stanford, 1980.
- T. Christaller/D. Metzing, "Parsing Interaction: a multilevel parser formalism based on cascaded ATNs." in: K. Sparck-Jones/Y. Wilks, Automatic Natural Language Parsing. Chichester, 1983.
- M. Gehrke, "Rekonstruktion aufgabenorientierter Dialoge mit einem mehrstufigen Parsing-Algorith- mus auf der Grundl age kaskadierter ATNs." in[9] .
- M. HuSmann, "Ell ipsenrekonstruktion im Kontext; Forderungen an eine Ellipsenkomponente und ihre Real isierung in SWYSS." in[9].
- W. Klein, "WegauskUnfte." Zeitschrift fUr Linguistik und Literaturwissenschaft, 9: 9 -57 ( 19 79) .
- W. Marslen-Wilson/L. K. Tyler, "The termporal structure of spoken language understanding." cognition, 8: 1 -72 (1980).
- D. Metzing, "Dialogue interactions in a Task Domain : Model and some Problems of Application," in: Hein/Hagglund (ed.), Models of Dialogue. MUnch en, 1982.
- F. di Primio/T. Christaller, "A poor man's flavor system." (to appear), ISSCO, Geneva, 1983.
- W. Wahlster (ed.), "Proc. of 6th German Work- shop on AI", Berlin-Heidelberg-New York, 1982.
- W. A. Woods, "Cascaded ATN Grammars." Journa 1 ~. 6:1 (1980) 1 -13.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
ABSTRACT
This paper gives an overview of a model for the reconstruction of task-oriented dialogues based on an interactive, multilevel parsing formalism. It is applied to route description dialogues. It will be shown, how the pragmatic aspects of such dialogues are taken into account on different levels of pro cessing. The approach described is based on an ex tension of the concept of cascaded ATNs. Further more this approach uses knowledge sources {KSs) for every participant in the dialogue in which knowledge about the world and a partner model is build up during the analysis of a dialogue. These KSs are supplied to the parsing process, as well. In this paper special importance is laid on the descrip tion of the interaction and cooperation of the different processing components of this formalism.
INTRODUCTION
In the project "Prozedurale Dialogmodelle" we have concerned ourselves with solving the follo wing problems by reconstructing task-oriented dia logues:
- What dependencies exist between the different linguistic processing components (LPCs) in rea listic task-oriented dialogues~ of the special type 'route description'?
- Which knowledge sources (KSs) participate in such dialogues?
- What informations do these KSs contain?
- At which point and with which information do the LPCs interact?
Furthermore, we assume that the cooperation of the various LPCs starts at a very early stage in processing an utterance in a dialogue. In our model the .LPCs are organized as coroutines, and not, as suggested by the results of Marslen-Wilson and Tyler [ 6], as parallel processes. So that, e. g. as soon as a partial syntactic description of an utterance is derived, the semantic components are triggered.
The linguistic basis of our approach is a scheme of dialogue interaction in a task domain(?), in which the purpose of an utterance in the dialogue context is de.termi ned with respect not only to the
· The research reported here is supported by a grant of the Deutsche Forschungsgemeinschaft.
~The dialogues, we are dealing with, were re corded in Frankfurt/Main for W. Klein [5).
SYNTAX, SEMANTICS AND PRAGMATics · IN CONCERT: AN INCREMENTAL, MULTILEVEL APPROACH IN RECONSTRUCTING TASK-ORIENTED DIALOGUES
Manfred Gehrke
Project "Prozedurale Dialogmodelle"· Department of Linguistics and Literature University of Bielefeld, Bielefeld, FRG
maintainance of the interaction.
THE MULTILEVEL PARSING FORMALISM
The reconstruction of dialogues is carried out by a multilevel parsing formalism [2, 3] in which the different components are independent, but closely cooperating. Under this viewpoint "parsing" is taken to mean 'application of recognition pro cedures' and not 'recognition of syntactic sen tence structures' .
The architecture of the formalism is based on the concept of cascaded ATN (CATN) [10] with the following important differences:
- the different levels of CATN are not neces sarily ATNs, and
- levels of an CATN can be 'by-passed'.
In addition to the usual means of communication in a CATN such as TRANSMIT there is a 'blackboard' (the common KS) in order to store the results of the various cascade levels. While the RUS Parser [1] is a 2-stage cascade, our formalism allows an arbitrary number of stages, in this application 6.
The syntactic component (SYNC) is an ATN. Due to our dialogues, which were uttered in a rather colloqual language and therefore overwhelmed with incomplete utterances, this component must be very robust. Therefore, a complete syntactic analysis is of minor importance. This is, e.g., obtained by enabling SYNC to reach a permissible final state, if its analysis of an utterance is not rejected by the semantic or pragmatic components. So the main purpose of SYNC is to detect phrases in a utterance (noun-, prepositional-phrase, adverb-preposition groups, etc.) and to make suggestions on their role in the utterance. Additionally it should detect wh- and imperative structures. As soon as a phrase or even a part of it is analysed, the semantic com ponent is triggered in order to see how it can process the results of SYNC. This includes an early detection of wrong pathes of analysis.
The semantic component (SEMC) is a case oriented production rule system. Because of the incremental manner of parsing, the semantic interpretation has to be tentative. The determination of a syntactic construct is - until the main verb is detected - more a suggestion rather than a categorization. Thus the case slots are defined in two ways. First ly, there are very general definitions of case slots applying before the main verb is detected which build up a proposed interpretation of the utterance.
Secondly, more specific definitions of these slots are connected to the respective verb frame, thus meeting its special demands. When the verb is trans mitted, the hitherto constituted case slots are tested and the definitions of the required case slots, which have not yet been constituted, are supplemented by the special requirements of the verb. Every constituted case slot is transmitted to the task-communication component.
The remaining levels of the cascade are respon sible for the interpretation of the dialogue from pragmatic points of view. For that purpose they are supported by KSs that contain knowledge of the world and a partner model (WPKSs) . The infor mation, gathered by the initiator of the direction giving dialogue (addresser) during the dialogue is stored in his/her WPKS as well as the assumptions on how the dialogue may proceed. The content of the addresser's WPKS could be stated after a first piece of route description, such as
Problem: addresser wants to get to the opera Addresser wants to know: solution of the problem Addresser believes: Addressee knows the solution Addresser believes: Addressee will tell him/her the solution
Addressee says: go straight ahead to the church Addressee says: arrive at Rathenauplatz
On the other hand, the addressee's knowledge about the course of dialogue is stored in his/her WPKS. In addition, each partner should have a pro blem solver which takes into account the appro priate KS for finding the way asked for. But this is beyond the scope of our project.
The Task-Communication-Component (TCC) con tains the task dependent, pragmatic categories. It is responsible for the detection of utterances bearing solutions to subproblems, while a general solution is the purpose of the dialogue. In the case of direction-giving dialogues the following task dependent categories are used :
- route description, i.e. utterances containing a piece of the proposed route, e. g. "You go straight ahead on that street."
- place descriptions, i.e. utterances bearing some information about where a change of di rection will take place, or confirming that one is on the right path, e. g. "Then there is Woolworth's to the left."
- partial goal determination, i.e. when an inter mediate goal is reached and where a change of direction often takes place, e. g. "Turn left at Wollworth' s."
- goal declaration, i.e. that point in which the reaching of the goal is stated, e. g. "Then you arrive at the station."
- destination specification, i.e. where the addresser utters the goal he will reach, e. g. "We are looking for the bus station.".
An Alteration of this type of task-oriented dia logues only requires a change in this component at the pragmatic level.
The Question-Answer-Interaction-Component (QAIC)
has to detect the patterns of speech acts used to control the cooperative organisation of mutual un derstanding. These are, e.g., the exchange of question, answer, assurance, confirmation, etc. The Task-Interaction-Component (TIC) provides a general scheme of task-oriented dialogues with categories like ask-for-information, information giving (e.g. "There you'd better ask someone else."). The final component is responsible for starting, keeping and ending the social contact (SCC).
All the components on the pragmatic level are ATNs, because task-oriented dialogues consist of rather fixed patterns of pragmatic categories. While it is true that a dialogue may be inter rupted by a clarification dialogue, such a clari fication has a fixed order, as well. Thus it is only necessary to detect the beginning of an em bedded dialogue. Markers ·for clarification are wh phrases inside the dialogue such as "Where?" or phrases like
"I don't get it! Would you please repeat that?"
INTERACTIONS WITH THE SYNTACTIC COMPONENT
The interaction of an ATN with its neighbouring components is carried out by the TRANSMIT-action. After traversing an arc with this action, pro cessing in the ATN is interrupted and control is passed over to the next component. Such TRANSMITs occur in SYNC for two reasons:
- at critical points in the analysis where a wrong decision can be rather resource consuming, e.g. when the head noun of a nominal phrase or a preposition after a nominal phrase is detected,
- at points where important results for SEMC are detected such as verbs, deictic or directional adverbs.
SYNC can receive two types of message from SEMC . On the other hand, if SEMC lacks further informa tion to process, then SYNC starts its analysis at the goal node of that arc on which the TRANSMIT occured. If, on the other hand, something is wrong with the analysis of SYNC, then backtracking is forced.
INTERACTIONS WITH THE SEMANTIC COMPONENT
SEMC can receive various types of message from SYNC. If a full verb is recognized, then it be comes head of the verb frame; previously filled case slots are matched against the specific require ments on case slots of that verb. In the case that further frames are possible for one verb these framshave to be instatiated and processed in a similar way.
All other syntactic constructs (nominal-, pre positional phrases, embedded sentences) are pro vided with case slots or categorized as modifiers of a verb, e.g. adverbs, auxiliary verbs, to mark tense, modality, etc.
When the end of an utterance is indicated, then the frame is checked for completion (i.e. a verb is recognized and all obligatory case slots are filled), otherwise possibly additional instantiated verb frames are not completed, it is assumed that the utterance is some sort of cotextual ellipsis [4] and an inference mechanism (see below) based on the pragmatic assumptions is initiated.
Finally SEMC can be "by-passed", when e.g. an interjection such as "yes, uhh" is detected by SYNC. An application of SEMC to such words doesn't contri bute to the whole understanding process. Therefore they are immediately routed to QAIC, in which they may be categorized as an assurance, confirmation, etc.
TCC may signal to SEMC that its result is not correct. This could be caused by a wrong choice of the appropiate frame, e.g. a camp 1 eted frame was sent as a interpretation, while an uncompleted frame held true for an elliptical utterance. Then the latter frame is transmitted instead, otherwise SY~C has to resume its analysis.
In cases where no verb frame can be completed, results are transmitted, too, but marked as uncom plete. Then several cases are to distinguish:
- To disambiguate elliptical utterances we had to consider that the missing phrase X is somehow described by the unsatisfied case slot and that X was in some way uttered earlier. Then the WPKS of the addressed participant is looked up, assuming that there is a referent for X. For an illustration, consider the following part of a dialogue :
"da kommen sie auf den Rathenauplatz, den Uberqueren s i e auch." ("then you come to the Ra thenaup 1 atz, you cross i t, too . ")
The second utterance has two alternative analysis. Either it is a relative clause with an incorrect word order. Or it is an independent utterance where the noun referring to the definite article "den" is missing. The frame of "Uberqueren" re quires as an object case slot a physical object that could be crossed. Secondly, the addresser has the strong assumption that the addressee will somehow conclude the route description. Des pite the missing phrase the utterance strengthens this assumption. Thus the addresser's WPKS is looked up in reverse order (see above) for a route or place description containing a referent to "den" . Here it will conclude in finding "Ra thenaupl atz" as referent.
- For deictical particles referents are similarly found, because they are often used to denote the last partial goal. A subgoal can be expressed as a named location in the GOAL case slot in the last route description, or as a location in the PLACE case slot in the last place description.
These particles (esp. "there") may have the special meaning of denoting the goal, as in "and then you are there."
- If a verb i s missing, then it is first assumed, that the utterance is a route description. There fore, this case frame is tested with verbs of the field "MOVE", otherwise the test is continued with verbs referring to location.
FURTHER WORK
Hitherto SYNC and SEMC are implemented together with the procedures to build up the KSs as well as the procedures for the interaction between the components. At present we begin to incorporate the pragmatic components, so we do not have confirmed results on the interaction between them and SYNC and SEMC.
The parsing formalism runs on a TR-440 at the University of Bielefeld. It is written in FLAVORS (BJ, an object-oriented language, embedded in MACLISP.
ACKNOWLEDGEMENTS
I would like to thank T. Christaller, D. Metzing and B. Terwey, without whose cooperation this work would not have been possible.
@inproceedings{ijcai1983p175,
title = {Syntax, Semantics, and Pragmatics in Concert: An Incremental, Multilevel Approach in Reconstructing Task-Oriented Dialogues},
author = {Gehrke, Manfred},
booktitle = {Proceedings of the Eighth International Joint Conference on
Artificial Intelligence, {IJCAI-83}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Alan Bundy},
pages = {721--722},
year = {1983},
note = {Natural Language},
url = {https://ijcai.org/proceedings/1983/175},
}
TY - CPAPER TI - Syntax, Semantics, and Pragmatics in Concert: An Incremental, Multilevel Approach in Reconstructing Task-Oriented Dialogues AU - Gehrke, Manfred T2 - Proceedings of the Eighth International Joint Conference on Artificial Intelligence, IJCAI-83 PB - International Joint Conferences on Artificial Intelligence PY - 1983 SP - 721 EP - 722 UR - https://ijcai.org/proceedings/1983/175 ER -Copied
Machine-readable versions of this paper, for tools and AI agents:
Welcome
Future Conferences
AI Journal Division
IJCAI Proceedings
About IJCAI
IJCAI Awards