JPH07160717A - Device for recognizing relation between adjacent speeches - Google Patents

Abstract

PURPOSE:To provide the device for recognizing relation between adjacent speeches which recognizes relations between adjacent speeches. CONSTITUTION:A speech semantic contents extracting means 1 extracts semantic contents of speeches consisting of the declaration object, the sentence pattern, functions of speeches, etc., from two inputted speeches. A speech premise extracting means 2 extracts the premise of speeches from semantic contents of speeches. A means 5 for recognizing relations between adjacent speeches refers to vocabulary knowledge derived from the concept structure of events stored in a vocabulary knowledge storage means 3, the structure where relations between the preceding speech and the succeeding speech stored in a storage means 4 of the structure of relations between adjacent speeches are described, and semantic contents and the premise of input speeches to recognize the relations between adjacent speeches.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for recognizing a relationship between adjacent utterances, and more particularly to a device used in a question answering system or the like using a natural language to process a dialogue in a natural language.

[0002]

2. Description of the Related Art Conventionally, there is no technique for recognizing the relationship between adjacent utterances. For example, the region information is assumed on the basis of the plan recognition of the speaker throughout the dialogue as disclosed in Japanese Patent Laid-Open No. 1-233619. There was a method of recognizing the content of dialogue by analysis using sufficient information including. This method classifies utterance types into "request", "response", "confirm", etc., and dynamically changes the plan according to the utterance type. The relationship between the "response" and "confirmation" utterances is recognized as a related utterance. The plan is, for example, an GET-VALUE-UNIT plan that includes an ASK-VALUE plan asking for the value of a certain thing and an INFORMATION-VALUE plan that conveys the value of the certain thing to the other party as an interaction plan that represents the exchange of information. There is.

[0003]

In the above method, only the relation between the "request" utterance and the corresponding "response""confirmation" utterance is recognized as a related utterance,
It is not intended to recognize the relationship between adjacent utterances. In addition, this method often misses utterance pairs that cannot be recognized at first glance. As an example, consider the following set of utterances. Where is A1 Kojien? B2 It's sold out now. In the utterance of B2, since the place of Kojien, which is asked for the value in the utterance of A1, is not communicated, the above-mentioned GET-
This related utterance cannot be recognized using the VALUE-UNIT plan.

The object of the present invention is to recognize the relationship between any adjacent utterances.

[0005]

In order to achieve the above-mentioned object, the present invention comprises utterance meaning content extraction means for extracting the meaning content of an utterance consisting of a statement object, sentence pattern, utterance function, etc. from an input utterance, Utterance premise extraction means for extracting the utterance premise from the utterance meaning content, vocabulary knowledge storage means, adjacent utterance relation structure storage means, utterance meaning content, utterance premise, vocabulary knowledge, and adjacent utterance knowledge With reference to the relationship structure,
And an adjacent utterance relation recognition means for recognizing the relation between adjacent utterances.

[0006]

With the above structure, the meaning and premise of the utterance are extracted from the input utterance set, and the adjacent utterance relation is recognized by referring to the information, the vocabulary knowledge, and the adjacent utterance relation structure. You can

[0007]

1 is a block diagram showing the configuration of an embodiment of the present invention.

First, the configuration of an embodiment of the present invention will be described with reference to FIG. The utterance meaning content extraction unit 1 extracts the meaning content of an utterance including an assertion target, a sentence pattern, a utterance function, and the like from the input utterance. Based on the information about the case structure of the word or phrase that is the center of the utterance, the assertion target is converted into a description format based on a list of predicate names, argument roles, and their value pairs. For example, as for the question utterance A1 described above, exist ((S Kojien) (P?)) Is a statement target. Here, S and P represent that the roles of the arguments are the subject and the place, respectively. The sentence pattern is indicated by any one of a plain text, an interrogative sentence, an imperative sentence, and a request sentence. The function of utterance is indicated by information transmission, information request, action request, and the like.

The following predicates content, type,
For each of the utterances Utto, utterance Utt represents that the utterance assertion target is Ucontents, the sentence pattern is Utype, and the utterance action is Uforce. contents (Utt, Ucontents) type (Utt, Utype) force (Utt, Uforce) The utterance premise extraction unit 2 extracts the utterance premise from the meaning content of the utterance obtained by the utterance meaning content extraction unit 1. FIG. 2 shows classification of utterance premise. Below, each premise is demonstrated.・ Presence of existence. The actual existence of the individual object definitely described in the utterance. -Narrative premise. In the utterance, the event that is the background of the utterance. In a question sentence, it is an event described other than the focus of the question.
The contents of the hearing correspond to the narrative premise. · Relevance of respondent knowledge. The responder knows the direct answer to the question. • Suitability for relationship with context. The target in the semantic content of the utterance can uniquely correspond to the target in the preceding conversation or the target of the utterance situation.・ Premises introduced by "mo", "even", "but", "suru", "dare", and "to". Something else that is similar to the one collected by the particle, and that satisfies the content of the question sentence.・ Premise introduced by “ha”. Something else that is similar to the one collected by the particle and does not satisfy the content of the question sentence.・ The premise to be introduced by “only”, “shika”, “just” and “only”. Others, similar to those picked up by particles, do not satisfy the content of the question text.・ Premise introduced by “even”, “up”, and “mo”. It is unlikely that the event represented by the meaning of the utterance will occur.

The following predicates prepr_log, prep
r_pra and prepr_int are utterances Ut, respectively.
In t, the semantic premise is Upprepr_log, the pragmatic premise is Upprepr_pra, and the premise leading the questioner's intention is Upprer_int. prepr_log (Utt, Upprer_log) prepr_pra (Utt, Upprer_pra) prepr_int (Utt, Upprer_int) The vocabulary knowledge storage unit 3 stores a conceptual structure of an event associated with a phrase representing an event. FIG. 3 shows a conceptual structure of events that describes causal relationships between events and relationships in time series.

For example, the case structure information of the word "return" representing an event is return ((S person 1) (O thing) (OT person 2)
(At T)). Here, O, OT, and T represent that the roles of the arguments are target, goal, and time, respectively. As an incidental event to this, FIG. 4 shows a conceptual structure of an event “return book” in which an object “book” is applied to a target. have, move
An event whose predicate is -1, read is a basic event.
(ΑX−β) represents that the case element α in the basic event is the case element β of the event “return book”. T (np)
N in represents that the event is the nth preceding event of the event "return book", and n in T (ns) is the nth subsequent event of the event "return book". It means that. Further, n in T (n) represents that the event is the nth basic event among the basic events constituting the event "return book".

As an example of vocabulary knowledge derived from the conceptual structure of this event, the following predicates cause, pre_cond
Respectively, the event Q is the successor state of the event P, that is, the event P causes the event Q, and the event P is the event Q.
It is a precondition of. cause (P, Q) pre_cond (P, Q) The adjacent utterance relation structure storage unit 4 stores a structure describing the relationship between the preceding utterance Pre and the subsequent utterance Suc. First,
The relation pair between two utterances is the relation I between the preceding utterance and the utterance I.
Inference ref_w leading the original response Org via nf_kn
It is described by a two-stage inference including hat and inference ref_how that guides a subsequent utterance from the original response via the inter-utterance relation Inf_rh. In the following, various relational structures will be described using the Horn clause format of the programming language Prolog. pair (Pre, Suc):-ref_what (Pre, Org, Inf_kn),
ref_how (Org, Suc, Inf_rh). A description example of the relational structure between the preceding utterance Pre and the original response Org is shown below. (W1) Normal response. Give direct answers to questions as they are. ref_what (Pre, Org, normal response):-force (Org, information transmission), contents
(Pre, P), contents (Org, Q), o
r (match (P, Q), match (P, not
(Q))). Here, match (P, Q) indicates that the event P and the event Q are substantially equal to each other, and specifically, indicates that the predicate and the argument in each event can be compared with each other while allowing similarities. A computer using the programming language Prolog can be realized by unification, which is one of the operations of Prolog. (W2) Misunderstanding pointed out. Tell them that the semantic premise of the preceding utterance is not met. ref_what (Pre, Org, misunderstanding):-force (Org, information transmission), prepr_log
(Pre, P), contents (Org, Q), m
match (not (P), Q). A description example of the relationship structure between the original response Org and the subsequent utterance Suc is shown below. (H1) Direct expression. Inference r that guides subsequent utterances from the original response
ef_how is an inference that is a rhetorical development of how to convey the original response to the dialogue partner, and when nothing is inferred, it becomes the following direct expression. ref_how (Org, Suc, direct expression):-force (Org, Forse), force (S
uc, Sforce), and Force = Sforce,
contents (Org, Q), contents
(Suc, R), match (Q, R). Also, the rhetorical expansion can take the following structure so that it is repeated any number of times. Mid is response information in the intermediate stage of inference. In practice, the rhetorical expansion is limited to two times as a practical allowable range, and therefore the following structure is applied only once. ref_how (Org, Suc, [Inf_rh1,
Inf_rh2]):-ref_how (Org, Mid, Inf_rh1),
ref_how (Mid, Suc, Inf_rh2). (H2) Reason. Communicate the reason for the original response. ref_how (Org, Suc, reason):-force (Suc, information transmission), contents
(Org, Q), contents (Suc, R), o
r (cause (R, Q), pre_cond (not
(R), not (Q))). The adjacent utterance relation recognition means 5 recognizes the adjacent utterance relation by referring to the meaning content of the utterance, the utterance premise, the vocabulary knowledge, and the adjacent utterance relation structure.

A specific operation of one embodiment of the present invention will be described using the utterances A1 and B2 described above.

The utterance semantic content extraction means 1 extracts the following information as the semantic content of the input two utterances. contents (A1, exist ((S Kojien)
(P?))) Contents (B2, sold_out ((S Kojien) (P bookstore))) force (B2, information transmission) The utterance premise extraction unit 2 extracts the meaning content of the preceding utterance A1.
The following information is extracted as a premise of existence. prepr_log (A1, exist ((S Kojien) (P bookstore))) The adjacent utterance-to-speech relationship recognition means 5 has the meaning and premise of the utterance, and the concept of the event "sold out" stored in the vocabulary knowledge storage means 3. Lexical knowledge derived from structure cause (sold_out ((S object) (P
Store)), not (exist ((S thing) (P
The adjacent utterance relation is recognized as follows by referring to the store)))) and the two adjacent utterance relation structures stored in the adjacent utterance relation structure storage means 4 as to the misunderstanding indication and the reason. ref_what (A1, Org, misunderstanding):-force (Org, information transmission), prepr_log
(A1, P), contents (Org, Q), ma
tch (not (P), Q). ref_how (Org, B2, reason):-force (B2, information transmission), contents (O
rg, Q), contents (B2, R), caus
e (R, Q). However, in the above relational structure, P: = exist ((S Kojien) (P store)) Q: = not (exist ((S Kojien) (P
Store))) R: = sold_out ((S Kojien) (P store)). Therefore, pair (A1, B2):-ref_what (A1, Org, misunderstanding), ref
_How (Org, B2, reason). Therefore, the relationship between adjacent utterances of the preceding utterance A1 and the subsequent utterance B2 can be recognized as a misunderstanding and a reason.

As a second embodiment of the present invention, the concept relation storage means 6 for storing the concept relation between the objects represented by the words and phrases.
And the area knowledge storage means 7 for storing knowledge depending on the area including the plan goal structure, and the adjacent utterance relation recognition means 5 recognizes the adjacent utterance relation by also referring to the conceptual relationship and the area knowledge. FIG. 5 is a block diagram showing the configuration of the embodiment. The second embodiment will be described below.

The conceptual relationship storage means 6 stores the conceptual relationship between the objects represented by the phrases. As knowledge derived from the conceptual relationship between the objects, the following predicate same_categor
y and is_a represent that the target A and the target B are direct subordinate concepts of the same target, and the target A is a subordinate concept of the target B, respectively. same_category (A, B) is_a (A, B) By introducing such a predicate is_a, the relation structure between the adjacent utterances and the following utterance can be described in the adjacent utterance relation structure storage means 4. It will be possible. (H3) Grounds for judgment. Communicate the basis for determining the content of the original response. ref_how (Org, Suc, basis of judgment):-force (Suc, information transmission), contents
(Org, Q), contents (Suc, R), o
r (is_a (Q, R), cause (R, Q), pr
e_cond (Q, R)). The area knowledge storage means 7 stores knowledge depending on the area including the plan goal structure. As the knowledge derived from the plan goal structure, the following predicate subplan is the event P
Represents that it is a subplan that constitutes the event Q. subplan (P, Q) By introducing such a predicate subplan,
The relationship structure between the adjacent utterances can be described in the relationship structure storage means 4 between the following utterances and the original response. (W3) Progress report. Planning is carried out with the goal of obtaining answers to questions, and the progress is reported. ref_what (Pre, Org, progress report):
-Force (Org, information transmission), contents
(Pre, P), contents (Org, Q), s
ubplan (Q, P).

[0017]

As described above, according to the apparatus for recognizing the relationship between adjacent utterances of the present invention, the semantic content and premise of the input utterance are extracted, and the information, the structure of the relationship between adjacent utterances, and the vocabulary knowledge are generally recognized. By referring to the knowledge, it is possible to recognize an arbitrary relationship between adjacent utterances.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is a diagram showing the classification of the premise of a question sentence.

FIG. 3 is a diagram showing a conceptual structure of an event.

FIG. 4 is a diagram showing a conceptual structure of an event “return book”.

FIG. 5 is a block diagram showing a configuration of a second exemplary embodiment of the present invention.

[Explanation of symbols]

 1 utterance meaning content extraction means 2 utterance premise extraction means 3 vocabulary knowledge storage means 4 adjacent utterance relation structure storage means 5 adjacent utterance relation recognition means 6 conceptual relation storage means 7 domain knowledge storage means

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Tadahiro Kitahashi 3- 1-43 Uenonishi, Toyonaka, Osaka

Claims (4)

[Claims] 1. An utterance meaning content extraction means for extracting the meaning content of an utterance consisting of a statement object, a sentence pattern, a utterance function, etc. from two input utterances, and an utterance premise for extracting an utterance premise from the meaning content of the utterance. Extraction means, vocabulary knowledge storage means for storing a conceptual structure of an event associated with a phrase representing an event, adjacent utterance relation structure storage means for storing a structure describing a relationship between adjacent utterances, and meaning of the utterance Between adjacent utterances, the content, the premise of the utterance, the vocabulary knowledge, and the adjacent utterance relation recognizing means for recognizing the adjacent utterance relation by referring to the adjacent utterance relation structure. Relationship recognizer. 2. A concept relation storing means for storing a concept relation between objects represented by phrases, wherein the adjacent utterance relation recognizing means also recognizes the adjacent utterance relation by also referring to the conceptual relation. The apparatus for recognizing the relationship between adjacent utterances according to claim 1. 3. A region knowledge storage unit for storing knowledge dependent on a region including a plan goal structure, wherein the adjacent utterance relation recognizing unit also refers to the region knowledge to recognize an adjacent utterance relation. The adjacent utterance relation recognition device according to claim 1 or 2. 4. A dialogue structure management means for dynamically managing a dialogue structure represented by a tree structure based on a plan goal structure, a question utterance, and a number corresponding to the question utterance with reference to the dialogue structure. 4. The apparatus for recognizing the relationship between adjacent utterances according to claim 1, further comprising utterance pair recognition means for recognizing an utterance pair consisting of the response utterance.