JP3009636B2 - Spoken language analyzer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic dictation and writing device, a dialogue system in which humans interact with a computer, a speech language analyzer used in an automatic translation system, and the like.

[0002]

2. Description of the Related Art A speech language translation system in which a speech recognition unit outputs a speech recognition result to a language analysis unit as a morpheme sequence in sentence units, and the language analysis unit analyzes with a unified grammar ( Hereinafter, this will be referred to as a first conventional example.) Reference 1 “Toshizawa Toshiyuki et al.,“ ATR Speech Language Translation Experiment System ASURA ”, IPSJ 46th Annual Convention, 6B
-5, March 1993 ". However, in the first conventional example, since the speech recognition unit outputs the speech recognition result to the language analysis unit as a morpheme sequence in sentence units, syntax analysis is performed twice by the speech recognition unit and the language analysis unit. We are inefficient. Further, there is a problem that the language analysis unit is not robust against nongrammatical utterances, in combination with the fact that statistical preferences are not taken into account.

[0003] In machine processing of spoken words, "disfluency" such as non-grammatical expression and slangness
Often causes normal phrase structure analysis to fail. It is desirable to realize a robust syntax and semantic analysis process that can analyze even such "non-fluent" utterances if they can be interpreted semantically. Here, "disfluency" has the phenomenon of rephrasing, inserting interjections such as "ano" and "erto", and omitting particles. For an analysis of these phenomena, see, for example, Reference 2, “Den Yasuharu,
"Ungrammatical phenomena in spoken language and its mechanical processing", SIG-SLUD-9503, SIG-SLUD-9503
1996 ".

[0004] In recent years, attempts have been made to improve the speech recognition rate by integrating speech recognition and phrase structure analysis. In some cases, it is better to analyze the utterance in units shorter than the sentence due to uncertainty or the like. Literature 3 “Toshizawa Toshiyuki et al.
"Language Phenomena of Natural Utterances and Japanese Grammar for Speech Recognition", IPSJ Spoken Language Processing Research Group, 6-5, 199
"Five years" reports that sections separated by short pauses appearing in natural utterances can often be described using context-free grammar. Such a phrase structure of a speech unit smaller than a sentence becomes a tree structure of a unit smaller than a sentence, that is, a partial syntax tree (hereinafter, referred to as a partial tree). If the result of sentence analysis can be obtained by merging these subtrees, it is possible to realize integrated speech language processing from speech recognition to sentence analysis, and it is also advantageous from the viewpoint of economy of computational resources.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above problems and to provide a speech and language analysis apparatus that performs robust and efficient speech recognition and syntax and semantic analysis that do not satisfy syntactic constraints. It is in.

[0006]

According to a first aspect of the present invention, there is provided a speech language analyzing apparatus for performing a morphological analysis and a syntax analysis with reference to a predetermined syntax rule for each partial utterance based on an input utterance voice. By performing, based on the speech recognition means to output the partial syntax tree, the restrictions on the connection of the partial syntax tree, and the dependency probability when each partial syntax tree obtained from a predetermined corpus relates to another word, A partial tree merging processing unit that merges the partial syntax trees output from the speech recognition unit and outputs a syntax and a semantic analysis result.

According to a second aspect of the present invention, in the speech language analyzing apparatus of the first aspect, the partial tree merging processing means is arranged such that each partial syntax tree obtained from a predetermined corpus has a different syntax. Calculating means for calculating the likelihood of dependence on all combinations of words in each partial syntax tree and a plurality of subsequent partial syntax trees based on the dependency probabilities when the words are involved; and A reordering unit that reorders the likelihood of the combination of the partial syntax tree and the word on the storage device in the order of the large likelihood of dependency and stores the reordered likelihood in the storage device; Of which
For the combination of the partial parse tree and the word having the maximum likelihood of dependency, if the target word is a declinable case, the case that can be related to the declinable word is determined in advance and the same plural cases are the same. Merging means for judging whether or not the partial syntax tree can be related based on the constraint on the connection of the partial syntax tree that cannot be related to the verb, and merging the combination of the partial syntax tree and the word which can be related while satisfying the above constraint. It is characterized by having.

Further, in the speech language analysis apparatus according to the third aspect, in the speech language analysis apparatus according to the second aspect, the dependency probability is obtained by modifying a partial syntax tree to be processed obtained based on the corpus. Statistical probabilities that can be associated with the word
After the likelihood depending on the mutual positional relationship between the partial syntax tree to be processed and the word of interest obtained based on the corpus, and the partial syntax tree to be processed obtained based on the corpus, It is characterized by being expressed as a product of the sentence end probability and the sentence end probability.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. When trying to perform ordinary phrase structure analysis on natural spoken words, the entire structure cannot be obtained for non-grammatical expressions, and often a fragmentary group of subtrees is obtained. In the embodiment according to the present invention,
A speech language analysis apparatus that combines these subtree structures and analyzes the structure of the entire sentence will be described. Here, a statistical method was used to realize a robust merging process. The statistical data for this was obtained from the travel conversation corpus. In the present embodiment, the merging method will be described in detail later, and an experimental result of reconstructing a tree structure database decomposed in units of clauses based on the present method will be described. In the present embodiment, the details of the merging method will be described later.

An embodiment according to the present invention is a speech language analyzing apparatus using a method for statistically analyzing a language phenomenon of a spoken language that is difficult to process by ordinary phrase structure analysis. For example, when trying to process natural spoken language by a machine, ordinary phrase structure analysis may not give the result of analyzing the entire sentence due to non-grammatical expressions, slang, etc. Even in such a case, as a result of the phrase structure analysis, it is possible to obtain a tree structure including morphemes, phrases, phrases, clauses, and the like in the utterance input. When a plurality of such tree structures are considered to constitute one sentence (or are governed by one predicate) in terms of meaning, if they can be combined and processed, it is useful for semantic analysis or translation processing. In this paper, a subtree refers to a tree structure of such a unit smaller than a sentence, and merging subtrees refers to a process of merging subtrees to form a structure equivalent to a sentence. The phrase structure of an utterance unit smaller than a sentence becomes a tree structure of a unit smaller than the sentence, that is, a subtree. If the subtree can be merged to obtain a sentence analysis result, from speech recognition to sentence analysis. An integrated form of speech language processing can be realized, and it is also advantageous in terms of economy of computational resources.

In the method proposed here, the partial tree output by the phrase structure analysis is converted into a dependent (dependency) structure, and the result is merged. Is merged. A dependency probability between each subtree and an expression in another subtree on which the subtree can depend is extracted from a corpus subjected to dependency structure analysis, and is used as preference information for merging. In addition, mutual positions of these subtrees, positions where sentence endings are assumed, and the like are also used to calculate the likelihood of dependence.

FIG. 1 is a block diagram showing a configuration of a speech language analyzer according to an embodiment of the present invention. In FIG. 1, an uttered voice is input to a microphone 10, converted into a voice signal, and then input to a voice recognition unit 1. The speech recognition unit 1 outputs, for example, L
After performing acoustic analysis such as PC analysis to extract acoustic feature parameters such as cepstrum coefficients and logarithmic power, based on the acoustic feature parameters, for example, an uttered speech using a phoneme HMM (phoneme hidden Markov model). To recognize the phoneme sequence corresponding to. Furthermore, the speech recognition unit 1 performs a morphological analysis and a syntax rule in a syntax rule memory 20 that stores, for example, vocabulary and grammatical rules in advance for each partial utterance based on the phoneme string recognized by the speech. , A partial parse tree is output for each partial utterance.

The output of the subtree in the speech recognition unit 1 is performed as follows. Reduction in LR parsing of phoneme sequences (re
As a result of performing the duce) process, even if the candidate does not reach the acceptance, the candidate that satisfies the following conditions is left connected to the partial solution list. (Condition 1) All symbol sequences are subjected to reduction processing. (Condition 2) The grammar history is not a subset of the grammar history of another partial decomposition in which the same symbol series is subjected to reduction processing. Condition 1 means that a subtree is not left unless it is at least one word (morpheme). Condition 2 is

This is for avoiding that many similar candidates remain due to a syntax rule such as [Expression 1] (<adverb phrase>-><adverb>). A plurality of solutions (subtrees) registered in the partial solution list are pruned (prunin) by beam search.
g) Unless otherwise performed, it is output as a recognition result candidate.

Next, the partial tree merging processing unit 2 performs the above-described speech recognition based on the restriction on the connection of the partial syntax tree and the dependency probability when each partial tree obtained from a predetermined corpus relates to another word. The partial syntax trees output from the unit 1 are merged to output a syntax and semantic analysis result. Here, in the subtree merging processing unit 2, all subtrees and a plurality of subsequent subtrees are obtained based on the dependency probability obtained when the subtrees relate to different words, obtained from a predetermined corpus. Is calculated, and the likelihoods for all the calculated subtrees and word combinations are rearranged in the order of the likelihood of dependence on the reordering memory 42 and stored. For the combination of the subtree and the word having the maximum likelihood of the combination of the subtree and the word, the case that can be related to the declinable word is determined in advance when the target word is a declinable word; Based on the restriction on the connection of subtrees that the same multiple cases cannot be related to the same declinable word, it is determined whether or not they can be related, and the combination of subtrees and words that can be related to satisfy the above restrictions is merged. Do In addition, the dependency probability is preferably a statistical probability obtained based on the corpus, in which a subtree to be processed can be related to a word of a destination, and a part to be processed obtained based on the corpus. It is expressed as the product of the likelihood depending on the mutual positional relationship between the tree and the word at the destination, and the sentence end probability obtained based on the corpus and which comes after the subtree to be processed.

FIG. 2 shows an example of the operation of the speech language analyzer of FIG. As shown in FIG. 2, the uttered speech input by the speech recognition unit 1 is a partial syntax tree group ((1) to (3)).
After that, the partial syntax tree group is converted by the partial tree merge processing unit 2 into a syntax and semantic analysis result. In this figure, the part connected by the arrow is one tree (partial tree)
In some cases, one partial syntax tree is composed of one word, such as partial syntax trees (1) and (2), and one partial syntax tree is composed of two clauses, such as partial syntax tree (3) Some do.

In the speech language analyzer of the present embodiment,
The speech recognition unit 1 and the subtree merging processing unit 2 are configured by, for example, a digital computer.
For example, it is constituted by a hard disk memory. In this embodiment, statistical information indicating the preterminal symbol bigram, the bigram of the phrase structure rule, the dependency likelihood between partial syntax trees, and the like may be used as the preference information. Here, regarding “bigram of phrase structure rule”, reference 4
“K. Kita al.,“ Continuously Spoken Sentence Recogn
ICS by HMM-LR ", ICSLP-92, pp. 305-308, 1992".

Next, an embodiment of the processing of the voice recognition unit 1 will be described. In this embodiment, the voice recognition unit 1
Uses LR parsing as a parsing method for outputting a subtree, in other words, a partial decomposition. Table 1 shows a description example of a grammar for this method. The grammar shown in Table 1 describes the grammar in units of “sections”.

[0018]

[Table 1] Description example of grammar ─────────────────────────────────── (<start> <−− > (<Section>)) (<section> <−−> (<predicate>))) (<section> <−−> (<postposition phrase>))) (<section> <−−> (<adverb phrase> )) (<Clause> <−−> (<noun modifier>)) (<clause> <−−> (<noun>))) (<clause> <−−> (<numeric>)) (<clause> <−−> (<compound>)) (<section> <−−> (<inflection>))) (<section> <−−> (<interjection>))) (<predicate> <−−> (< (Verb phrase))) (<predicate> <−−> (<verb>))) (<verb> <−−> (<sa-variable noun> <auxiliary verb>))) (<verb phrase> <−−> (<verb > <Auxiliary verb>)) (<postposition phrase> <−−> (<noun> <particle>))) (<adverb phrase> <− −> (<Adverb>)) (<noun> <−−> (<compound>))) (<adverb> <−−> (ima));; now (<noun> <−−> (ny uu y oo ku sh itihoteru)) ;; New York City Hotel (<compound> <−−> (ro Q py akunig oo shi ts u)) ;; Room 602 (<particle> <−−> (ni)) , (<Sa-noun> <−−> (taizai)) ;; stay (<auxiliary verb> <−−> (sh i)) ;; shi (<auxiliary verb <−−> (teimasu)) ;; I (<interjection> <−−> (an oo)) ;; Ah (<noun> <−−> (wata sh i)) ;; ; There (<noun> <−−> (no)) ;; no (<noun> <−−> (seki)) ;; cough (<sa variable noun> <−−> (yoyaku)) ;; <Auxiliary verb> <−−> (tai N desuga)) ;; Mostly (<Particle> <−−> (wa)) ;; is (<noun> <−−> (suzuki)) ;; Suzuki ( <Particle> <−−> (to)) ;; and (<verb> <−−> (m oo sh i)) ;; Speaking (<auxiliary verb> <−−> (masu)) ;; Hamburger (<noun> <−−> (ha N b aa g aa)) ;; hamburger (<particle> <−−> (o)) ;; , (<Sa-noun> <−−> (ch uu mo N)) ；； Order (<auxiliary verb ＞ ＜ −− ＞ (ta N desu)) ；； ───────────────────────

In the HMM-LR continuous speech recognition method in which the HMM phoneme recognition and the LR syntax analysis method are combined, as a result of performing a reduction process, only candidates leading to acceptance are stored as sentences and a beam search is performed. In this embodiment, in order to output a partial tree, in other words, a partial solution, even a candidate that has not been accepted is left in the process by connecting it to the partial solution cell list. Of course, the accepted candidates are also left as before, and the two are combined to perform pruning by beam search. Then, the one with the higher score is output as the recognition result.

Here, if all the partial solutions are linked to the partial solution cell list, for example,

Many similar candidates remain due to the syntax rule of [Expression 2] (<adverb phrase><−−>(<adverb>)).
Table 2 shows an example of a similar subtree.

[0021]

[Table 2] Examples of similar subtrees 節 clause │ adverb phrase │ adverb │ now ────── adverb phrase │ adverb │ now ────── adverb │ now ───── ─

The adverb "now" goes up to the adverb phrase, and the adverb not. In such a case, the largest partial tree is stored, and the other partial trees can be expressed as the partial tree. Therefore, only when the grammar history of a certain cell is not a partial subtree of the grammar history of another cell in which the same symbol series is subjected to the reduction processing, the grammar history is connected to the partial decomposition cell list to improve the processing efficiency. .

Next, an embodiment of the processing of the subtree merging processing unit 2 will be described. In this embodiment, a sequence of a dependency structure tree (subtree) is used as an input. Therefore, in the present embodiment, the speech recognition unit 1 converts the phrase structure (sub) tree generated by the syntax analysis into the dependency structure (sub) tree. In order to convert a phrase structure (sub) tree into a dependency structure (sub) tree, rules for determining a grammatical head in the phrase structure are used. In a dependency structure, merging subtrees is equivalent to finding a dependency between subtrees.

The dependency structure of the example sentence "I am currently staying at Room 602 of New York City Hotel" is shown in FIG. Dependencies are indicated by arrows. Note that <TLO
C>, <ADJUNCT>, and <LOCT> are labels indicating the type of dependency.

FIG. 2 (upper row) shows a dependency structure of three partial trees corresponding to the example sentence "I am currently staying at Room 602 of New York City Hotel". Due to the presence of the pause delimiter “▼”, the subsequent part is a partial tree different from the previous part. Also, since there is no grammatical rule connecting the adverb (“now”) and the noun (“New York City Hotel”), the “now New York City Hotel” part becomes two subtrees. Fig. 2
Dependencies not found in (upper) are created, and Fig. 2
(Lower).

The following algorithm can be used for subtree merging using the dependency structure tree as an input. <Step SS1> Each subtree and subsequent subtrees (plural)
Calculate the dependency likelihood for all the words in the combination. <Step SS2> Among the combinations for which the dependency likelihoods have been calculated in step SS1, those having a dependency likelihood of a certain value or more are arranged in the order of the dependency likelihood, and they are merged in order. At the time of verbal modification and merging, case frames are checked.

The operation of the above algorithm will be described with reference to the following example. Let's look at an example of how this process works. Example: "I ordered a hamburger and a hamburger." Consider merging subtrees made from this sentence split at the punctuation mark. First, two clauses ending with the pronouns "I" and "wo" are statistically given a high likelihood of being likely to depend on the word "order". In this example, the restatement appears, but "Hamburgers"
And "ordered" are close in distance, so "hamburger"
, And a stronger likelihood of dependence than "ordered" are given. First, "hamburger" and "ordered" are merged. As a result of this merger, the case of “ordered” was consumed, so the interpretation that “hamburger” was “ordered” was rejected by the case frame check and “hamburger”
Is determined to be garbage without a destination.

Next, the calculation of the dependency likelihood will be described. In the present embodiment, the dependency likelihood is a product of three likelihood functions (each taking a value of 0 or more and 1 or less). Hereinafter, each function will be described.

(A) Statistical Dependency by Syntactic Structure
Function for consistency which is bigram likelihood function of a kind, clause at the top of the originating candidate (or smaller partial structure than clause)
And the likelihood that they are dependent on each other from the heading and part-of-speech pattern of each node of the dependency destination candidate (usually a word). Here, the dependency destination is a synonym of the dependency destination, and the dependency source is a synonym of the dependency source. The likelihood is a conditional probability PR that a dependency is established when a pattern of a pair of dependent candidates occurs in the same sentence in a corpus in a form in which a candidate of a dependent source precedes. This probability PR is expressed by the following equation, where PP is a source and destination pattern of a dependent candidate pair.

[0030]

[Mathematical formula-see original document] PR = (number of times PP is actually dependent) /
(The number of times the dependent pattern appears earlier in the same sentence)

The probability value PR is obtained by collecting statistics separately from the merging process and using a value stored in advance in a table (details will be described later). Since the heading information includes semantic information, the relationship between the meaning and the dependency can be grasped within the range of the combination appearing in the corpus by this method. Pattern matching at the time of likelihood table search is based on a kind of longest match method. An explanation will be made below using the example of FIG.

Example: The likelihood that “reservation” depends on “cancel” is obtained from a predetermined likelihood database. In the example of FIG. 3, the upper row finds the pattern of the dependency candidate pair, and the lower row shows the part of speech. At each stage, if the pattern is in the likelihood database, use the likelihood in the likelihood database,
If not found, proceed to the next step.

The above-mentioned probability value PR is calculated from a predetermined corpus.
It is calculated as a statistical probability that the subtree to be processed can be related to the dependent word.

(B) of the likelihood depending on the position of the candidate
The dependency likelihood is calculated from the mutual positional relationship between the function dependent source candidate subtree and the dependent candidate word or the positional relationship between the dependent candidate candidate and the sentence end. For this purpose, as shown below, a histogram of the physical distance of the dependency relationship is obtained from a corpus, which is a text database containing the text data for learning, and is used.

(1) If it can be determined from the histogram of the dependency destination top node and the part of speech of the dependency source that the dependency relationship is always an adjacent relationship (for example, a direct dependency relationship between an auxiliary verb and a verb), the adjacent dependency candidate is determined. Gives likelihood 1 and gives likelihood 0 to non-adjacent dependency candidates. (2) When it is determined from the histogram that the dependency has a non-adjacent dependency, an exponential function ex is used as the likelihood.
p (−kλd) is used. Here, λ is the reciprocal of the average dependency distance for the type of dependency obtained from the histogram, d is the distance between dependency candidates, and k is a predetermined constant. Subtrees closer to this likelihood setting are preferentially merged. (3) Since words related to the entire sentence, such as inflections and conjunctions, often depend on the last expression of the sentence in dependency structure analysis, in the present embodiment, it is assumed that conjunctions, inflections, and the like are associated with the last sentence expression. . Therefore, the probabilities (see the following section (C)) that the dependence destination candidate is the end-of-sentence expression are given to those words as the dependence likelihood.

The likelihood depending on the position of the dependency destination candidate is
From a predetermined corpus, the likelihood is calculated as a likelihood depending on the mutual positional relationship between the subtree to be processed and the dependent word.

(C) Function of Sentence End Probability In speech language processing, one person may speak a plurality of sentences in succession, and the end of a sentence is not always clear. Since the dependency does not hold beyond the end of the sentence, if the probability that the end of the sentence is between the dependency candidates increases, the likelihood of dependency between those candidates decreases accordingly. Therefore, the dependency likelihood according to the sentence end probability is defined as the probability that the end of the sentence does not come between two subtrees (dependent candidate pairs). Assuming that there are some breaks in the subtree between the source candidate and the destination candidate, the likelihood is the product of the probabilities that the end of the sentence does not come at each of those breaks (1−the probability that the end of the sentence comes). The sentence end probability that a sentence end comes between each subtree is calculated as a function of the part of speech of the last word of the preceding subtree and the part of speech of the first word of the next subtree, and using statistical data extracted from the corpus.

The sentence end probability is calculated as a sentence end probability that a sentence end comes after a subtree to be processed from a predetermined corpus.

Next, the case frame check will be described. The verb (adjective) has a limitation that a case that can be involved is determined, and that the same case cannot be used more than once. In other words, the constraint is that if the word to which the word is related is a declinable word, the case that can be related to the declinable word is determined in advance, and the same plurality of cases cannot be related to the same declinable word. It is desirable to consider these restrictions when merging subtrees. In this experiment, we investigated the surface case pattern of the case particle "ga, wo, ni" for each verb appearing in the corpus, and examined whether the expressions governed by those case particles could be related to each verb in the merging process. I checked it. In addition, it is prohibited that expressions having the same surface case are related to the same verb more than once (at this time, the polysemy of the case of the particle is also taken into consideration). In addition, the experiment was conducted in consideration of passive, causative, and changes in case patterns of verbs governed by the auxiliary verb "tai".

Next, data extracted from the corpus will be described. The data extracted from the corpus among the data used in the merging process are summarized as follows. The corpus used here is a speech language database owned by the present applicant (for example, see Reference 5 “T. Morimoto e
t al., “A Speech and Language Database for SpeechT
ranslation Research ", Proceeding of ICSLP '94 pp.17
91-1794, 1994. " ) (Travel conversation: 375 conversations, 1
This dependent structure analysis, which is a result of performing the dependent structure analysis in (3647 sentences), is a correct answer that has been inspected and corrected manually.

Then, when all the dependencies in the dependency structure database are examined, and the dependency source pattern and the dependency destination pattern appear in the same sentence, the dependency source pattern appears earlier. The probability PR that the dependency relationship is established is obtained using the above equation (2). Dependency pattern is basically a node of subtree (=
(Word) heading and part-of-speech pair. The dependency source pattern is a clause or a structure within a clause. For example, in response to the expression "I am going", two patterns ("go" / verb) and ("ha" / particle) are obtained as dependent patterns, and ("ha" / Particles ("I" / pronoun) and ("I" / pronoun) are obtained. In addition, in order to avoid the sparse data problem, statistics were also obtained for those in which the headword of the independent word was replaced with a wildcard “*”.

Further, a histogram of the mutual distance when the two parts of speech are dependent on each other was obtained. In the case of many functional words (for example, particles), they depend only on the preceding word, but verbs and the like have long-distance dependencies in the form of dependency. As described above, the dependency distance information obtained from the corpus is used to calculate the dependency likelihood.

Here, the sentence ending model obtained the probability that the end of a sentence comes between two parts of speech from the corpus. In addition, we examined whether the words appearing in the corpus could take on the surface cases "ga", "wo", and "ni", and used them as restrictions when merging.
For example, the verb "meet" takes the case particles "ga", "wo", and "ni".

FIG. 4 is a block diagram showing details of the configuration of the subtree merging unit 2 of FIG. In FIG. 4, the subtree merge processing unit 2 includes a merge processing controller 30 which is an arithmetic and control unit which is a digital computer such as an MPU that executes the merge processing.
0, (a) a processing memory 31 used for executing the merging process, (b) a distance attenuation rate table 32 used for calculating likelihood based on the position of the dependence destination candidate, and (c) a case frame check. Case frame table 33 used;
(D) End-of-sentence / start-of-sentence and part-of-speech statistical file 34 used for calculating end-of-sentence probability and (e) Dependency probability table 35 used for calculating statistical consistency of dependency by syntactic composition Is done. In addition, the merge processing controller 30 includes (f) an input file 21 for storing the result of the partial syntax tree output from the speech recognition unit 1,
(G) An output file 22 that stores the syntax and semantic analysis results output from the merge processing controller 30 is connected. Further, the processing memory 31 includes a partial tree memory 41
, A rearrangement memory 42, a dependency likelihood memory 43, and a mandatory case list memory 44. Here, the processing memory 31 is composed of, for example, a hard disk memory,
These files, tables and files indicated by 22, 32 to 35 are stored in a storage device such as a hard disk memory.

Further, the record format and record example of each table and file in the subtree merge processing unit 2 of FIG. 4 are shown. Table 3 shows the distance attenuation rate table 32 used for calculating the likelihood based on the position of the dependence destination candidate, and Table 4 shows the case frame table 33 used for the case frame check.
Table 5 shows the sentence end / sentence head and part-of-speech statistical file 34 used for calculating the sentence end probability, and Table 6 shows a dependency probability table 35 used for calculating the statistical consistency of the dependency relationship based on the syntactic structure. Shown in These tables and files are obtained by performing a predetermined grammatical analysis based on a corpus of a database of travel conversations owned by the present applicant, and are in the form of a predetermined dependency analysis tree. This is the grammatical analysis data described.

[0046]

[Table 3] Distance attenuation table ─────────────────────────── Record format :: = (Dependent part of speech Dependent part of speech Average dependent distance) ─────────────────────────── Record example: (<person name> <adjunct particle> 0) (<adverb> <savari noun> 396 / 287) (<conjunction> <auxiliary verb stem> 412/51) (<conjunction particle> <main verb> 657/260) (<conjunction particle> <main verb> 1347/664) (<case particle> <main verb> (6028/7099) ───────────────────────────

[0047]

[Table 4] Case frame table ─────────────────────────────────── Record format :: = (word case Frames) words :: = (headline part of speech) case frames :: = ('case frames (case frames *)) case frames :: = (' case frames (case slots *)) case slots :: = (case Particle {+,-}) ─────────────────────────────────── Record example: ((broadcast <sa-change Noun>) (case frame group ((case frame ((ga-))))))) ((overnight <sa inflection noun>) (case frame group ((case frame ((ga-) (ni-))))) ) ((Pay <Main verb>) (Case frame group ((Case frame ((GA-) (A--)))) (Case frame ((GA-) (NI-))))))))) Verb>) (case frame group ((case frame ((ga-) (ni-)))))))) ((decoration <main verb>) (case frame group ((case frame ((ga-) (wo-) ))))) ──────────────── ───────────────────

[0048]

[Table 5] End-of-sentence / beginning-of-sentence, part-of-speech statistical file ────────────────────────────── Record format :: = (sentence-end-of-speech part-of-sentence Part-of-speech (end sentence frequency overall frequency)) Record example: (final particle main verb (71 100)) (common noun interjection (3 48)) (personal common noun (3 23)) (suffix connective (17 53)) (normal Noun Person (7 27)) (Conjunctive Interjection (1 209)) (Inflection Address (1 1)) ───────────────────────── ─────

[0049]

[Table 6] Dependency probability table ─────────────────────────────────── Record format :: = [[' Dependency Occurrence N] ['Pattern Occurrence N] [' Depend on [['Speech Part of Speech] [' Heading Heading]]] ' ] ['Argument [[' Head [['Part of speech Part of speech] [' Heading Heading]]]]]]] ──────────────────────── ─────────── Record example: ─────────────────────────────────── [ [Dependency Appearance 3] [Pattern Appearance 3] [Dependent [[part of speech <auxiliary verb stem>] [headline arima]]] [Dependence source [[head [[part of speech <auxiliary verb>]] [heading in]] ] [Argument [[head [[part of speech <proper noun>]] [headline New York City Hotel]]]]]]]] ────────────────── ──────────────── [[Dependent appearances 14] [Pattern appearances 25] [Depends on [[part of speech <final particle>] [heading is]]] [Dependent [[Head [[part of speech <auxiliary verb stem>] [heading in]]]] [argument [[head [[part of speech <quasi-body particle>]] [heading of]]]]]]]] ─────────────────────────── [[Dependency Appearance 10] [Pattern Appearance 10] [Dependency [[Part of speech <Main verb >] [Heading]]] [Dependent Source [[Head [[Part of speech <case particle>]] [Heading wo]]] [Argument [[Head [[Part of speech <Sa-noun>]] [Heading reserved]]]]] ]] ───────────────────────────────────

FIG. 5 is a flowchart showing the merging processing for one utterance unit executed by the subtree merging processing section of FIG. As shown in FIG. 5, first, in step S11, a partial tree for one utterance is read from the input file 21 storing the output from the speech recognition unit 1 to the partial tree memory 41 in the processing memory 31. Next, in step S12, the sentence end probability that the end of the sentence comes after each subtree is calculated with reference to the sentence end / head of sentence and the part of speech statistics file 34 for the processing target subtree. Then, in step S13,
For each declinable word, a list of mandatory cases that are always required by the declinable word is found and stored in the mandatory case list memory 44. Further, in step S14, as described in detail above,
As in the above step SS1, the dependency likelihood is calculated for the combination of each subtree and all the words in the plurality of succeeding subtrees and stored in the dependency likelihood memory 43. In step S15, the dependency likelihood is calculated. Of the calculated combinations are arranged in ascending order of the likelihood of a certain value or more and stored in the reordering memory 42.

Next, in step S16, it is determined whether or not there is a combination of words that have not been merged. If there is, the process proceeds to step S17. The output is output to the output file 22, and the merging process ends.

In step S17, it is determined whether or not the combination having the maximum likelihood depends on the case particle depending on the word. If it depends, the process proceeds to step S18. If not, the process proceeds to step S20. In step S18, the case frame of the combination having the maximum likelihood is checked with reference to the essential case list in the essential case list memory 44, and it is determined whether or not the above-described case frame constraint is satisfied. If the constraint is not satisfied, the process proceeds to step S21. If the constraint is satisfied, the case of the case particle is deleted from the essential case list memory 44 in step S19.
Go to 0. In step S20, a merging process for merging the combinations is performed, and the process returns to step S16. In step S21, after the combination is rearranged and removed from the memory 42, the process returns to step S16.

As described above, according to the present embodiment of the present invention, morphological analysis and syntactic analysis with reference to a predetermined syntactic rule are performed for each partial utterance based on the input utterance voice. Outputting the partial syntax tree, and then based on the constraints on the connection of the partial syntax tree and the dependence probability when each partial tree obtained from a predetermined corpus relates to another word, And outputs the syntax and semantic analysis results. As a result, it is possible to robustly perform the recognition of the utterance that does not satisfy the syntactic constraint and the syntax and semantic analysis. In addition, in order to transfer a partial tree from the speech recognition unit to the partial tree merging unit,
Processing efficiency is good, and processing can be performed at a higher speed than in the conventional example.

[0054]

As described in detail above, in the speech language analyzing apparatus according to the first aspect of the present invention, morphological analysis and predetermined syntax rules are referred to for each partial utterance based on the input utterance voice. By performing the syntax analysis, a speech recognition unit that outputs a partial syntax tree, constraints on connection of the partial syntax tree, and a dependency probability when each partial syntax tree obtained from a predetermined corpus relates to another word are obtained. A partial tree merging unit that merges the partial syntax trees output from the speech recognition unit and outputs a syntax and a semantic analysis result. Therefore, according to the present invention, the recognition of utterances that do not satisfy the syntactic constraints and the syntactic and semantic analysis are performed robustly,
And, from the speech recognition means to the subtree merging processing means,
Since the delivery is performed using the subtree, a speech language analysis device with high processing efficiency can be provided.

Further, in the speech language analysis device according to the second aspect, in the speech language analysis device according to the first aspect,
The subtree merging processing means, based on a dependency probability when each partial syntax tree relates to another word, obtained from a predetermined corpus, all the partial syntax trees and all the subsequent partial syntax trees in a plurality of partial syntax trees Calculating means for calculating a likelihood of dependence on a combination of words; and reordering the likelihoods of dependence on all combinations of partial syntax trees and words calculated by the calculating means in the order of the likelihood of dependence on the storage device. The reordering means stored in the storage device and the combination of the partial parse tree and the word having the maximum likelihood of the combination of the partial parse tree and the word rearranged by the reordering means are used by the word of the destination. If the case is a case, the case that can be related to the declinable is determined in advance, and based on the constraint on the connection of the partial parse tree that the same plural cases cannot be related to the same declinable, Ri obtained whether a judgment, and a merging unit for merging the partial parse tree and word combinations obtained relates satisfy the above constraints. Therefore, according to the present invention, recognition of an utterance that does not satisfy the syntactical constraints, robust syntax and semantic analysis, and transfer of the partial tree from the speech recognition unit to the partial tree merging unit are performed. Therefore, it is possible to provide a speech language analysis device with high processing efficiency. Further, the processing of the partial tree merging processing means can be easily and automatically performed efficiently.

Further, in the speech language analysis apparatus according to the third aspect, in the speech language analysis apparatus according to the first aspect, the dependency probability is based on a partial syntax tree to be processed obtained based on the corpus. The statistical probability that can be related to the preceding word, the likelihood that is obtained based on the corpus, and the likelihood that depends on the mutual positional relationship between the partial syntax tree to be processed and the related word, and the likelihood obtained based on the corpus. And the sentence end probability that the sentence end comes after the partial syntax tree to be processed. Therefore, according to the present invention, recognition of an utterance that does not satisfy the syntactical constraints, robust syntax and semantic analysis, and transfer of the partial tree from the speech recognition unit to the partial tree merging unit are performed. Therefore, it is possible to provide a speech language analysis device with high processing efficiency. Further, the processing of the partial tree merging processing means can be easily and automatically performed efficiently.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech language analysis device according to an embodiment of the invention.

FIG. 2 is a diagram illustrating an operation example of the speech language analysis device in FIG. 1;

FIG. 3 is a diagram showing an example of pattern matching processing at the time of searching a likelihood table in a subtree merging processing unit in FIG. 1;

FIG. 4 is a block diagram illustrating details of a configuration of a subtree merging processing unit in FIG. 1;

FIG. 5 is a flowchart showing a merging process for one utterance performed by the subtree merging unit of FIG. 4;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Speech recognition part, 2 ... Subtree merge processing part, 10 ... Microphone, 20 ... Syntax memory, 21 ... Input file, 22 ... Output file, 30 ... Merge processing controller, 31 ... Processing memory, 32 ... Distance attenuation rate Table, 33: case frame table, 34: sentence end / sentence, part of speech statistics file, 35: dependency probability table, 41: partial tree memory, 42: rearrangement memory, 43: dependency likelihood memory, 44: required case list memory.

Continuation of the front page (72) Inventor Tadashi Morimoto 5th Sanriya, Seiya-cho, Soraku-cho, Kyoto, Japan , A) Arakawa, T., "Aggregating Subtrees by Statistical Method," IEICE Technical Report, May 17, 1996, NLC96-8, p. 47-54 Takezawa, et al., "Continuous Speech Recognition Using Syntax Rules and Bigrams of Preterminal Symbols in Subtrees", IEICE Technical Report, 1995, NLC95-54 (SP95- 89), p. 55-62 Tashiro, et al., "A Method of Merging Subtrees for Spoken Language Processing", Information Processing Society of Japan, 1995, Vol. 95, No. 89 (NL-109), p. 27-32 Takezawa, G., "Dialogue Speech Recognition Method Using Syntactic Rules Based on Subtrees and Preterminal Bigram", IEICE Transactions D- ▲ II ▼, December 1996, Vol. J79-D. -▲ II ▼, No. 12, p. 2078-2085 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 17/27 G06F 17/28 G10L 3/00 561 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. A speech recognition unit for outputting a partial syntax tree by performing a morphological analysis and a syntax analysis with reference to a predetermined syntax rule for each partial utterance based on an input uttered speech, Based on the connection-related constraints and the dependency probabilities when each partial syntax tree obtained from a predetermined corpus relates to another word, the partial syntax trees output from the speech recognition unit are merged, and the syntax and semantics are combined. A speech language analyzing apparatus, comprising: a subtree merging processing means for outputting an analysis result. 2. The partial tree merging processing means according to claim 1, wherein each partial syntax tree and a plurality of subsequent partial syntaxes are obtained based on a dependency probability obtained from a predetermined corpus when each partial syntax tree relates to another word. Calculating means for calculating the likelihood of dependence on all combinations of words in the tree; and determining the likelihood of dependence on all combinations of partial syntax trees and words calculated by the calculating means with a large likelihood of dependence on the storage device. A reordering means for reordering in the order and storing in the storage device; and a combination of the partial syntax tree and the word having the maximum dependency likelihood among the combinations of the partial syntax tree and the words reordered by the reordering means. When the preceding word is a verb, a case related to the sub-syntactic tree is determined in which cases that can relate to the verb are determined in advance and the same plural cases cannot relate to the same verb. 2. The speech language according to claim 1, further comprising: merging means for judging whether or not it can be related based on the approximation, and merging a combination of a partial parse tree and a word which can be related while satisfying the above constraint. Analysis device. 3. The above-mentioned dependency probability is a statistical probability that a partial parse tree to be processed can be related to a word of a destination obtained based on the corpus, and a statistical probability of the processing target obtained based on the corpus. It is expressed as the product of the likelihood depending on the mutual positional relationship between the partial parse tree and the word to be borrowed, and the sentence end probability obtained after the partial parse tree to be processed, obtained based on the corpus. 3. The speech language analyzer according to claim 2, wherein: