JP3181465B2 - Language processor - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a language processing apparatus for performing a semantic analysis of a sentence input in a natural language in a speech dialogue system, a natural language interface, and the like.

[0002]

2. Description of the Related Art It is highly desired that a computer can be operated by a natural language such as voice as a natural and easy-to-use interface for ordinary users. 2. Description of the Related Art Conventionally, in order to interpret a sentence input by a user on a computer, techniques such as syntax analysis and semantic analysis have been used as typical language processing techniques.

[0003] In order to process a natural language sentence by a computer, a technique for recognizing words and clauses constituting the sentence and analyzing and defining the sentence structure when they constitute the sentence is described in, for example, The Institute of Electronics, Information and Communication Engineers, edited by Hirosato Nomura, "Basic Technology of Natural Language Processing", first edition (March 30, 1988) [
Reference 1], edited by IEICE, edited by Makoto Nagao,
"Japanese Information Processing" 3rd Edition (May 30, 1987) [Hereafter,
Numerous technologies have been devised as described in [2]. Among them, a technique that is relatively frequently used to construct a dialogue system is a parsing technique based on a regular grammar or a context-free grammar.

In these techniques, it is customary to prearrange the arrangement (sentence pattern) for expressing a sentence of words, phrases, and the like according to syntax rules. The specified syntax rules are converted into a finite state network or a table for parsing and used by a parser.

In the syntactic analysis processing, first, as preprocessing for analyzing the structure of a sentence, a morphological analysis is performed on an input sentence, and constituent elements of the sentence such as words and phrases are extracted. By morphological analysis, information such as the part of speech and the use of the constituent elements of a sentence can be obtained. Next, the sequence of the component elements is parsed using the information such as the part of speech and utilization of the component elements and the information of the syntax rules. If a given sequence of components is accepted in a prescribed sentence pattern, the qualification relationship between components,
The sentence structure such as the relationship between the subject and the predicate is output as a result of the syntax analysis. If the given sequence of components is arranged in a manner other than the predefined sentence pattern, the parsing fails. In this case, it is determined that a sentence not accepted by the interactive system has been input. The above is the first related art.

In a system in which a user and a computer work interactively, there is a need for a semantic analysis technique for converting the meaning of the work that the user wants the computer to execute into a semantic expression format executable by the computer. is there.

For semantic analysis technology, see Reference 1 (p. 229)
−272) and Reference 2 (p. 141−166), the semantic analysis method using case frame format that expresses the co-occurrence relationship between verbs and noun phrases or prepositional phrases,
Generally used widely.

FIG. 19 is a block diagram showing a typical process of semantic analysis using this case frame. The input unit 1 converts information that the user wants to input into the computer into a sentence expressed in characters, and outputs the sentence to the morphological analysis unit 2. The morphological analysis unit 2 is a grammar that defines morphological information such as headwords, parts of speech, and inflections of words described in the vocabulary knowledge storage unit 3 and how words are connected to form extraction units such as phrases. Based on the rules, the input sentence is subjected to morphological analysis, and the extracted words and phrases are output to the meaning extracting means 9 together with the morphological information. Meaning extraction means 9
Refers to the semantic feature knowledge storage means 5 which defines the correspondence between the word and the semantic feature expressing the semantic concept of the word,
After giving a semantic feature to the phrase with morpheme information input by the morphological analysis means 2, the case frame knowledge storage means 1
The meaning expression of the sentence is extracted by using the case frame defined as the meaning expression format of the sentence at 0.

The case frame in this method is used as a semantic structure pattern expressing the meaning of the entire sentence. A plurality of case frames are defined for each verb usage, and are composed of a plurality of slots representing case elements. A slot is a portion in which a noun phrase or the like constituting a sentence is embedded, and is described for each case element of a verb such as an action nominative case, a target case, and a place case. A noun phrase embedded in a slot is called a filler. As the filler, a short unit up to about a word or a phrase is usually used, and each filler is given a semantic feature using a semantic dictionary that describes the correspondence between words and semantic features. Also, each slot has
The semantic features and prepositions that are allowed for the filler embedded therein are set.

[0010] Based on the above semantic expression, fillers such as words and phrases extracted from the input sentence are embedded in slots whose semantic features and prepositions match (slot filling). In this way, a case frame in which all fillers that can form a sentence are embedded in a slot is extracted, and this is used as a result of extracting a semantic structure expressing the meaning of the entire sentence. The above is the second related art.

[0011]

However, the above-mentioned language processing technology has the following problems.

(1) A sentence uttered by a general user includes a very large number of phrases, and particularly in the case of Japanese, the order between phrases is relatively free compared to other languages, and the subject Omission of particles and particles also occurs on a daily basis. For this reason, it is considered that there are enormous types of sentence patterns input to the dialogue system. Therefore, in the first related art in which it is required to explicitly define the sentence structure for each sentence pattern, it has been difficult to define the sentence structure so as to sufficiently cover various types of input sentences.

(2) On the other hand, in the second conventional technique, only the semantic relation between the components is extracted by the case frame without specifying the arrangement of the components. In this method, the sentence pattern is not explicitly defined, and even if the sentence pattern is different, it is regarded as the same if the semantic relationships between the components are equal. Many types can be expressed. However, for each verb usage, all combinations of fillers that can constitute a sentence are subject to semantic verification,
Semantic ambiguity generated so many semantic interpretation hypotheses that semantic verification was inefficient.

Therefore, the present invention provides a language processing apparatus which has a sentence pattern expression capability capable of sufficiently covering various expressions contained in a general user input sentence and which can efficiently perform semantic verification of the entire sentence. The purpose is to provide.

[0015]

According to a first aspect of the present invention, there is provided a language processing apparatus comprising: an input unit for inputting a sentence expressing a natural language; A morphological analysis unit that outputs candidates such as words, a semantic feature knowledge storage unit that stores semantic feature information for each word or phrase, and a semantic relationship between words and phrases are described. The semantic feature is given to each candidate output by the morphological analysis means by referring to the concept element knowledge storage means, which is expressed in a frame structure as a component and stored as a concept element, and the semantic feature knowledge storage means, For a candidate group forming a part, a concept element extracting means for extracting a concept element by referring to the concept element knowledge storage means, and a semantic expression of the entire sentence as a combination of the conceptual elements and being recognized as a semantic expression of the entire sentence Suta A sentence sense determination knowledge storage unit that describes conditions for a combination of concept elements, and a concept element stored in the sentence sense determination knowledge storage unit for a combination of concept elements extracted by the concept element extraction unit. Sentence meaning determination means for outputting a combination of concept elements that meet a condition for the combination , wherein the language processing device stores the concept element knowledge storage
Means require multiple slots to define conceptual elements
Slots and non-essential slots are classified and stored.
The element extraction means determines if the candidate is registered in the required slot.
Only in this case, the concept element is output .

[0016]

In the language processing apparatus according to a second aspect of the present invention, the morphological analysis means gives positional information in the sentence as candidate positional information, and the conceptual element extracting means determines that a candidate constituting a certain conceptual element is an input sentence. The feature element is characterized in that the concept element is output only when it is one continuous section.

In the language processing apparatus according to a third aspect of the present invention, the sentence meaning determination knowledge storage means stores knowledge of a positional order relationship between candidates constituting a certain concept element, and the sentence meaning determination means includes: It is characterized by selecting and outputting a combination of concept elements including only the concept elements satisfying the order relation of the positions among the candidates among the combinations of concept elements input from the concept element extraction unit.

According to a fourth aspect of the present invention, there is provided a language processing apparatus, comprising: an intention extraction knowledge storage unit for storing the knowledge of the correspondence between the intention and the predicate in the candidate; Intention extraction that extracts a predicate from among the candidates input from, determines the intention corresponding to the extracted predicate using knowledge of the correspondence between the intention and the predicate, and outputs the intention to the sentence meaning determination unit. Means, the sentence meaning determination knowledge storage means stores a combination of concept elements classified according to intention, and the sentence meaning determination means includes a combination of the concept elements input from the concept element extraction means. It is characterized in that only a combination of concept elements corresponding to the intention is selected and output.

According to a fifth aspect of the present invention, there is provided a language processing apparatus comprising semantic feature relationship knowledge storing means for storing knowledge of a relationship between semantic features, wherein the intention extracting means is configured to select one of the candidates input from the concept element extracting means. When there is no predicate, the intention is determined using the semantic features assigned to the candidates and the knowledge of the relationship between the semantic features, and the intention is output to the sentence semantics determination means.

According to a sixth aspect of the present invention, in the language processing apparatus, the concept element extracting means refers to the knowledge of the correspondence between the concept element and the intention stored in the sentence meaning determination knowledge storage means, and And extracting only the concept elements corresponding to the intention determined in (1) and outputting the combination of the concept elements to the sentence meaning determination unit.

According to a seventh aspect of the present invention, in the language processing apparatus, the sentence meaning determination knowledge storage means stores linguistic knowledge for candidates constituting a concept element, and the sentence meaning determination means stores the linguistic meaning in the stored language. The linguistic validity of the concept elements is evaluated based on the statistical knowledge, the likelihood of the combination of the concept elements is calculated, and the combination of the concept elements having the maximum likelihood is output.

[0023]

Next, the operation of the present invention will be described.

FIG. 5 is a diagram showing an example of adding semantic feature information.
FIG. 6 is a diagram showing an example of a frame structure expressing a concept element,
FIG. 7 is a diagram showing an example of a semantic understanding result expressed by a combination of extracted conceptual elements, FIG. 8 is a diagram showing an example of a conceptual element frame expressing “place (point)”, and FIG. the first aspect of the present invention
FIG. 10 is a diagram showing an example of an erroneous semantic comprehension candidate for explaining a modified example of FIG. 10, FIG. 10 is a diagram showing an example of setting an essential slot in a conceptual frame, and FIG.
FIG. 12 is a diagram showing an example of a conceptual element frame expressing
Illustrates an example of a wrong meaning understood candidates for explaining the second invention, FIG 13 is a view showing an example of a semantic understanding candidates for explaining the third invention, Figure 14, the third invention FIG. 15 shows an example of an incorrect semantic understanding candidate for explanation.
FIG. 16 is a diagram showing the relationship of semantic features expressing the semantic world (hotel reservation) of the target conversation, and FIG. 16 is a diagram showing an example of a procedure for extracting intention. FIG. 17 is a diagram showing an example of a penalty rule based on a syntactic viewpoint, FIG. 18 is a diagram showing an example of a penalty rule based on a semantic viewpoint, and FIG.
Is a configuration diagram illustrating an example of a second conventional technique.

For example, if the sentence “Please tell us a hotel with a beautiful pool near Shonan Coast” is input, the morphological analysis means will use “Shonan Coast”, Extract “nearby”, “beautiful”, “in the pool”, “along”, “hotel”, “tell me” and so on. These phrase candidates are given semantic features as shown in Fig. 5, and by applying the concept element frame shown in Fig. 6, the concept "shonan coast" and "near by" represent a place (far and near). As a whole,
Similarly, “beautiful”, “of a pool”, “a”, “at a hotel”, etc. are grouped as conceptual elements representing facilities.
By combining a combination of these two conceptual elements and a verb “Tell me please” representing the intention of the input sentence, a semantic understanding result as shown in FIG. 7 is obtained.

As described above, according to the first aspect of the present invention, a partial semantic unit of an input sentence can be extracted as a concept element. The order of appearance of conceptual elements is free, so for example, input sentences in word order such as "Tell me a beautiful pool hotel, near the Shonan coast" or "Teach me, a beautiful pool hotel, Shonan An input sentence in word order such as "near the coast" can be analyzed without explicitly defining such a word order sentence pattern. Therefore, the sentence pattern can be expressed more efficiently than in the first related art in which it is difficult to define the input sentence so as to cover the wording sufficiently.

In the case of a concept element, only phrase candidates constituting a sentence part are subjected to meaning extraction, and the meaning of the entire sentence is represented by a combination of concept elements. For this reason, the case frame defined to cover the usage of a certain verb is applied to all phrase candidates constituting a sentence directly, compared to the second conventional technique.
The number of semantic comprehension candidates resulting from semantic ambiguity can be reduced, and semantic verification can be performed efficiently.

Next, in the semantic understanding candidate of the entire sentence output according to the present invention, not only a correct semantic understanding result but also an erroneous semantic competing candidate may be generated at the same time.

For example, when a conceptual element frame representing a place (point) as shown in FIG. 8 is defined, "on Shonan coast" and "near by" are determined as conceptual elements representing a place (perspective). As shown in Fig. 9, “Shonan Kaigan” is a conceptual element of a place (point), and “Nearby” is determined as a conceptual element of a place (perspective). Incorrect semantic competing candidates are generated.

At this time, in the first invention, FIG.
As shown in (3), among the slots in the conceptual element frame, the slots that must be filled in order for the conceptual element to be satisfied are specified as required slots, and the conceptual elements that do not include the required slots are rejected. By doing so, it becomes possible to reject an erroneous semantic competing candidate.

If a conceptual element frame representing an atmosphere (hotel) as shown in FIG. 11 is defined,
As shown in Fig. 2, two clauses, "beautiful" and "hotel", which are separated from each other in the input sentence, are collected and extracted as concept elements that define the atmosphere of the hotel, and are erroneous semantic competing candidates. May be generated.

At this time, in the second invention, the erroneous meaning comprehension as shown in FIG. 11 is performed by using the constraint that the phrase constituting the concept element is positioned in the input sentence. Can be rejected.

In the case of the example sentence "I want to reserve a hotel conference room near the station" as shown in FIG. 13, the two phrases "at the station" and "at the vicinity" represent a concept representing a place (perspective). In addition, the two phrases “hotel” and “meeting hall” are grouped as concept elements representing facilities. However, as shown in FIG. 14, the "hotel" which should be included in the qualifying slot of the conceptual element representing the equipment
May be included in the qualifying slot of the conceptual element representing the place (perspective) with the positional order reversed, and an erroneous semantic competing candidate may be generated.

At this time, in the third invention, there is no “hotel” clause that modifies “nearby” after the “nearby” clause. Using the knowledge, it becomes possible to reject an erroneous semantic competing candidate as shown in FIG.

By the way, the work requested by the user from the system includes, for example, reservation, change, WH question (when, where,
Questions such as how many, etc.), YN questions (questions for answering yes or no), and information retrieval can be classified. If these are set as intentions that the input sentence means, and predicates that are deeply semantically related to the intentions are defined in advance as keywords, the intentions can be determined from the keywords in the input sentence, and the tasks requested by the user Thus, appropriate system operation can be performed. For example,
"I'm looking for a cheap hotel near Yokohama station."
Input sentence such as "I'm looking for it."
Is stored as a keyword indicating an information search, when this sentence is input, the system determines the intention as the information search. Then, it is possible to perform information search for hotels that meet the conditions such as “(close to Yokohama Station)” and “(price) inexpensive” which are separately extracted as concept elements.

However, when an intention and a conceptual element are extracted from an input sentence, a contended candidate with a meaning understanding in which the combination of the intention and the conceptual element is semantically inappropriate may be generated. For example, if the intention is a WH question about the travel time, the candidate with the meaning of "Yokohama Hotel is [concept element: hotel name], how long is it walking from the station [concept element: traffic (time required)]" is semantic Correct. However, "On Monday next week [
[Date], how long is it walking from the station?
(Required time)], the combination of conceptual elements representing date and time and required time may be semantically inappropriate for the intention of the WH question.

At this time, in the fourth invention, a combination of concept elements that semantically correctly corresponds to a certain intention is classified and stored in the sentence meaning determination knowledge storage means. The determination means can determine that the concept element representing the date and time does not hold under the condition of the WH question regarding the required time, and can reject an erroneous semantic competing candidate.

When an intention is extracted from an input sentence, a predicate set in advance as having a deep semantic relationship with the intention may be omitted in the input sentence, and the intention may not be determined. For example, "A cheap hotel close to Yokohama station"
When the predicate of the keyword does not exist in the sentence as in the input sentence, the intention cannot be determined.

At this time, in the fifth invention, as shown in FIG. 15, the abbreviated intention is inferred using the knowledge expressing the semantic world of the target dialogue sentence as the relation of the semantic feature, and the intention extraction is performed. Can be determined even when the predicate for does not exist in the input sentence.

The combination of concept elements input from the concept element extraction means to the sentence meaning determination means includes a correspondence between the intention and the combination of concept elements defined in the sentence meaning determination knowledge storage means. A concept element that satisfies and a concept element that does not satisfy are mixed.

At this time, in the sixth invention, by using the intention information in the concept element extracting means, it is possible to reject an inappropriate concept element candidate which does not correspond to the intention before being input to the sentence meaning determination means. become. As a result, concept element candidates that do not correspond to the intention can be rejected by the concept element extraction means, and the concept element extraction processing can be performed efficiently.

By the way, when a plurality of conditions relating to the establishment of the above conceptual elements are used at the same time, and only the correct meaning understanding candidate is obtained by rejecting the meaning understanding candidate which does not meet the condition, the meaning which satisfies all the conditions is obtained. If there is no understanding candidate, the final semantic understanding result will not be output. For example, a user tries to say, "I'm looking for a place near Yokohama station for about 10,000 yen per night" and say, "I'm near a station and looking for about 10,000 yen per night in Yokohama." There are cases. If the condition of the positional order relationship between the slots in claim 3 is “Yokohama”, which modifies “to a station”
Does not allow him to be behind "at the station"
The part of “close to the station, in Yokohama” cannot be extracted as one conceptual element, and a correct meaning understanding result cannot be obtained.

At this time, in the seventh invention, instead of rejecting semantic comprehension candidates that do not meet the condition, for example, a penalty is added one point each time the condition is violated, and the candidate with the least penalty is finally determined. Output. As a result, it is possible to calculate the likelihood of the semantic understanding candidate from the viewpoint of a plurality of concept element formation conditions, and even if a sentence contrary to a certain concept element formation condition is input, the correct meaning is not rejected. Understanding results will be obtained.

[0044]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a language processing apparatus according to the present invention will be described below with reference to the drawings.

First, an embodiment according to the first invention will be described. FIG. 1 is a functional block diagram showing an embodiment of the present invention. In FIG. 1, an input unit 1 converts information that a user wants to input into a computer into a sentence expressed in characters, and outputs the result to a morphological analysis unit 2. When the user inputs a sentence to the computer using the keyboard, the input means 1 is a keyboard, and when the user inputs a sentence by voice,
Input means 1 is a speech recognizer.

The morphological analysis means 2 includes a vocabulary knowledge storage means 3
Based on vocabulary information such as headwords and parts of speech of words described in, and grammatical rules that define the connection between words and words, morphological analysis is performed on the input sentence, and possible phrase candidates that can appear Are extracted from the input sentence, and a set of those clause candidates is extracted from the concept element extracting means 4 in the meaning extracting means 9.
Output to

The concept element extracting means 4 stores semantic feature knowledge storing means 5 which describes the correspondence between independent words and semantic features.
, A semantic feature corresponding to the independent word in the phrase is given to each phrase candidate input from the morphological analysis means 2. A connectable combination is selected in temporal order from the phrase candidates given semantic features, and each phrase candidate in the combination and a certain concept element frame defined in the concept element knowledge storage unit 6 are selected. Verify slot compatibility. As a result of the verification, if it matches, the phrase candidate is registered in the slot. If not, the same verification is performed for the other conceptual element frames. When the verification of the compatibility with the slot has been completed for all the selected phrase candidates, all the concept element frames having the slots in which the phrase candidates are registered are extracted from the combination of the phrase candidate sequences constituting the sentence. The combination is output to the sentence meaning determination unit 7 as concept elements.

The sentence meaning determination means 7 includes the concept element extraction means 4
It is determined whether or not the combination of concept elements input from satisfies the condition for the combination of concept elements described in the sentence meaning determination knowledge storage means 8, and the combination that satisfies the condition is determined as the meaning of the entire sentence. Is output as a semantic understanding result expressing.

In the above configuration, the operation of the present invention will be described with an example of an input sentence "Tell me a hotel with a beautiful pool near the Shonan coast". The morphological analysis means 2 recognizes the phrase candidates that can appear in the input sentence as “Shonan coastal”, “Nearby”, “Pretty”, “Pool”, “Always”, “Hotel”, “Tell me”
And so on. In addition, the same meaning word, a phrase candidate due to a recognition error in the case of speech input (for example, “stairs” for “shore”), and the like are also extracted. Part-of-speech information such as "Shonan Coast (noun), no (case particle)" and information on the temporal position of the phrase candidate in the input sentence are also extracted from these phrase candidates.

A set of these phrase candidates is input to the concept element extraction means 4, and the information on the independent words in each of the phrase candidates and the correspondence between the independent words and the semantic features stored in the semantic feature knowledge storage means 5. , Semantic features are given to the respective phrase candidates as shown in FIG. This semantic feature is used to verify the semantic constraints described in each slot of the conceptual frame, and whether a phrase candidate can semantically satisfy a slot of the conceptual frame. Is used to determine

Subsequently, the concept element extracting means 4 applies a concept element frame to combinations of phrase candidates that can be connected in a temporal order, and combines the phrase candidates into units of concept elements. The specific definition of the concept element is described as a frame structure as shown in FIG. In the example of FIG. 6, as a basic concept that can appear in a sentence, a concept element frame describing a concept expressing the perspective of a place such as “close to a station” or “away from a downtown area”; A conceptual element frame that represents the existence of facilities, such as "there is a pool" or "there is a parking lot", is shown. In each conceptual element, a slot that can sufficiently cover the wording for expressing the conceptual element is described. In each slot, semantic features and particles are described as matching conditions to be satisfied by the filler. When registering a phrase candidate in a slot, a phrase candidate including a particle is registered in a slot in which the semantic feature and the particle match, and a phrase candidate in which the particle is omitted is registered in a slot in which the semantic feature matches.

FIG. 7 shows an example in which the phrase candidates of FIG. 5 are grouped into units of concept elements in this way, and the meaning understanding determination means 7 generates the meaning understanding candidates. In this example, the concept element
The expression of (the perspective of the place) is "on Shonan coast",
"Nearby," etc. are summarized, and expressions that represent conceptual elements (equipment) are "clean", "pool",
"Aru" and "Hotel" are included. Also, the predicate "Please tell me" is registered as a meaning understanding candidate as an expression representing intention. By combining these conceptual elements, an expression form of the meaning of the entire sentence can be obtained.

Next, a modified example according to the first invention will be described. In the slots that make up a conceptual frame,
A slot in which a phrase candidate must be registered in order for the concept element to be established is defined as an essential slot, and other slots are defined as non-essential slots. The designation of the essential slot is performed in the definition of the concept element frame in the concept element knowledge storage means 6. FIG. 10 shows an example of the definition of a conceptual element frame described in the conceptual element knowledge storage unit 6.
For example, in (1), fillers with * semantic features such as "far", "close", and "distant" are registered as mandatory slots of conceptual frames that represent perspective of places. The attribute slot to be specified is specified. In (2), a point slot is specified as a mandatory slot of a conceptual frame at a point of a place. When the concept element hypothesis is generated, the concept element extraction unit 4 checks whether or not a phrase candidate is registered in an essential slot of the concept element frame. If you are registered in the required slot,
The hypothesis of the conceptual element is output to the sentence meaning determination means 7, and the hypothesis of the other conceptual element is rejected.

For example, in the semantic comprehension candidate shown in FIG. 7, since the phrase candidate “nearby” is registered in the attribute slot, which is a required slot of the concept element frame representing the place (far and near), this concept element hypothesis Is determined to be satisfied, and the meaning understanding candidate is output to the sentence meaning determination means 7. However, in the semantic comprehension candidate shown in FIG. 9, the phrase candidate “Shonan Kaigan” is registered in the qualifying slot of the conceptual element frame representing the place (point). Since this decoration slot is a non-essential slot and cannot be established independently, this concept element hypothesis is determined as not being established and rejected. Other processes are the same as those of the first embodiment.

Next, an embodiment according to the second invention will be described. The morphological analysis unit 2 gives the information on the position of the phrase candidate in the input sentence to the phrase candidate as, for example, the start and end information of the phrase candidate. The concept element extraction means 4 refers to the position information of the phrase candidates registered in the slot of the concept element frame, and only when all the phrase candidates constituting the concept element are positionally consecutive, the concept element Is output to the sentence definition determination means 7.

This condition does not allow one input sentence to have a plurality of the same conceptual elements at distant positions. For example, in the combination of conceptual elements shown in FIG. 12, two phrase candidates “Shonan Kaigan” and “Hotel”, which are distant from each other, are put together in a conceptual element frame representing an atmosphere (hotel). This hypothesis is interpreted as "Please tell me a hotel on the Shonan coast near (something) with a beautiful pool." This semantic understanding result is not semantically valid as compared with the result of FIG.
Such a misinterpretation candidate whose meaning is understood can be rejected by the present invention. Other processes are the same as those of the first embodiment.

Next, an embodiment according to the third invention will be described. Each slot of the conceptual element frame includes, for example, a nominative case (~, ~), a starting case (~, from ~),
Terminal case (to, to), target case (to), instrument case (to
), Adnominal case (adnominal phrase, adverb form of adverb), adjunct modifier (adverb form of adverb, adverb), declarative case (verb, adjective appearing in conceptual element), etc. Is defined for each Case properties are assigned to each slot.
The sentence meaning determination knowledge storage means 8 stores, as knowledge that defines the order relation of each slot of the conceptual elementary frame, "the adnominal qualified case exists before the qualified case (primary case, object case, etc.)". , "The nominative case, the starting case, the end case, the object case, the instrumental case, etc. exist before the narrative case" in the form of rules. The sentence meaning determination means 7 applies the rule of the slot order relation to the combination of concept elements input from the concept element extraction means 4, and rejects semantic comprehension candidates including concept elements that do not conform to the rule. I do.

For example, in the case of an input sentence "I want to reserve a hotel conference hall near a station", the correct meaning understanding result is as shown in FIG. At this time, as shown in FIG. 14, a semantic understanding candidate may be generated in which the phrase candidates “near” and “hotel” are put together as conceptual elements representing the perspective of a place. This means that the "hotel"
In this example, the temporal order of phrase candidates is reversed, and the meaning is interpreted as "near the hotel." In this case, the sentence meaning determination means 7 stores the rule that the union qualified case is stored in the sentence meaning determination knowledge storage means 8 in front of the case to be qualified (main case, target case, etc.). It is determined that this concept element candidate violates the rule by using the information on the temporal position held by the phrase candidates “hotel” and “nearby”, and the semantic understanding candidate in FIG. 14 is rejected. can do. The other processing is the first
This is the same as one embodiment according to the invention of the present invention.

Next, an embodiment according to the fourth invention will be described. In this embodiment, the intention of the input sentence can be extracted by providing the intention extraction knowledge storage means 11 and the intention extraction means 12 as shown in FIG.

The intention extraction knowledge storage means 11 stores knowledge of (1) intention, and (2) keywords associated with the intention (intention extraction keywords). The intention indicates the type of meaning of the entire sentence of the input sentence. For example, if the target conversational sentence relates to hotel reservation, there are utterance types such as reservation, change, cancellation, and guidance request. Information requests can be further categorized into WH questions such as where, when, how many, etc., YN questions that ask for yes, no as answers, and searches that require information from the database. . These types of utterance types are defined as intentions of the input sentence in hotel reservation. Further, in order to determine an intention from an input sentence, a predicate that is semantically related to each intention is defined as an intention extraction keyword. The intention extraction keywords are, for example, "reserve", "stay", "apply" if the intention is a reservation, and "search", "check", "teach" if the intention is a search. And so on. Each intention extraction keyword is given a corresponding intention as a semantic feature.

The intention extraction means 12 is based on the intention stored in the intention extraction knowledge storage means 11 and the knowledge of the intention extraction keywords, and is a phrase candidate sequence to which a semantic feature input from the concept element extraction means 4 is added. To determine the intention.

An example of a procedure for determining the intention is shown as a procedure when the predicate in FIG. 16 exists. If the intention extraction keyword included in the clause candidate string is a predicate indicating reservation, change, or cancellation, the intention is determined as reservation, change, or cancellation, respectively. If the predicate is a guidance request, the phrase column candidate is "
If it contains a word that indicates a WH question such as `` when '', the intention is a WH question.If there is no word that indicates a WH question and the final particle `` ka '' is included, the intention is a YN question. I do. Otherwise, the intention is to retrieve information from the database. According to the intention extraction procedure,
For example, in an input sentence such as "I'm looking for a cheap hotel near Yokohama station", the predicate "I'm looking for" is the intention extraction keyword of the guidance request, and the WH question sign, In addition, since the final particle "ka" is not included, the intention is determined by the search.

The sentence meaning determination knowledge storage means 8 stores combinations of concept elements that are permitted to be established under certain intention conditions as knowledge of the correspondence between combinations of intention and concept elements.
The expression of the correspondence between the intention and the conceptual element is based on a rule or a table. The sentence meaning determination means 7 refers to this knowledge and rejects a combination of concept elements whose correspondence with the intention is not defined.

As an example of the concept element, a place (point), a place (perspective), a facility (existence), a facility (use), a traffic (required time), a traffic (means), a cost, a dish, a date, a hotel name,
The number of people, room (type), room (number), atmosphere (landscape), atmosphere (hotel), user name, user address, etc. can be defined. Among these, for example, when the intention is to search for a hotel, the concept element representing the name and address of the user does not have the property as the attribute of the hotel, and therefore should not be included in the input sentence for searching for the hotel. Therefore, the intention of "hotel search" is to allow only the concept elements (location, cost, facilities, etc.) that represent the attributes of the hotel,
It is defined by the sentence definition determination knowledge storage means 8.

Also, for example, the intention is that the WH
In case of question, place (point), place (perspective), equipment (existence), equipment (use), traffic (time required), traffic (means)
, Hotel name, etc. are defined by the sentence definition determination knowledge storage means 8 as corresponding concept elements. At this time, for the meaning understanding candidate such as “[concept element: date and time] next Monday, how long is it walking from the station [concept element: traffic (required time)]”, the intention is required by the intention extraction means 12. Since it is determined as a WH question relating to time and the concept element: date and time is not defined by the knowledge of the correspondence between the intention and the concept element, such a semantically unsuitable semantic understanding candidate can be rejected. . Other processes are the same as those of the first embodiment.

Next, an embodiment according to the fifth invention will be described. In this embodiment, as shown in FIG. 3, by providing a semantic feature relationship knowledge storage unit 13, the intention can be determined even when a predicate does not exist in the phrase candidate sequence.

The semantic feature relationship knowledge storage means 13 defines the knowledge expressing the semantic world of the target dialogue by the relationship between the semantic features. For example, a hierarchical structure of semantic features as shown in FIG. 15 is defined as the relationship knowledge between the semantic features expressing the semantic world of hotel reservation. In this hierarchy, the required information to book a hotel is "Hotel Name"
"Date", "Room", "Number of people", etc. are defined, and the hotel attribute information required to determine the hotel name includes "Location", "Facility", "Cost", "Meal", "Transportation" , “Atmosphere”, etc. are defined.

When the predicate does not exist in the clause candidate sequence input from the concept element extracting means 4, the intention extracting means 12 compares the semantic feature assigned to the phrase candidate with the semantic feature relation knowledge storage means 13. The intention is determined using the stored knowledge of the relationship between the semantic features, and the intention is output to the sentence meaning determination unit 7.

An example of a procedure for inferring an intention is shown as a procedure in the case where the predicate in FIG. 16 does not exist. In this procedure, if the input phrase candidate sequence includes a phrase candidate to which the semantic feature of the reservation essential information is added, the intention is reserved. Otherwise, if there is a hotel name and there is information on hotel attributes, the intention is made a guide request. If there is a hotel name and there is no information on hotel attributes, the intention is made a reservation. If there is no hotel name and there is information on hotel attributes, the intention is set as a guidance request. If the intention is inferred as a guidance request, it is further determined to be one of a WH question, a YN question, and a search according to the procedure of (b) of 1. in FIG. The intention determined by the above procedure is output to the sentence meaning determination unit 7.

For example, "a cheap hotel close to Yokohama station"
In the case of an input sentence where the predicate is omitted, there is no clause candidate and hotel name for the reservation essential information, and the information of the hotel attribute includes “Yokohama (place)”, “cheap (cost)”, etc. The intention is inferred as a search because it is a request and does not include the WH question sign and the final particle "ka". Thus, the intention can be determined even if the predicate is omitted in the input sentence. Other processes are the same as those of the first embodiment.

Next, an embodiment according to the sixth invention will be described. In the present embodiment, as shown in FIG. 4, the concept element extracting means 4 uses the intention information output from the intention extracting means 12 so that the concept element extracting process can be performed efficiently. It is.

The sentence meaning determination knowledge storage means 8 stores knowledge of the correspondence between the intention defined in the intention extraction knowledge storage means 11 and the concept element defined in the concept element knowledge storage means 6 in a certain intention. Is stored as a condition that defines a concept element that is satisfied under. The concept element extracting means 4 is composed of an intention extracting means 12
With reference to the intention input from the above and the knowledge of the correspondence between the intention and the concept element stored in the sentence meaning determination knowledge storage means 8, only the concept element corresponding to the intention is selected, and Output to means 7. As a result, the concept element extraction unit 4 can reject a concept element candidate that does not correspond to the intention during the concept element extraction process, so that the processing in the concept element extraction unit 4 can be performed efficiently. Other processes are the same as those of the first embodiment.

Next, an embodiment according to the seventh invention will be described. The present embodiment evaluates the linguistic validity of the concept element and gives the likelihood to the semantic understanding candidate, so that even if a sentence contrary to the condition for establishing the concept element is input, the semantic understanding result can be output, A plurality of conflicting semantic comprehension candidates are ranked so that a semantic comprehension result can be uniquely determined.

The sentence semantics determination knowledge storage means 8 stores the condition of linguistic knowledge for the candidate group forming the concept element.
For example, (1) a penalty rule from a syntactic point of view and (2) a penalty rule from a semantic point of view are set as the linguistic knowledge stored in the sentence meaning determination knowledge storage means 8. FIG. 17 shows an example of a penalty rule from a syntactic viewpoint, and FIG. 18 shows an example of a penalty rule from a semantic viewpoint. The penalty rule is described for each type of viewpoint, and includes an if section indicating a penalty condition and a then section specifying a penalty operation when the condition is satisfied.

The sentence semantics determination means 7 compares each conceptual element constituting a certain meaning understanding candidate with reference to FIGS.
Apply the penalty rule of, and if the if part is satisfied,
The penalty operation of the then part is executed, and the penalty is added as a score of the meaning understanding candidate. When the application of the penalty rule is completed for all semantic understanding candidates, only the semantic understanding candidate with the smallest penalty is selected and output as the semantic understanding result of the input sentence.

For example, the user may say, “Near Yokohama Station,
If I try to say, "I'm looking for around 10,000 yen per night", and enter a sentence such as "I'm near the station in Yokohama, I'm looking for around 10,000 yen per night" Is extracted as a concept element candidate representing the perspective of the place. In the case of this concept element candidate, "Yokohama's"
There is no subscript behind the phrase candidate.

At this time, the sentence meaning determination means 7 stores the "if the adjunct form, adjunct modifier, and cosoad descriptive of the verb exist, and If it does not exist, a penalty is added to the semantic comprehension candidate by one point. ", It is determined that the conceptual element candidate conforms to the above penalty rule, and the semantic understanding including the conceptual element candidate is performed. One point is added to the candidate. Referencing the other penalty rules in the same way, evaluating the linguistic validity of all concept element candidates included in the semantic understanding candidate, and calculating the score of all added penalties as the likelihood of the semantic understanding candidate I do. This process is performed for all the semantic understanding candidates, and the semantic understanding candidate with the minimum penalty is output. Other processes are the same as those of the first embodiment.

[0078]

As described above, according to the present invention, unlike the method using the syntax rules that define the pattern of the entire utterance sentence, it is not necessary to describe all the sentences as sentences, so that the expression efficiency of the sentences is improved. Can be increased. Also,
Since only the phrase candidates that form part of the sentence are targeted for concept element extraction, and the meaning of the entire sentence is expressed by a combination of concept elements, a case frame is applied to all the phrase candidates that make up the sentence. In this way, the semantic verification can be performed efficiently.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing one embodiment of the first to third and seventh aspects of the present invention.

FIG. 2 is a configuration diagram showing an embodiment of the fourth invention.

FIG. 3 is a configuration diagram showing an embodiment of the fifth invention.

FIG. 4 is a configuration diagram showing an embodiment of the sixth invention.

FIG. 5 is a diagram showing an example of giving semantic feature information.

FIG. 6 shows an example of a frame structure expressing a concept element,
(1) is a conceptual element frame that represents the perspective of a place, and (2) is a diagram that represents a conceptual element frame that represents equipment.

FIG. 7 is a diagram illustrating an example of a semantic understanding result expressed by a combination of extracted concept elements.

FIG. 8 is a diagram showing an example of a conceptual element frame expressing “place (point)”.

FIG. 9 is a diagram showing an example of an incorrect meaning comprehension candidate for explaining a modification of the first invention.

FIG. 10 shows an example of setting a required slot in a concept element frame, where (1) is a concept element frame expressing the perspective of the place, (2) is a concept element frame expressing the point of the place, and (3)
FIG. 4 is a diagram showing a conceptual element frame representing equipment.

FIG. 11 is a diagram showing an example of a conceptual element frame expressing “atmosphere (hotel)”.

FIG. 12 is a diagram illustrating an example of an incorrect semantic understanding candidate for describing the second invention.

FIG. 13 is a diagram showing an example of a meaning understanding candidate for describing the third invention.

FIG. 14 is a diagram illustrating an example of an incorrect semantic understanding candidate for explaining the third invention.

FIG. 15 is a diagram showing a hierarchical structure of semantic features expressing the semantic world (hotel reservation) of the target dialogue.

FIG. 16 is a diagram showing an example of a procedure for extracting an intention.

FIG. 17 is a diagram showing an example of a penalty rule based on a syntactic viewpoint.

FIG. 18 is a diagram showing an example of a penalty rule based on a semantic viewpoint.

FIG. 19 is a configuration diagram illustrating an example of a second conventional technique.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input means 2 Morphological analysis means 3 Vocabulary knowledge storage means 4 Concept element extraction means 5 Semantic feature knowledge storage means 6 Concept element knowledge storage means 7 Meaning establishment judgment means 8 Sentence establishment judgment knowledge storage means 9 Mean extraction means 10 Case frame Knowledge storage means 11 Intention extraction knowledge storage means 12 Intention extraction means 13 Semantic feature relation knowledge storage means

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 17/27-17/28

Claims (7)

(57) [Claims] An input means for inputting a sentence expressing a natural language, a morphological analysis means for performing a morphological analysis on the sentence input by the input means and outputting candidates such as words and phrases, Semantic feature knowledge storage means for storing semantic feature information for phrases, and semantic relationships between words and phrases are described. These are expressed in a frame structure as a partial semantic unit of sentences and stored as concept elements. The concept element knowledge storage means and the semantic feature knowledge storage means are referred to, and each candidate output by the morphological analysis means is given a semantic feature, and the concept element knowledge is assigned to a candidate group constituting a sentence part. A concept element extraction means for extracting a concept element by referring to a storage means, and a sentence meaning that describes a condition for a combination of concept elements to be recognized as a semantic expression of the entire sentence as a combination of the semantic expressions of the entire sentence. For the combination of the concept element extracted by the concept element extraction means, the combination of the concept element that meets the condition for the combination of the concept element stored in the sentence meaning determination knowledge storage means is output. Means for determining sentence semantics , and the concept element knowledge storage means comprises:
Multiple slots defining mandatory elements are mandatory and non-mandatory
The data is classified into slots and stored.
Only when a candidate is registered in the
A language processing apparatus, which outputs 2. The method according to claim 1, wherein the morphological analysis means includes :
Is given as position information of a candidate, and the concept element extracting means comprises:
Candidates constituting a certain conceptual element are defined as a continuous section of the input sentence.
2. The language processing apparatus according to claim 1 , wherein a concept element is output only when the following condition is satisfied . 3. The storage device according to claim 1, wherein
Memorizes knowledge of position order relation between candidates that constitute conceptual elements
And the sentence meaning determination means is more than the concept element extraction means.
In the combination of input conceptual elements, the position of the candidate
Select combinations of conceptual elements that contain only conceptual elements that satisfy the order relation.
3. The language processing device according to claim 1 , wherein the language processing device outputs the selected language. 4. The intention of the sentence type of the input sentence,
Intention extraction knowledge that memorizes knowledge of correspondence with predicates in candidates
Storage means and a candidate input from the concept element extraction means.
Extract the predicate from the database and obtain the knowledge of the correspondence between the intention and the predicate.
Determining an intention corresponding to the extracted predicate,
An intention extracting means for outputting an intention to sentence meaning determining means;
In addition, the sentence meaning determination knowledge storage means stores information according to intention.
Storing a combination of the classified concept elements, and
The stage is a combination of conceptual elements input from the conceptual element extracting means.
Select only combinations of conceptual elements that correspond to the intention from the set
4. The language processing device according to claim 1 , wherein the language processing device outputs the data. 5. A meaning for storing knowledge of a relationship between semantic features.
Feature information storage means, wherein the intention extraction means
Predicates exist in candidates input from notation element extraction means
If not, the semantic feature assigned to the candidate and the meaning
The intention is determined using knowledge of the relationship between the features, and the sentence
5. The method according to claim 4, wherein the intention is output to the standing determination means.
2. A language processing device according to claim 1. 6. The concept element extracting means, wherein:
Of the correspondence between concept elements and intentions stored in the fixed knowledge storage means
Referring to the knowledge, the intention determined by the intention extraction means
Extracts only the corresponding conceptual element and sends it to the sentence meaning determination means.
The language processing apparatus according to claim 5 , wherein a combination of concept elements is output . 7. The knowledge meaning determination knowledge storage means includes:
Linguistic knowledge of the candidates constituting the element is stored,
The decision making means is based on the stored linguistic knowledge.
Evaluate the linguistic validity of conceptual elements and evaluate the likelihood of
Calculation and output the combination of conceptual elements with the highest likelihood.
The language processing device according to claim 1, wherein