JP3526095B2 - Rule generator - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rule generation device,
More specifically, the present invention relates to a rule generation device in sentence recognition of voice recognition and dialogue processing.

[0002]

2. Description of the Related Art When stipulating a sequence within a word such as a phrase in speech recognition, it is generally described in the form of a rewriting rule. The rewriting rules are machine-friendly and are a data description method often used in machine processing. The rewriting rule is very difficult to write in a human intuitive description when targeting a linguistic phenomenon. This is one of the reasons that it is impossible to describe a linguistic phenomenon from one viewpoint, and multiple viewpoints such as syntax and meaning are mixed.

For example, Reference 1 “Comparison and Examination of Syntactic Analysis Methods for Continuous Speech Recognition / Understanding Systems” (Nakakawa, 1 person)
According to IPSJ Transactions vol.30, No.8, 1989.10), HMM (Hidd (Hidd
en Markov Model) and a method of recognizing it decisively is proposed. Looking at the grammar used,
The two viewpoints of syntactic properties (nouns) and meanings are mixed.

Reference 2 "Language model in speech recognition" (1 person from outside of the Kita SIG-SL
UD-9201,9201), it has been reported that a lot of non-sentences are generated when trying to describe a complicated language phenomenon by rewriting rules. That is, in the HMM-LR, the search range of the speech recognition is reduced by using the context-free grammar that describes the task. Sentences are randomly generated from the grammar used in HMM-LR, and it is checked what percentage of the generated sentences is correct in Japanese. In voice recognition, the number of recognition candidate words affects the recognition rate and recognition time. It is important to narrow down the candidate words as much as possible. In addition to being syntactically correct, you need to narrow down the words that you recognize in terms of task meaning and other aspects.

Further, according to Document 3, "2-stage LR parser considering meaning in free speech recognition" (3 outside the 3rd session of Acoustical Society of Japan, 3-4-10, 1993.3), grammar,
The LR table is divided into different levels of meaning. In other words, in order to support free-voicing sentences in the task of telephone number guidance, the context-free grammar generates many candidates with slightly different phoneme sequences even if they have exactly the same inquiry content during recognition. Cause an explosion. For this reason, candidates having an effective meaning are pushed to the outside of the beam width, which is one of the causes of deterioration of recognition performance. Therefore, it has been proposed to consider meanings in the recognition process and merge candidates of phoneme sequences having the same meanings to leave valid candidates within the beam width.

[0006]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention As mentioned above, reference 1
When adding or correcting words, things must always be written in consideration of multiple classification points of view. In addition, the statement 2 has a point that the rewriting rule has low descriptiveness in that a correct candidate is generated. Furthermore, the literature
Depending on the task, the items of No. 3 are not limited to being classified by one criterion, and there may be cases in which a plurality of criteria are required, or cases in which only words of a specific condition are desired to be collated first.

In the context-free grammar described above, the rules must be rewritten as the number of classification criteria increases. In addition, the number of branches has increased and the types of nodes have also increased, making the description of data very complicated. In particular, in the case of voice recognition, it may not be always the same as the syntactical unit that regularizes a sentence, and it may be easier to recognize a short word by combining a plurality of words. Also in this case, it was necessary to prepare a rule for that.

The present invention has been made in view of the above circumstances, and describes words and expressions having a plurality of classification criteria without mixing the criteria, and makes them syntactically and semantically correct. The object is to provide a rule generation device that generates existing words or word strings as rules.

[0009]

SUMMARY OF THE INVENTION The present invention, in order to solve the above problem, display the (1) the recognition target words, word groups
Group names or their combinations, a word dictionary consisting of notations, readings, and group names of words to be recognized , and a sequence of words from the left of the template.
Refer to the above word dictionary, with the term extracted in the left side of the top
Then, if the term is a word, the reading is extracted, or if the term is
When it is a group name, the notation is extracted as the right side, and then the right
For each term on the side, refer to the word dictionary with the term as the left side
Then, if the term is a word, the reading is extracted, or if the term is
When it is a group name, the notation is extracted and used as the right side.
Repeat until the right side is read
And a rule generation unit that generates a derivation rule for deriving a reading from the text. Further, (2) in (1), the sentence template is described by adding a restriction to the word or group name to be generated. , The word dictionary is limited to words
Register the limit value to add the
For added words or group names, see the word dictionary above.
When you look at the word with a limit value that matches this limit,
Generate a derivation rule by extracting only the group name , and (3) in (1) above, the sentence template connects a plurality of words or group names.
Described as one term , the rule generation unit
Or a group of group names, select this section at the top
Uses the word or group name included in this section as the left side
Align them in order and arrange them on the right side, then on the right side.
Make each solid word or group name the left side
To generate a derivation rule, and (4) above.
In (1), the sentence template is a word or group.
Group name is described as optional and the rule generator
Replaces the word or group name specified as optional
If it is a containing term, it can be omitted on the right side of the top
With or without the specified word or group name
To generate two types of derivation rules, i.e.
(5) In (1) above, the sentence template is a word
Or add the paired condition name and its value to the group name
The word dictionary puts restrictions on words
The condition name and its condition value are stored, and the rule generation unit
For a word or group name with a subject and its value appended,
To when you see the word dictionary, the condition value under the same conditions name
Extract only matching words or group names
Generating a derivation rules, further, in (6) above (5), the sentence template, multiple describes the limits in pairs condition name and its value to the word or group name, the regulations
The rule generation unit refers to the word dictionary and makes a plurality of restrictions.
Extract only words or group names that meet all
To generate a derivation rule, and (7) above (5)
In, the sentence template is a word or group name
By adding the possible range of conditions names and their values described before
The notational rule generating unit refers to the word dictionary and refers to the same condition name.
Only words or group names whose values are within the possible range of
To generate a derivation rule by extracting as a target , or (8) a group representing a word or a group of words to be recognized
A sentence template describing names or combinations thereof, a word dictionary consisting of notations, readings, parts of speech, and group names of words to be recognized , a connection pattern dictionary describing parts of speech and their connection patterns, and parts of speech and conjugation parts of speech. When a plurality of words or group names are generated by connecting a usage dictionary in which the usage ending is described, the word dictionary and the connection pattern dictionary are combined.
And referring to the above-mentioned inflection dictionary , test whether the words can be connected,
A rule generating section for generating a derivation rule for deriving a reading based on the rule; and (9) in (8), the rule generating section generates a plurality of words or group names by connecting them. , the connection pattern is more than Oh lever, this
It is characterized in that the derivation rule is generated by repeating the number of connection patterns of .

[0010]

[0011]

[0012]

[0013]

[0014]

[0015]

Embodiments will be described below with reference to the drawings. Figure 1 is a block diagram for <br/> explaining an embodiment of a rule generating apparatus according to the present invention, in the figure 1 sentence template,
Reference numeral 2 is a word dictionary, and 3 is a rule generation unit. The sentence template 1 describes a word or a group of words to be recognized. The word dictionary 2 is composed of group names that represent words and groups of words. The rule generation unit 3 generates rules according to the description of the sentence template 1. That is, the rule generation device of this embodiment includes the rule generation unit 3, the word dictionary 2, and the sentence template 1. A word to be recognized is described in the sentence template 1, and based on this, a rule is generated with reference to the word dictionary 2.

FIG. 2 is a diagram showing an example of a sentence template. Describe the word itself or the group name of the word to be recognized. A rule is generated based on this description, but the description of the template is generated on the left side of the top level, and while referring to the dictionary, rules up to the reading level are generated.

FIG. 3 is a diagram showing an example of a word dictionary. The word notation is associated with the group name (meaning here) and reading. The word specified by the sentence template is extracted from the dictionary and a rule is generated as a recognition target word.

In the processing of this embodiment, the word and the group of words described in the sentence template are set to the left side of the top, and if these are words, the reading is extracted, and if they are group names, the word (notation) is extracted from the word dictionary. And make it the right side. Next, for the right side,
The readings are extracted from the word dictionary, respectively, and are taken as the left and right sides. Repeat until the right side is clear. Reading, in this case, is all written in hiragana.

Example 1 Since "$ out" is described in the template, the left side of the rule is set to "$ out", the description of "$ out" is checked from the word dictionary, the word notation is obtained, and those are written on the right side. Generate a rule as. The rules are as shown in Table 1 below.

[0020]

[Table 1]

Next, looking at each item on the right side, since it is not a reading level, the dictionary is searched for "going out" and "going out", and "gaishutsu" and "dekake" are obtained. The rules shown in Table 2 are generated with the left side and the right side respectively.

[0022]

[Table 2]

Looking at the right side, since it is a reading level, the generation is terminated. The finally generated rules are (left side → right side column) as shown in Table 3.

[0024]

[Table 3]

[0025] Next, a description will be given in Example 2. The configuration diagram of the rule generation device of the present embodiment is the same as that of FIG. Figure 4
FIG. 8 is a diagram showing an example of a sentence template. A word itself or a word group name of a word to be recognized and a condition for restricting a set of words pointed to by the word itself or a word group name are additionally described. In the example, a condition of "noun" is described as a restriction for "$ out" in front of ":". A rule is generated based on this description. The template description is generated on the left side of the top level, and the rules up to the reading level are generated while referring to the dictionary.

FIG. 5 is a diagram showing an example of a word dictionary. The notation of words, the group name (meaning here), the reading and the part of speech are described in association with each other. Here, the part of speech is described as a condition in the template. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

The processing of this embodiment is performed in order from the left of the sentence template. The words, groups of words, and conditions described in the sentence template are set to the left side of the top, and words are extracted from the word dictionary according to the description of the template and are set to the right side. Next, with respect to the right side, words are extracted from the word dictionary in the same manner as the left side and the right side. Do this until the right side is visible. In this case, Yomi means that everything is written in Hiragana.

Example 2 Since "$ Go: Noun" is described in the template, the left side of the rule is set to "$ Go: Noun". First, the word having the description "$ Go" is searched from the word dictionary, and " Go out, get out. Since the condition is after ":", it is checked whether or not the description "noun" is present in the dictionary.
This is the right side because it is on the go. I'm not at the reading level yet, so I'll pull "out" from the word dictionary,
Obtain "Gaishutsu" and set it on the left and right sides. Since "Gaishutsu" is a reading level, the processing for this description ends here. The same processing is performed for the description of the remaining templates. The final rules obtained are shown in Table 4 below.

[0029]

[Table 4]

Next, a third embodiment (claim 3) will be described. The configuration diagram of the rule generation device of the present embodiment is the same as that of FIG. FIG. 6 is a diagram showing an example of the sentence template of claim 3. The words themselves to be recognized or the group names of the words are described in the order in which they appear, but a term with a "+" sign is connected to the preceding term to generate a rule in one unit. The example means that a rule is generated in one unit by connecting “$ ga” with “+” to “$ meeting”. A rule is generated based on this description. The template description is generated on the left side of the top level, and the rules up to the reading level are generated while referring to the dictionary.

FIG. 7 is a diagram showing an example of a word dictionary. The word notation, the group name (meaning here), and the reading are described in association with each other. The word specified by the sentence template is extracted from the dictionary and a rule is generated as a recognition target word.

In the processing of this embodiment, the words and the group of words described in the sentence template are set to the left side of the top, and the words are extracted from the word dictionary as the right side according to the description of the template. At this time, if there is a "+" sign in the rear word, the words with "+" are grouped together to generate a rule. On the left side, generate a sign that connects two terms with "+",
Put each term on the right side. Next, for each term on the right side, words are extracted from the word dictionary, and are defined as the left side and the right side. Next, look up the reading on the right side from the word dictionary, and place the reading on the right side. In this case, Yomi means that everything is written in Hiragana.

Embodiment 3 Since "$ meeting" is described in the template and there is a "+" at the back, it is generated by connecting it to the word at the back. The back is "$ ga", so the left side of the rule is "$ meeting + $ ga"
And put “$ meeting” and “$ ga” on the right side. This is shown in Table 5.

[0034]

[Table 5]

The right side is searched with a word dictionary. First, look up the word that has the description "$ meeting", and then use the words "meeting" and "meeting"
To get Do the same for "$ ga", and use the words "ga" and "ha".
To get “Meeting” on each right-hand side that is not a reading level
Draw a word dictionary about "meeting" and then "kaigi"
Get a "housekeeping". Since the right side is all reading levels, the processing for this description ends here. The rules finally obtained are as shown in Table 6 below.

[0036]

[Table 6]

Next, a description will be given in Example 4. The configuration diagram of the rule generation device of the present embodiment is the same as that of FIG. Figure 8
FIG. 8 is a diagram showing an example of a sentence template. The words themselves of the words to be recognized or the group names of the words are described in the order of appearance. "%" Represents that the word may be omitted. A rule is generated based on this description, but the description of the template is generated on the left side of the top level, and while referring to the dictionary, rules up to the reading level are generated.

FIG. 7 is a diagram showing an example of a word dictionary. The word notation, the group name (meaning here), and the reading are described in association with each other. The word specified by the sentence template is extracted from the dictionary and a rule is generated as a recognition target word.

In the processing of this embodiment, the words and the group of words described in the sentence template are set to the left side of the top, and for these, words are extracted from the word dictionary according to the description of the template and set to the right side. At this time, if there is a "%" sign in the word, it may be omitted. For example, if the word has a "+", the word connected to this word and the word without this word Generate two types of rules. If there is no "+",
Generate a symbol without the "%".

Example 4 Since "$ meeting" is described in the template and there is a "+" at the back, it is generated by connecting it to the word at the back.
Since "% $ is" in the back, the left side of the rule is "$ meeting +% $
"," And "$ ga" are generated.
This is shown in Table 7.

[0041]

[Table 7]

Look up the right side in the word dictionary. First, the word with the description of "$ meeting" is checked, and "meeting" and "meeting" are obtained. Do the same for "$ ga" and get "ga" and "ha". For each right side, draw the word dictionary for "meeting" and "meeting" that are not at the reading level, and obtain "kaigi" and "chief". Since the right side is all reading levels, the processing for this description ends here. The rules finally obtained are as shown in Table 8 below.

[0043]

[Table 8]

Next, a description will be given in Example 5. The configuration diagram of the rule generation device of this embodiment is the same as that of FIG. Figure 9
FIG. 8 is a diagram showing an example of a sentence template. The word itself of the word to be recognized or the group name of the word, the condition name that limits the set of words that are pointed to by the word itself or the group name of the word, and the value thereof are described together. In the example, by specifying a condition name "frequency" and its value "1" as a restriction for "$ meeting" in front of ":", the Only those that meet the conditions are generated as rules.

FIG. 10 is a diagram showing an example of a word dictionary.
The word notation, the group name (meaning here), the reading and the condition are described in association with each other. Here, the value is described under the condition of frequency. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

In the processing of the present embodiment, the words, word groups and conditions described in the sentence template are set to the left side of the top, and for these, words are extracted from the word dictionary according to the template description and are set to the right side. Next, with respect to the right side, words are similarly extracted from the word dictionary, and set as the left side and the right side. Do this until the right side is visible. In this case, Yomi means that everything is written in Hiragana.

Example 5 Since the template "$ meeting: frequency = 1" is described, the left side of the rule is set to "$ meeting: frequency = 1", and the word having the description of "$ meeting" is first set from the word dictionary. Find out, "meeting"
Get a "meeting". Since the condition after ":" is, it is checked whether or not there is a description in the dictionary that the value of the item "frequency" is "1". Since "conference" is applicable, this is the right side. I'm not at the reading level yet, so I'll pull "meeting" from the word dictionary and then "kaigi"
To the left and right sides. Since "Kaigi" is a reading level, the processing for this description is ended here. The same processing is performed for the description of the remaining templates. The final rules obtained are shown in Table 9 below.

[0048]

[Table 9]

Next, a description will be given in Example 6. The configuration diagram of the rule generation device of this embodiment is the same as that of FIG. Figure 11
FIG. 8 is a diagram showing an example of a sentence template. The word itself of the word to be recognized or the group name of the word, the condition name that limits the set of words that are pointed to by the word itself or the group name of the word, and the value thereof are described together. In the example, the condition name "part of speech", the value "noun", the condition name "frequency", and the value "1" are specified as restrictions on "$" in front of ":". From the words pointed to by "$ meeting",
Only those that meet the conditions are generated as rules.

FIG. 12 is a diagram showing an example of a word dictionary.
The word notation, the group name (meaning here), the reading and the condition are described in association with each other. Here, the value is described under the condition of frequency. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

In the process of this embodiment, the words, the groups of words and the conditions described in the sentence template are set to the left side of the top, and the words are extracted from the word dictionary according to the description of the template and are set to the right side. Next, with respect to the right side, words are similarly extracted from the word dictionary, and set as the left side and the right side. Do this until the right side is visible. In this case, Yomi means that everything is written in Hiragana.

Example 6 Since "$ meeting: part of speech = noun, frequency = 1" is described in the template, the left side of the rule is set to "$ meeting: part of speech = noun, frequency = 1". Look up the word that describes "$ meeting" and get "meeting""meeting""meeting". Since the condition is after ":", it is checked whether or not there is a description in the dictionary that the value of the "part of speech" item is "noun" for these. “Meeting” and “meeting” correspond. Then, it is checked whether or not the dictionary has a description that the value of the item "frequency" is "1". "Meeting" is applicable. This is the right side. Since it is not at the reading level yet, next, "meeting" is drawn from the word dictionary to obtain "kaigi", which is the left side and the right side. Since "Kaigi" is a reading level, the processing for this description is ended here. The same processing is performed for the description of the remaining templates. The final rules obtained are as shown in Table 10 below.

[0053]

[Table 10]

Next, a seventh embodiment will be described. The configuration diagram of the rule generation device of this embodiment is the same as that of FIG. FIG.
FIG. 8 is a diagram showing an example of a sentence template. The word itself of the word to be recognized or the group name of the word, the condition name that limits the set of words that are pointed to by the word itself or the group name of the word, and the range of its value are described together. In the example, as a restriction on “$ meeting” in front of “:”, by specifying a condition name “frequency” whose value range is larger than “0”, the “$ meeting” is specified. Only the words that meet the conditions are generated as rules from among the words. In addition to “>” (greater than specified) in the example, “<” (less than specified) “! =” (Value other than specified) can be specified as the value range.

FIG. 14 is a diagram showing an example of a word dictionary.
The word notation, the group name (meaning here), the reading and the condition are described in association with each other. Here, the value is described under the condition of frequency. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

In the processing of this embodiment, the words, the groups of words and the conditions described in the sentence template are set to the left side of the top, and for these, the words are extracted from the word dictionary according to the description of the template and set to the right side. Next, with respect to the right side, words are similarly extracted from the word dictionary, and set as the left side and the right side. Do this until the right side is visible. In this case, Yomi means that everything is written in Hiragana.

Example 7 Since "$ Meeting: Frequency>0" is described in the template, the left side of the rule is set to "$ Meeting: Frequency>0", and "$ Meeting" is first described from the word dictionary. Look up words and get "meetings" and "meetings". Since ":" is the condition,
For these, it is checked whether the dictionary has a description that the value of the item "frequency" is "greater than 0". Since "conference" is applicable, this is the right side. Since it is not at the reading level yet, next, "meeting" is drawn from the word dictionary to obtain "kaigi", which is the left side and the right side. Since "Kaigi" is a reading level, the processing for this description is ended here. The same processing is performed for the description of the remaining templates. The rules finally obtained are as shown in Table 11 below.

[0058]

[Table 11]

Next, a description will be given in Example 8. Figure 15
Is a diagram for explaining another embodiment of the rule generating apparatus according to the present invention, in the figure, is connected pattern dictionary 4, 5 is utilized dictionary, other portions of the same functions as in FIG. 1 is the same It is attached with a code. The sentence template 1 describes a word or a group of words to be recognized. The word dictionary 2 is composed of group names that represent words and groups of words. The connection pattern dictionary 4 describes words and their connection patterns. The utilization dictionary 5 describes a part of speech, a utilization part of speech, and an inflection ending. When a plurality of words are connected and generated, the rule generation unit 3 tests whether or not the words can be connected, and generates a rule based on the result.

That is, the rule generation device of the eighth embodiment comprises the sentence template 1, the word dictionary 2, the connection pattern dictionary 4, the utilization dictionary 5, and the rule generation unit 3. Words to be recognized are described in the sentence template 1, and based on this, the word dictionary 2
Generate rules by referring to. When there is a "+" sign, two words are combined to generate a rule, but in this case, whether or not they can be connected is tested by a connection pattern, and only rules that can be connected are Depending on the case, it is generated by being expanded into a practical form by the practical dictionary 5. Especially, in Japanese, the auxiliary verb has a severe connection limitation, so this processing is performed especially for the auxiliary verb.

FIG. 16 is a diagram showing an example of a sentence template. The words themselves of the words to be recognized or the word group names are described in the order in which they appear, but a term with a "+" sign is connected to the preceding term to generate a rule in one unit. The example is to connect “$ will” with “+” to “$ out” and generate a rule in one unit.
At this time, some words may or may not be connected. They are described in a connection pattern, connection verification is performed based on this description, and rules are generated.

FIG. 17 is a diagram showing an example of a word dictionary.
The word notation, the group name (meaning here), the reading and the part of speech are described in association with each other. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

FIG. 19 is a diagram showing an example of a utilization dictionary.
Describe inflectional forms and endings for words that have conjugation, such as verbs and auxiliary verbs. It consists of part-of-speech name, inflection part-of-speech name, and ending. The ending is a word added by taking one letter from the basic form. "*" Means to take only one character and add nothing.

FIG. 18 is a diagram showing an example of a connection pattern dictionary. It consists of part-of-speech name and connection pattern. Write the part of speech or inflection that the part of speech can connect to the front as a connection pattern. The part-of-speech or inflectional form is one or more characters of the part-of-speech name in the word dictionary or the in-use part-of-speech name in the in-use dictionary (identifiable range). In the processing of this embodiment, the words and the group of words described in the sentence template are set to the left side of the top, and words are extracted from the word dictionary according to the description of the template and are set to the right side.
At this time, if there is a "+" sign in the rear word, the words with "+" are generated as a group.

(1) The symbol on the left side of the template description is generated as a symbol connecting two terms with "+". (2) If the part of speech of the backward word is in the connection pattern, that connection pattern is obtained. (3) For the preceding word, refer to the word dictionary and the inflection dictionary by its part of speech, and if there is a connection pattern (if the connection pattern is in a partial character string), the inflectional form notation and reading are generated. (4), ":" is used as a constraint name for the preceding word with the connection pattern as the constraint name.
Generate by separating with. (5) and (1) are left sides, and a front word and a rear word with restrictions are placed on the right side. (6) Generate a rule with the notation and reading corresponding to each of the right sides.

Example 8 (1), "$ out" is described in the template,
And because there is a "+" in the back, it is generated by connecting it to the word in the back. “$ Go + $ will” (2), the back is “$ will”, and when the word behind is searched in the word dictionary, the word “tai” and part-of-speech “auxiliary tai” are obtained. Looking up in the connection pattern dictionary, the part-of-speech “auxiliary verb want” is “for”
Is. (3) The dictionary “forward” is searched for “$ out”, and the word “out” part-of-speech “sahen noun” and the word “go out” part-of-speech “one-stage verb” are obtained. Since there is no description of "use", the in-use dictionary is searched next by the part-of-speech to obtain "continuous use" of "one-stage verb" (if it is in a partial character string). (4), The ending of the one-stage verb "going out" is taken in accordance with the utilization dictionary to obtain "going out". Yomi also obtains a "dekake" in the same way. (5) Putting a restriction on the word "$ out" in the front, "$ out:
Is generated. The left side of (6) and (1), the constrained (5), and the rear word are set. This is shown in Table 12.

[0067]

[Table 12]

(7) The notations on the right side and the readings are linked. The rules finally obtained are as shown in Table 13 below.

[0069]

[Table 13]

[0070] Next, a description will be given in Example 9. The configuration diagram of the rule generation device of this embodiment is the same as that of FIG. Figure 2
0 is a diagram showing an example of a sentence template. The words themselves to be recognized or the group names of the words are described in the order in which they appear, but a term with a "+" sign is connected to the preceding term to generate a rule in one unit. The example means that “$ reservation” is connected to “$ passive” with “+” to generate a rule in one unit. At this time, some words may or may not be connected. They are described in a connection pattern, connection verification is performed based on this description, and rules are generated.

FIG. 21 is a diagram showing an example of a word dictionary.
The word notation, the group name (meaning here), the reading and the part of speech are described in association with each other. The word specified in the sentence template and its condition are referred to and extracted from the dictionary to generate a rule as a recognition target word.

FIG. 23 is a diagram showing an example of a utilization dictionary.
Describe inflectional forms and endings for words that have conjugation, such as verbs and auxiliary verbs. The ending consisting of the part-of-speech name, the part-of-speech name and the ending is a word added by taking one character from the basic form. "*" Means to take only one character and add nothing.

FIG. 22 is a diagram showing an example of the connection pattern dictionary. It consists of part-of-speech name and connection pattern. Write the part of speech or inflection that the part of speech can connect to the front as a connection pattern. The part-of-speech and inflectional form are one or more characters of the part-of-speech name in the word dictionary or the in-use part-of-speech name in the in-use dictionary (identifiable range).

In the processing of this embodiment, the words and the group of words described in the sentence template are set to the left side of the top, and the words are extracted from the word dictionary according to the description of the template and are set to the right side. At this time, if there is a "+" sign in the rear word, the words with "+" are generated as a group.

(1) The symbol on the left side of the template description is generated as a symbol in which two terms are connected by "+". (2) If the part of speech of the rear word is in the connection pattern, repeat the following for the number of connection patterns. (3) With respect to the preceding word, the word dictionary and the inflection dictionary are referred to by the part of speech, and if there is a connection pattern, the inflection of the inflection form and the reading are generated. (4), ":" is used as a constraint name for the preceding word with the connection pattern as the constraint name.
Generate by separating with. (5) In the latter word, a connection pattern is generated by delimiting with "<" as a constraint name for which a connection pattern is requested. (6) and (1) are left sides, and the front word (4) and the rear word (5) with restrictions are placed on the right side. (7) Generate a rule with the notation and reading corresponding to each of the right sides.

Example 9 (1), "$ reservation" is described in the template,
And because there is a "+" in the back, it is generated by connecting it to the word in the back. "$ Reserved + $ Passive" (2), the back is "$ Passive", and when the word behind is searched in the word dictionary, the word "re" part of speech "auxiliary verb le", word "ru" part of speech "auxiliary verb le" obtain. When each is examined in the connection pattern dictionary, the part-of-speech “auxiliary verb le” is “not yet”, and the part-of-speech “auxiliary verb le” is “for 1”.

(Regarding the connection pattern "not yet") (3), a dictionary search for "$ reservation " at the front to obtain the word "use" part-of-speech "wa line 5 verb" and the word "put" part-of-speech "first-stage verb" . Since there is no description of "un", next, the in-use dictionary is searched by the part-of-speech, and "un" of "wa line 5 verb" is obtained. (4) Take the one-letter ending of the wa row five-stage verb "use" and add the ending "wa" to get "use". Yomi also obtains "tsukawa" in the same way. (5) Put a restriction on the word "$ reservation" in front of the word "$ reservation:
"No" is added to the word "$ Passive" in the rear, and the restriction of "$"
Passive <not yet ”is generated. The left side of (6) and (1) and (5) are set. Table 14
Shown in.

[0078]

[Table 14]

(7) The notations on the right side and the readings are linked.

(Regarding connection pattern "1") (8), search the front "$ reservation " in the dictionary, and find the word "use" part-of-speech "wa line five-stage verb" and the word "insert" part-of-speech "one-stage verb". obtain. Since there is no description of "for 1", the in-use dictionary is next searched for part-of-speech to obtain "for 1" of "1 stage verb". (9) Take the ending of the one-stage verb "put" one letter, get "put", and read "yure" in the same way. (10) Put a restriction on the word "$ reservation " in front of the word "$ reservation:
"For 1", put a restriction on the request in the backward word "$ passive" and "$
Passive <1 ”is generated. The left side of (11) and (1) and (10) are set. Table 1
5 shows.

[0081]

[Table 15]

(12) Connect the notations on the right side with the readings. The final rule obtained is Table 16 below.

[0083]

[Table 16]

[0084]

As is apparent from the above description, the present invention has the following effects. (1) By generated automatically from the dictionary rules for the recognition, what was the omission or error at the time of creating the human hand, only these can be avoided, to describe the wish to recognize words (Sentence template here) It is efficient because there is no need to write rules separately. (2) Symbol when necessary another aspect the predicate, without mixing it, it is possible to generate only the correct rules. Since it can be described from different viewpoints, the descriptiveness and extensibility of the data are enhanced. (3) Symbol can combine more than one word in units of the predicate, since the rules for its possible automatically generated, is efficient. (4) The word with the Ministry unillustrated possibility, just describe its presence in the top-level data, because it generates a rule in consideration of the presence or absence, it eliminates the need of description, is efficient, the data serial Increased efficiency and expandability. (5 ) When a description from another point of view is required, correct rules can be generated without mixing them and without manually classifying them. In addition, at that time, a necessary viewpoint and its value can be specified, so that only the rule to be preferentially recognized can be generated, and the descriptiveness and expandability of the data are enhanced. When (6) Symbol viewpoint of predicate there are a plurality, without mixing it, without classifying manually, you can generate the correct rules. Data description and extensibility are enhanced. Also, at that time, a necessary viewpoint and its value can be specified, so that in that case, only the rule desired to be recognized can be generated. The recognition target can be narrowed down at the time of recognition. (7 ) When a description from another point of view is required, a correct rule can be generated without mixing it and without manually classifying it. By being able to describe the possible range of restrictions from another perspective, the descriptiveness and extensibility of data is enhanced. Against what is needed is a description of a different viewpoint (8) co verbs such meaning and part of speech other than the connection by generating it automatically, without classifying manually, you can correct rules, efficiency Target. Against what is needed is a description of a different viewpoint (9) co verbs such meaning and part of speech other than the connection by generating it automatically, without classifying manually, you can correct rules, efficiency Target. In particular, words having the same meaning but different in connection method do not need to be manually described, which is efficient and enhances the descriptiveness of data.

[Brief description of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a rule generation device according to the present invention.

FIG. 2 is a diagram showing an example of a sentence template in the present invention.

FIG. 3 is a diagram showing an example of a word dictionary in the present invention.

FIG. 4 is a diagram showing another example of a sentence template according to the present invention.

FIG. 5 is a diagram showing another example of a word dictionary in the present invention.

FIG. 6 is a diagram showing still another example of a sentence template according to the present invention.

FIG. 7 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 8 is a diagram showing still another example of a sentence template according to the present invention.

FIG. 9 is a diagram showing still another example of a sentence template according to the present invention.

FIG. 10 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 11 is a diagram showing still another example of the sentence template according to the present invention.

FIG. 12 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 13 is a diagram showing still another example of the sentence template in the present invention.

FIG. 14 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 15 is a configuration diagram for explaining another embodiment of the rule generation device according to the present invention.

FIG. 16 is a diagram showing still another example of the sentence template according to the present invention.

FIG. 17 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 18 is a diagram showing an example of a connection pattern dictionary in the present invention.

FIG. 19 is a diagram showing an example of a utilization dictionary in the present invention.

FIG. 20 is a diagram showing still another example of the sentence template in the present invention.

FIG. 21 is a diagram showing still another example of the word dictionary in the present invention.

FIG. 22 is a diagram showing another example of the connection pattern dictionary in the present invention.

FIG. 23 is a diagram showing another example of a utilization dictionary in the present invention.

[Explanation of symbols]

1 ... sentence template, 2 ... word dictionary, 3 ... rule generation unit,
4 ... Connection pattern dictionary, 5 ... Utilization dictionary.

─────────────────────────────────────────────────── ─── Continuation of the front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/18 G10L 15/22 G06F 3/16

Claims (9)

(57) [Claims] 1. A group representing a word or a group of words to be recognized.
Sentence templates that describe group names or their combinations, a word dictionary that includes notations, readings, and group names of words to be recognized , and the sentence templates are extracted in order from the left.
The left side of the top is the selected item, and referring to the word dictionary,
If the term is a word, the reading is extracted, or if the term is a group
When the name is a group name, the notation is extracted as the right side, and then each of the right side
For each term, refer to the word dictionary with the term on the left side,
If the term is a word, the reading is extracted, or if the term is a group
When the name is a group name, the notation is extracted and set as the right side.
Repeat until you read
And a rule generation unit that generates a derivation rule for deriving a reading . 2. The sentence template is a word or glue
The word dictionary is described by adding restrictions to the group name.
Registers the limit value that limits the word, and
The generator uses the word or group name with restrictions.
If you refer to the word dictionary,
Select only words or group names that have
The rule generation device according to claim 1, wherein the rule generation device generates the derivation rule. 3. The sentence template comprises a plurality of words or
Describe the rules by connecting the group names as one item.
The generator uses the term in the case of connecting multiple words or group names.
In this case, make this section the left-hand side of the top, and
Or group the group names according to the order of appearance
And then name each word or group name on the right side
2. The rule generation device according to claim 1, wherein the derivation rules are generated so that the derivation rules are on the left side of each . 4. The sentence template is a word or glue
Describe flop name as being optional, the rule generation unit
Replaces the word or group name specified as optional
If it is a containing term, it can be omitted on the right side of the top
With or without the specified word or group name
The rule generation device according to claim 1 , wherein two types of derivation rules, i.e., one and the other, are generated. 5. The sentence template is a word or glue.
Description, adding the paired condition name and its value to the group name,
The word dictionary, that the conditions subject that imposes limits on the word
The condition value is stored, and the rule generation unit stores the condition name and its value.
In the case of a word or group name added with, the word dictionary
When referring to, the word that matches the condition value with the same condition name
2. The rule generation device according to claim 1, wherein only the group name is extracted as a target to generate the derivation rule. 6. The sentence template is a word or glue
The limitations in pairs condition name and its value is more described flops name,
The rule generation unit refers to the word dictionary and controls a plurality of rules.
Target only words or group names that meet all the limits
6. The extraction rule is extracted to generate a derivation rule.
The described rule generation device. 7. The sentence template is a word or glue
A condition name and a possible range of its value are added to the group name and described, and the rule generation unit refers to the word dictionary and the same
A word or group whose values are in the range possible for the condition name
The rule generation device according to claim 5, wherein only the first name is extracted to generate the derivation rule. 8. A group representing a word or a group of words to be recognized.
Sentence templates that describe group names or their combinations, a word dictionary of notations, readings, parts of speech, and group names of words to be recognized , and connection patterns that describe parts of speech and their connection patterns. In the case where a dictionary, a part-of-speech, a part-of-speech, a part-of-speech, and a part-of-speech, and a plurality of words or group names are connected to each other, the word dictionary and the connection pattern are used.
A rule generation unit for verifying whether or not a word can be connected with reference to a book and the inflection dictionary, and generating a derivation rule for deriving a reading based on the rule. 9. The rule generator is configured to include a plurality of words or groups.
When connecting and creating group names , there are multiple connection patterns.
If so, the rule generation device according to claim 8, wherein the derivation rule is generated by repeating the number of connection patterns .