JP2007066237A - Symbol string conversion method, voice recognition method, voice paraphrase method, symbol string conversion device, program and storage medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a symbol string conversion method and device, in which a conversion designation can be easily changed. <P>SOLUTION: Hypothesis generation units for converting an input symbol string to a hypothesis that is a candidate of a target symbol string are cascade-connected in multistage, so that the conversion destination can be freely changed by properly changing the number of stages of the hypothesis generation units. The procedure of extracting a hypothesis whose accumulate weight shows a specific value from hypotheses generated by the hypothesis generation unit of the final stage of the multistage cascade-connected hypothesis generation units and correcting the weight of hypotheses generated by the hypothesis generation unit of the previous stage using the accumulated weight of hypothesis is repeated to the hypothesis generation unit of the first stage. When all input symbol strings are read, the procedure of extracting a hypothesis whose corrected accumulate weight shows the specific value from hypotheses generated by the hypothesis generation unit of the first stage, inputting this hypothesis to the next stage and extracting a hypothesis whose accumulated weight shows the specific value from hypotheses generated in the next stage is repeated to the hypothesis generation unit of the final stage. A hypothesis whose accumulated weight shows the specific value among the hypotheses generated in the final stage is outputted as a conversion result. According to this, conversion accuracy can be improved. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention minimizes the cumulative value of the weights of conversion rules to be applied from a large number of output symbol strings that can be generated for an input symbol string by a symbol string conversion rule expressed by a weighted finite state converter. The present invention also relates to a symbol string conversion method capable of efficiently finding an output symbol string exhibiting the maximum singular value, a speech recognition method using the same, a speech paraphrase method, a symbol string conversion device, a program, and a recording medium.

  Weighted Finite-State Transducer (WFST) is a method for expressing rules for converting a symbol string into another symbol string in a state and state transition diagram. The weighted finite state converter is disclosed in Non-Patent Document 1, for example. Hereinafter, this weighted finite state converter is referred to as a hypothesis generation rule, and a portion that executes this hypothesis generation rule is referred to as a hypothesis generation unit.

The hypothesis generation rule is a set of a state, a state transition indicating that the transition from the state to the state can be performed, an input symbol accepted in the state transition, an output symbol output at that time, and a weight of the state transition Defined by The hypothesis generation rule repeats a state transition while outputting an output symbol according to a state transition that sequentially accepts the symbols of the input symbol sequence from an initial state when a certain input symbol sequence is given, and ends when the end state is reached. It is a calculation model. Formally, the hypothesis generation rule is defined by the following eight sets (Q, Σ, Δ, i, F, E, λ, ρ).
1. Q is a set of finite states.
2. Σ is a finite set of input symbols.
3. Δ is a finite set of output symbols.
4). iεQ is the initial state.
5. FεQ is a set of end states.
6). EεQ × Σ × Δ × Q is a set of state transitions in which an output symbol is output from the current state according to an input symbol to transition to the next state.
7). Λ is the initial weight.
8). p (q) is the end weight of the end state q. qεF.

An example of a hypothesis generation rule is shown in FIG. In FIG. 1, 10 indicates a state represented by a circle (“◯”), and the number in the circle represents the state number. Reference numeral 11 denotes an end state represented by a double circle (“◎”), and the number in the double circle is the number of the end state and the end weight accumulated at the end of the state transition. Is expressed as “(state number) / (end weight)”. Hereinafter, when the state is indicated using the state number, the state and the number are simply referred to as “state 0” or “state 3”.
Reference numeral 12 denotes a state transition represented by an arrow (“→”) connecting each state, and a symbol or a number given to each state transition is an input symbol, an output symbol, or a weight associated with the state transition. Is expressed as “(input symbol) :( output symbol) / (weight)”.

  The hypothesis generation rule of FIG. 1 can also be defined by a table. In FIG. 2, each row represents one state transition, and the state number of the transition source, the state number of the transition destination, the input symbol, the output symbol, and the weight in the state transition are described. The final state (state 3 in FIG. 1) has a format in which the transition destination, the input symbol, and the output symbol are empty, and the weight (end weight) accumulated at the end of the state transition is entered. In general, the initial state of the hypothesis generation rule is state 0, and the initial weight Λ is often omitted. Therefore, in this patent, the initial state is set to state 0, and the initial weight is omitted and not specified.

  The hypothesis generation rule of FIG. 1 can convert, for example, the input symbol string a, a, b, c into the output symbol string d, d, c, b, and the state transition process at that time is a series of state numbers. Is expressed as 0, 0, 1, 3, and the cumulative weight value (hereinafter referred to as “cumulative weight”) is 0.5 + 0.5 + 0.3 + 1 + 0.5 = 2.8. However, in the hypothesis generation rule of FIG. 1, two state transition processes of 0, 0, 1, 3 and 0, 0, 2, 3 are considered for the input symbol strings a, a, b, and c. It is done. In general, when there is a possibility of multiple state transitions for an input symbol string (this is called nondeterminism), the state transition process in which the cumulative weight in the state transition process is the minimum (or maximum) singular value is used. An output symbol string corresponding to the state transition process with the minimum (or maximum) cumulative weight is selected. Whether to choose the state transition process with the smallest cumulative weight or the largest state transition process depends on the nature and purpose of the weight.

ここで、Ｗ（Ｘ→Ｙ，Ｔ）は、仮説生成規則Ｔによって記号列Ｘが記号列Ｙに変換されるときの状態遷移過程における累積重みを表す。この累積重みＷ（Ｘ→Ｙ，Ｔ）の最小値Ｗ（Ｘ）を求めて、その最小値を与えるＹが記号列変換結果となる。
このＷ（Ｘ）を効率的に求めるには、一般にグラフのコスト最小探索の技術の一つである横型探索法を利用する。例えば、グラフの横型探索法の手順は、非特許文献２に開示されている。 Also in the example of FIG. 1, the state transition processes 0, 0, 1, 3 having the smallest cumulative weight are selected for the input symbol strings a, a, b, c, and the conversion results are d, d, c, b. And
When there is a certain hypothesis generation rule (this is assumed to be T) and a symbol string X is given as an input symbol string for this T, an output symbol string (that is, a symbol string conversion result) having a minimum cumulative weight is obtained. Therefore, it is necessary to calculate the minimum value W (X) of the next cumulative weight.

Here, W (X → Y, T) represents the cumulative weight in the state transition process when the symbol string X is converted to the symbol string Y by the hypothesis generation rule T. The minimum value W (X) of the cumulative weight W (X → Y, T) is obtained, and Y giving the minimum value is the symbol string conversion result.
In order to efficiently obtain W (X), a horizontal search method, which is one of the techniques for searching a graph with a minimum cost, is generally used. For example, the procedure of the horizontal search method of the graph is disclosed in Non-Patent Document 2.

The symbol string conversion based on the hypothesis generation rule is performed by searching for a state transition process with the minimum cost (cumulative weight) from the initial state to the end state by the input symbol string.
The weighted finite state converter includes a weighted finite state acceptor (WFSA) that does not define an output symbol and a directed graph that defines neither an input symbol nor an output symbol. In the case of WFSA, the input symbol and output symbol in each state transition are made the same, and in the case of a directed graph, it can be expressed by assuming that all input symbols and output symbols are one symbol. Therefore, all the methods for hypothesis generation rules handled in the present invention can be applied to WFSA and directed graphs.

An example of symbol string conversion using one hypothesis generation rule is shown in FIG. First, in this specification, “hypothesis” means that a certain symbol string is input in order, and the state transition is repeated from the initial state in the hypothesis generation rule for the input symbol string read up to the present time from the initial state. It represents one possible state transition process.
The symbol string conversion unit 102 includes a hypothesis generation rule storage unit 100, a hypothesis generation unit 200, and a hypothesis narrowing unit 400. The hypothesis generation unit 200 reads the hypothesis generation rule from the hypothesis generation rule storage unit 100 and executes a hypothesis generation operation according to the hypothesis generation rule. The hypothesis generation unit 200 updates the state transition process of each hypothesis using the symbol newly received from the symbol string input unit 103 according to the hypothesis generation rule and the newly received set of hypotheses for the symbol string read so far. Generate new hypotheses. All hypotheses generated by the hypothesis generation unit 200 are sent to the hypothesis narrowing unit 400.

  The hypothesis narrowing unit 400 narrows down hypotheses by deleting hypotheses other than the hypothesis having the smallest cumulative weight among hypotheses reaching the same state from the hypothesis set received from the hypothesis generation unit 200. If the input symbol string has been read to the end, the output symbol string corresponding to the hypothesis with the smallest cumulative weight is sent to the symbol string output unit 106. If the input symbol string has not been read to the end, the hypothesis is sent to the hypothesis generation unit 200. The symbol string output unit 106 outputs the output symbol string received from the hypothesis narrowing unit 400.

Next, an example of a procedure for converting a symbol string based on this embodiment will be described.
First, when a state transition having a hypothesis generation rule is represented as e, n [e] is a transition destination state (next state), i [e] is an input symbol, o [e] is an output symbol, and w [e] is Define as weight. In addition, when a certain hypothesis is expressed as h, a state in which s [h] is reached, W [h] is a cumulative weight in the state transition process, and O [h] is a symbol string output in the state transition process. To do.
In this procedure, hypotheses are managed using a list of hypotheses (hereinafter referred to as a hypothesis list). Hypotheses can be inserted into and extracted from the hypothesis list. However, when a hypothesis is inserted into the hypothesis list, if there is a hypothesis that reaches the same state in the hypothesis list, only the smaller cumulative weight is left in the hypothesis list to narrow down the hypotheses.

The symbol string conversion procedure using the hypothesis generation rule is shown in FIG.
Hereinafter, the procedure of FIG. 4 will be described in comparison with an embodiment (FIG. 3) of a symbol string conversion method provided with a hypothesis generation rule.
Starting from step S101, as an initial setting, empty hypothesis lists H and H ′ are generated in step S102. An initial hypothesis h is generated, and s [h] = 0 (initial state of the hypothesis generation rule), W [h] = 0, and O [h] = φ (empty symbol string) are inserted into the hypothesis list H. .
In step S102, the symbol string input unit 103 reads one symbol and substitutes the symbol for x. The following steps S105 to S108 are executed in the hypothesis generation unit 104.

In step S105, one hypothesis is extracted from the hypothesis list H and substituted into the hypothesis h, and a list E of state transitions whose input symbol is equal to x is prepared from the state s [h].
If E = φ (empty list) in step S106, the process proceeds to step S110. Otherwise, the process proceeds to step S107, and one state transition is extracted from the state transition list E and substituted into the state transition e.
In step S108, a new hypothesis f is generated, and s [f] = n [e], W [f] = W [f] + w [e], O [f] = O [h] · o [e] To do. Here, “·” represents an operation of connecting two symbols or a symbol string into one symbol string.

Step S109 is executed by the hypothesis narrowing unit 105 to narrow down the hypotheses by inserting the hypothesis f into the hypothesis list H ′.
Returning from step S109 to step S106, a hypothesis is developed for the next state transition.
In step S110, if the hypothesis list H = φ (all hypotheses have been expanded), the process proceeds to step S111. Otherwise, the process returns to step S105, and the next hypothesis is developed.
In step S111, all the elements of the newly generated hypothesis list H ′ are moved to the already empty hypothesis list H, and the process proceeds to step S112.

In step S112, if the next input symbol exists in the symbol string input unit 103, the process returns to step S104. Otherwise, it is determined that all the input symbol strings have been read, and the process proceeds to step S113.
In step S113, after adding the end weight of the end state to the cumulative weight of the hypothesis reaching the end state in the hypothesis list H, the cumulative weight (W The hypothesis h that minimizes [h]) is selected, and the output symbol string O [h] is output as a symbol string conversion result in the symbol string output unit 106.

In step S114, the symbol string conversion procedure using the hypothesis generation rule is terminated.
The process of obtaining the output symbol string when the input symbol strings a, a, b, and c are given to the hypothesis generation rule of FIG. 1 according to this symbol string conversion procedure will be described in order. However, if there is a certain hypothesis (current state number s, output symbol string O, cumulative weight W), the hypothesis is represented as (s, O, W). Further, a state transition with a hypothesis generation rule (current state number s, next state number n, input symbol x, output symbol y, weight w) is represented as <s → n, x: y / w>.

Starting from step S101, empty hypothesis lists H and H ′ are created in step S102.
In step S103, a hypothesis (0, φ, 0) is inserted into the hypothesis list H.
The symbol a is read in step S104, and a hypothesis (0, φ, 0) is extracted from the hypothesis list H in step S105. A state transition list E including a state transition <0 → 0, a: d / 0.5> whose input symbol is equal to a is created from the current state 0 of this hypothesis.
Since E = φ is not satisfied in step S106, the process proceeds to step S107, the state transition <0 → 0, a: d / 0.5> is extracted, and a new hypothesis (0, d, 0.5) is generated in step S108. In step S109, it is inserted into the hypothesis list H ′.

Returning to step S106, since E = φ, the process proceeds to step S110, and since hypothesis list H = φ, the process proceeds to step S111. The element (0, d, 0.5) of H ′ is moved to H, and since the next input symbol exists in step S112, the process returns to step S104.
Subsequently, the symbol a is read in step S104, and a hypothesis (0, d, 0.5) is extracted from the hypothesis list H in step S105. A state transition list E including a state transition <0 → 0, a: d / 0.5> whose input symbol is equal to a is generated from the current state 0 of this hypothesis.
Since E = φ is not satisfied in step S106, the process proceeds to step S107, and the state transition <0 → 0, a: d / 0.5> is extracted from the state transition list E. A new hypothesis (0, dd, 1) is generated in step S108, and inserted into H ′ in step S109.

Returning to step S106, since the state transition list E = φ, the process proceeds to step S110, and since H = φ, the process proceeds to step S111. The element (0, dd, 1) of the hypothesis list H ′ is moved to the hypothesis list H. Since the next input symbol exists in step S112, the process returns to step S104.
Subsequently, the symbol b is read in step S104, and a hypothesis (0, dd, 1) is extracted from the hypothesis list H in step S105. A state transition list E including state transitions <0 → 1, b: c / 0.3> and <0 → 2, b: b / 1> whose input symbol is equal to b is created from the current state 0 of this hypothesis.

Since E = φ is not satisfied in step S106, the process proceeds to step S107, and the state transition <0 → 1, b: c / 0.3> is extracted from the state transition list E. A new hypothesis (1, ddc, 1.3) is generated in step S108, and inserted into H ′ in step S109.
Returning to step S106, since E = φ is not true, the process proceeds to step S107, and the state transition <0 → 2, b: b / 1> is extracted from the state transition list E. A new hypothesis (2, ddb, 2) is generated in step S108, and inserted into H ′ in step S109.
Returning to step S106, since E = φ, the process proceeds to step S110, and since H = φ, the process proceeds to step S111. Elements (1, ddc, 1.3) and (2, ddb, 2) of H ′ are hypothesis lists. Since the next input symbol exists in step S112, the process returns to step S104.

Subsequently, the symbol c is read in step S104, and a hypothesis (1, ddc, 1.3) is extracted from the hypothesis list H in step S105. A state transition list E including a state transition <1 → 3, c: b / 1> whose input symbol is equal to c is created from the current state 1 of this hypothesis.
Since E = φ is not satisfied in step S106, the process proceeds to step S107, and the state transition <1 → 3, c: b / 1> is extracted from the state transition list E. A new hypothesis (1, ddcb, 2.3) is generated in step S108, and inserted into H ′ in step S109.
Returning to step S106, since E = φ, the process proceeds to step S110. Since H ≠ φ, the process returns to step S105, and a hypothesis (2, ddb, 2) is extracted from the hypothesis list H. A state transition list E including a state transition <2 → 3, c: a / 0.6> whose input symbol is equal to c is created from the current state 2 of this hypothesis.

  Since E = φ is not satisfied in step S106, the process proceeds to step S107, and the state transition <2 → 3, c: a / 0.6> is extracted from the state transition list E. In step S108, a new hypothesis (3, ddba, 2.6) is generated and inserted in the hypothesis list H 'in step S109. At this time, the hypothesis list H ′ already includes the hypothesis (3, ddcb, 2.3), and the hypothesis (3, ddba, 2.6) has reached the same state 3, so The hypothesis (3, ddcb, 2.3) with a small weight is left, and the hypothesis (3, ddba, 2.6) is deleted from the list.

Returning to step S106, since E = φ, the process proceeds to step S110, and since hypothesis list H = φ, the process proceeds to step S111, the element (3, ddcb, 2.3) of the hypothesis list H ′ is moved to H, and step S112 is performed. In step S113, the next input symbol does not exist.
In step S113, since the arrival state 3 of the hypothesis (3, ddcb, 2.3) in the hypothesis list H is an end state, the end weight is added to (3, ddcb, 2.8), and this hypothesis ends. Since this is the only hypothesis that has reached the state and the accumulated weight is minimum, the output symbol string ddcb is output as a conversion result, and the symbol string conversion process is terminated in step S114.

を計算し、このＷ（Ｘ）を与えるＹとＺの組のうちＺを記号列変換結果とする。 Next, there are two hypothesis generation rules, and consider the problem of converting a symbol string in order. That is, when there are two hypothesis generation rules, T ₁ and T ₂ and an input symbol string X is given, the symbol string X is _first converted to the symbol string Y using T ₁ , and the symbol string Y as an input symbol string of T _2, further it means that the conversion to obtain an output symbol sequence Z. In the theory of hypothesis generation rules, the problem is that all Y that can be the output symbol string when transforming X by T ₁ is considered as an input symbol string of T ₂ , and there is a possibility for these input symbol strings. From a set of output symbol strings Z of a certain T ₂ , there is a problem of obtaining a conversion result that minimizes the sum of the accumulated weight of the state transition process at T ₁ and the accumulated weight of the state transition process at T ₂ . Therefore,

And Z in the set of Y and Z that gives W (X) is the symbol string conversion result.

  As an example, consider the two hypothesis generation rules shown in FIGS. FIG. 5 is an example of a hypothesis generation rule for converting a kana symbol string into a kanji symbol string. However, the symbol “ε” appearing in FIG. 5 indicates that a state transition is made without an input symbol when it is on the left side of “:”, and nothing is output in the state transition when it is on the right side.

  In this hypothesis generation rule, for example, the symbol string “A, ME” is converted to “rain” by the state transition processes 0, 1, 5 (cumulative weight 1), and “A, ME, DA, MA” is the state transition process. It is converted into “rain, ball” by 0, 1, 5, 0, 4, 5 (cumulative weight 3). However, in Japanese, the kanji corresponding to “Amedama” is generally “Kamatama” rather than “Amatama”, but the hypothesis generation rule of FIG. Since the cumulative weight of the state transition processes 0, 2, 5, 0, 4, and 5 to be output is 4 while the cumulative weight to be converted to “rain, ball” is 3, the cumulative weight is small. The conversion result is “rain, ball”.

  On the other hand, FIG. 6 shows a hypothesis generation rule that gives weight to kanji concatenation. In this hypothesis generation rule, the input symbol and the output symbol are the same in all state transitions, that is, the symbol string is output as it is without conversion, but the symbol concatenation of the input symbol string (Kanji string) is Japanese in the state transition process. The weight is small when it is plausible, and the weight is large when it is not plausible. For example, the cumulative weights of the state transition processes 0, 1, and 3 that accept the input symbol string “rain, descend” are 0, and the cumulative weights of the state transition processes 0, 2, and 3 that accept the input symbol string “飴, descend” Is 3, indicating that the input symbol string “rain, rain” is the most likely connected symbol string in Japanese.

When the symbol string is converted using the hypothesis generation rules of FIGS. 5 and 6, for example, when the kana symbol string “A, ME, DA, MA” is input, the hypothesis generation rule of FIG. 5 is obtained. The output symbol string may be “rain, ball” with a cumulative weight of 3 and “carp, ball” with a cumulative weight of 4. When these output symbol strings are respectively input to the hypothesis generation rule of FIG. 6, the cumulative weight 5 for “rain, ball” becomes 1, and the cumulative weight for “carp, ball” becomes 1, and the two hypothesis generations of FIG. 5 and FIG. The sum of the cumulative weights according to the rules is calculated as follows.
Cumulative weight in “A, Me, Da, Ma” → “Rain, Jade” 3 + 5 = 8
Cumulative weight in “A, Me, Da, Ma” → “Aoi, Jade” 4 + 1 = 5
As a result of comparing these accumulated weights, an output symbol string having a smaller accumulated weight is “飴, ball”. In order to obtain such a conversion result, the hypothesis generation rule of FIG. 6 having a weight related to kanji connection is indispensable.

  FIG. 7 shows an example of conventional symbol string conversion using two hypothesis generation rules. First, when performing two-stage symbol string conversion using two hypothesis generation rules, the part that executes the hypothesis generation rule used in the previous stage is the preceding hypothesis generation part, and the part that executes the hypothesis generation rule used in the subsequent stage is the subsequent stage. It will be called a hypothesis generation unit. The “hypothesis” in the case of using two hypothesis generation rules means that a certain symbol string is input in order and the input symbol string read from the initial state in the previous hypothesis generation unit with respect to the input symbol string read so far. One state transition process that may occur when the state transition is repeated by, and an output symbol string corresponding to the state transition process of the preceding hypothesis generation unit as an input symbol string of the subsequent hypothesis generation unit Let us denote a set of two state transition processes.

  As another means of converting a symbol string using two hypothesis generation rules, a method of combining two hypothesis generation rules in advance into one hypothesis generation rule and applying the symbol string conversion procedure for one hypothesis generation rule There is. For example, Non-Patent Document 3 discloses a method for synthesizing a hypothesis generation rule. However, when a hypothesis generation rule is synthesized, states and state transitions can be made with respect to a combination of two state transitions, so the number of hypotheses generation rules and the number of state transitions may become very large. Therefore, when handling hypothesis generation rules in a computer, it may be difficult to execute symbol string conversion due to restrictions such as memory size.

  On the other hand, in the standard method of performing the minimum cost search using two hypothesis generation rules, when each hypothesis is updated and a new hypothesis is generated, possible state transitions in the previous hypothesis generation rule, Since new hypotheses are generated for all combinations of possible state transitions in the hypothesis generation rule, there is a problem that the number of hypotheses increases and the amount of calculation increases. An efficient symbol string conversion method using two hypothesis generation rules to solve this has been proposed. This method is described with respect to both a standard symbol string conversion method based on a minimum cost search and an efficient symbol string conversion method when two hypothesis generation rules are used. It is shown that when these methods are applied to speech recognition, the conversion time of the efficient symbol string conversion method relative to the standard symbol string conversion method is reduced to about one third.

Hereinafter, an efficient symbol string conversion method using the two previously proposed hypothesis generation rules will be described.
A symbol string input unit for reading a symbol string in order, a preceding hypothesis generation unit used in the previous stage for converting the symbol string in two stages using two hypothesis generation rules for converting the symbol string by state transition, and a subsequent hypothesis used in the subsequent stage A symbol string conversion unit that converts the symbol string using the generation unit, and a symbol string output unit that outputs a conversion result by the latter-stage hypothesis generation unit,
When symbols are read in order from the symbol string input unit and the input symbol string is read, cumulative values of weights (cumulative weights) for state transitions applied respectively in the preceding and succeeding hypothesis generation units in the symbol string conversion unit In the symbol string conversion method for outputting the output symbol string corresponding to the state transition process of the subsequent hypothesis generation unit that minimizes from the symbol string output unit,
While reading the symbols in order, the post-stage hypothesis generation when the output symbol string in the hypothesis state transition process is the input symbol string of the post-hypothesis generation section is the cumulative weight for the hypothesis representing one state transition process of the pre-stage hypothesis generation section The state transition process that minimizes the cumulative weight among the possible state transition processes in the unit is obtained, and the cumulative weight is corrected by adding the cumulative weight to the hypothetical cumulative weight.

  When the input symbol string is completely read, it is possible in the latter hypothesis generator when the hypothesis with the minimum or maximum accumulated weight and the output symbol string corresponding to the state transition process of that hypothesis are used as the input symbol string of the latter hypothesis generator. The symbol string conversion method is characterized in that an output symbol string corresponding to the state transition process having the smallest cumulative weight among the various state transition processes is used as a symbol string conversion result. Convert the column.

  Hereinafter, an efficient symbol string conversion method using two hypothesis generation units will be described with reference to FIG. The symbol string input unit 103 reads the symbols in order and sends them to the symbol string converter 102. In this example, the symbol conversion unit 102 includes a hypothesis generation unit 201-1 and 201-2, a hypothesis correction unit 301-1 and 301-2, and a hypothesis narrowing unit 401-1 and 401-2. In the hypothesis generation units 201-1 and 201-2, the preceding hypothesis generation unit 201-1 determines the hypothesis according to the preceding hypothesis generation rule read from the preceding hypothesis generation rule storage unit 101-1 of the symbol received from the symbol string input unit 103. Generate. The subsequent hypothesis generation unit 201-2 updates the hypothesis generated by the previous hypothesis generation unit 201-1 with the newly received symbol in accordance with the subsequent hypothesis generation rule read from the subsequent hypothesis generation rule storage unit 101-2. To generate new hypotheses and send them to hypothesis correction units 301-1 and 301-2. In the hypothesis correction units 301-1 and 301-2, possible state transition processes in the subsequent hypothesis generation unit when the output symbol string in the hypothesis state transition process is the input symbol string of the subsequent hypothesis generation unit 201-2. The state transition process in which the cumulative weight is the minimum is obtained, and the cumulative weight is corrected by adding it to the hypothetical cumulative weight, and is sent to the hypothesis narrowing down sections 401-1 and 401-2.

  Hypothesis narrowing-down sections 401-1 and 401-2 leave only hypotheses having the smallest cumulative weight among hypotheses with overlapping state transition processes from the hypotheses received from hypothesis correction sections 301-1 and 301-2. Refine the hypotheses by deleting other hypotheses. If the input symbol string has been read to the end, the subsequent hypothesis when the hypothesis having the smallest cumulative weight and the output symbol string corresponding to the state transition process of the hypothesis are used as the input symbol string of the subsequent hypothesis generation unit 201-2. An output symbol string for a state transition process having a minimum accumulated weight among the possible state transition processes in the generation unit 201-2 is sent to the symbol string output unit 106. If the input symbol string has not been read to the end, the remaining hypotheses are sent to the hypothesis generation unit 201-1. The symbol string output unit 106 outputs the output symbol string received from the hypothesis narrowing unit 401-2.

FIG. 8 shows a specific procedure of an efficient symbol string conversion method by the two hypothesis generation units 201-1 and 201-2.
First, when the “hypothesis” in the preceding hypothesis generation unit 201-1 and the symbol string output from the state transition process of the hypothesis are used as the input symbol string of the subsequent hypothesis generation unit 201-2, the subsequent hypothesis generation is performed. One possible state transition process of the unit 201-2 is distinguished from the “hypothesis” and is referred to as an “auxiliary hypothesis”. Since at least one auxiliary hypothesis always exists for one hypothesis, in this embodiment, a set of auxiliary hypotheses for the hypothesis h is represented by an auxiliary hypothesis list L [h].

Further, when there is a hypothesis h, s ₁ [h] is the state reached last in the state transition process of hypothesis h, and W [h] is the (corrected) cumulative weight in the state transition process of hypothesis h. On the other hand, when there is an auxiliary hypothesis g (gεL [h]) with respect to hypothesis h, s ₂ [g] is the last state reached in the state transition process of auxiliary hypothesis g, and W [g] is the hypothesis h. The sum of the cumulative weight in the state transition process and the cumulative weight in the state transition process of the auxiliary hypothesis g, O [g], is a symbol string output in the state transition process of the auxiliary hypothesis g. Then, the sets of end states of the hypothesis generation units 201-1 and 201-2 are set as F ₁ and F ₂ , respectively.

Hereinafter, the procedure of FIGS. 8 and 9 will be described in comparison with the embodiment of FIG.
Starting from step S201, as an initial setting, empty hypothesis lists H and H ′ are generated in step S202, an initial hypothesis h is generated, and s ₁ [h] = 0, W [h] = 0, L [ h] = φ (empty list) and insert into hypothesis list H. Also, an auxiliary hypothesis g is generated, and s ₂ [g] = 0, W [g] = 0, O [g] = φ (empty symbol string) is inserted into the auxiliary hypothesis list L [h] of hypothesis h To do.
In step S204, the symbol string input unit 103 reads one symbol and substitutes the symbol for x. The following steps S205 to S208 are executed in the hypothesis generation unit 201.

In step S205, substituted hypothesis taken out one hypothesis from hypothesis list H h, to generate a list E ₁ transition possible state transitions in the input symbol x from the state s ₁ [h].
If E ₁ = φ (empty list) in step S206, the process proceeds to step S211. Otherwise, the process proceeds to step S207.
In step S207, fetches one state transition from _{E 1,} is substituted into _{e 1,} generates new hypotheses f at step _{S208, s 1 [f] =} n [e 1], W [f] = W [h] + w [e ₁ ], L [f] = L [h].

In step S209, in the hypothesis correction unit 301, if o [e ₁ ] = ε, means for correcting the hypothesis f (described separately after the description of this procedure) is executed.
In step S210, the hypothesis narrowing unit 401 narrows down the hypotheses by inserting the hypothesis f into the hypothesis list H ′.
In step S211, if H = φ (all hypotheses have been developed), the process proceeds to step S212. Otherwise, the process returns to step S305 to generate the next hypothesis.
In step S212, all the elements of the newly generated hypothesis list H ′ are moved to the already empty hypothesis list H, and the process proceeds to step S213.

を計算し、ｇの累積重みＷ[ｇ]が最小の補助仮説ｇ’を選び、その出力記号列Ｏ[ｇ’]を変換結果として記号列出力部１０６において出力する。
ステップＳ２１５にて二つの仮説生成部２０１−１と２０１−２を用いた記号列変換手順を終了する。 In step S213, if there is a next input symbol in the symbol string input unit 103, the process returns to step S204. Otherwise, it is determined that all the input symbol strings have been read, and the process proceeds to step S214.
In step S214, the end weight ρ (s ₁ [h]) is set to the cumulative weight W [h] for all hypotheses h (s ₁ [h] εF ₁ ) that have reached the end state from the hypothesis list H. In addition, the auxiliary hypothesis g (s ₂ [g] ∈F ₂ ) that has reached the end state in the auxiliary hypothesis list L [h] of the auxiliary hypothesis h

The auxiliary hypothesis g ′ having the smallest accumulated weight W [g] of g is selected, and the output symbol string O [g ′] is output as a conversion result in the symbol string output unit 106.
In step S215, the symbol string conversion procedure using the two hypothesis generation units 201-1 and 201-2 is terminated.

となり、その補正累積重みが最小である仮説の補正累積重みは

となり、先に述べた式（２）における二つの仮説生成部の累積重みの和の最小値と一致する。 Next, a hypothesis correction procedure in the hypothesis correction units 301-1 and 301-2 will be described.
The correction of the hypothesis is a subsequent stage when the cumulative weight for the state transition process of the hypothesis in the preceding hypothesis generation unit 201-1 is used as the output symbol string in the state transition process of the hypothesis as the input symbol string of the subsequent hypothesis generation unit 201-2. This means that the hypothesis generation unit 201-2 obtains a state transition process having the smallest cumulative weight among the possible state transition processes, and adds the cumulative weight to the hypothetical cumulative weight. When the input symbol string X is read, the corrected cumulative weight of the hypothesis that accepts the symbol string X and outputs the symbol string Y is:

The corrected cumulative weight of the hypothesis with the minimum corrected cumulative weight is

Thus, the value coincides with the minimum value of the sum of the cumulative weights of the two hypothesis generation units in the above-described equation (2).

Therefore, by using the corrected cumulative weight, the output symbol string Y of the preceding hypothesis generation unit that minimizes it is obtained, and the possible state transition process when Y is used as the input symbol string of the subsequent hypothesis generation unit Among them, the output symbol string corresponding to the state transition process with the smallest cumulative weight is the symbol string conversion result.
In this symbol string conversion method, correction of hypotheses is efficiently performed using an auxiliary hypothesis list associated with each hypothesis. The hypothesis to be corrected is f, the auxiliary hypothesis list is L [f], and the output symbol in the last state transition of the hypothesis f is y.

From the state s ₂ [g] (g represents one auxiliary hypothesis included in L [f]) of the post-stage hypothesis generation unit 201-2, the state n [e ₂ ] is passed through the state transition e ₂ by the input symbol y. An auxiliary hypothesis j that reaches is generated, and its arrival state is s [j] = n [e ₂ ]. In the state transition process of hypothesis f, the latest state transition before the last state transition (e ₁ ) of hypothesis f and whose output symbol is not ε is denoted by t.
At this time, the cumulative weight W [j] of the auxiliary hypothesis j can be calculated by adding the following three elements.

(A1) Cumulative weight of auxiliary hypothesis g generated in state n [t] in the state transition process of hypothesis f:
This value is equal to W [g]. That is, in the state transition process in which the output symbol is ε, the auxiliary hypothesis list is directly inherited by the hypothesis of the subsequent state transition process, so that the auxiliary hypothesis list generated in the state n [t] is equal to L [f]. The cumulative weight W [g] of each auxiliary hypothesis in the auxiliary hypothesis list L [f] is used as it is.

に等しく、これを現在の仮説ｆの累積重みＷ[ｆ]から差し引くことにより、状態ｎ[ｔ]から状態ｓ_１[ｆ]までの間の累積重みが求まる。すなわち、

である。 (A2) Weights accumulated from state n [t] to state s ₁ [f]:
However, in the state transition process in which the output symbol is ε, the auxiliary hypothesis list is carried over to the newly generated hypothesis as it is, so the corrected cumulative weight of the hypothesis that has reached the state n [t] is the current auxiliary hypothesis. Minimum cumulative weight in hypothesis list L [f]

And subtracting this from the cumulative weight W [f] of the current hypothesis f, the cumulative weight between state n [t] and state s ₁ [f] is obtained. That is,

It is.

を代入することにより行われる。 (A3) Weight in state transition e ₂ from state s ₂ [g] to state s ₂ [j] of post-stage hypothesis generation unit 201-2: w [e ₂ ]
Therefore, the cumulative weight of the auxiliary hypothesis j is calculated as W [j] = W [g] + AW [f] + w [e ₂ ].
Then, the correction of the hypothesis f is performed by the minimum value of the cumulative weights W [j] of the auxiliary hypothesis j newly generated as W [f].

This is done by substituting

The hypothesis correction procedure will be described below with reference to FIG. This hypothesis correction procedure is executed in step S208 shown in FIGS. The hypothesis corrected in this procedure is f, and the output symbol from the previous hypothesis generation unit is y (= o [e ₁ ]).
In step S301, a procedure for correcting hypothesis f is started.
In step S302, an empty auxiliary hypothesis list L ′ is generated.
In step S303, the weight correction term AW [f] is obtained as AW [f] = W [f] −W [g ′]. However, g ′ indicates an auxiliary hypothesis having the smallest cumulative weight in the auxiliary hypothesis list L [f] of the hypothesis f.
In step S304, one auxiliary hypothesis is extracted from L [f] and substituted into g. A list E _{2 of} state transitions that can be transitioned with the input symbol y is generated from the state s ₂ [g], and the process proceeds to step S305.

In step S305, taken out one state transition from the list _{E 2,} it is substituted into _{e 2.}
In step S306, an auxiliary hypothesis j is generated, and s ₂ [j] = n [e ₂ ], W [j] = W [g] + AW [f] + w [e ₂ ], O [j] = O [g ] · O [e ₂ ].
In step S307, the auxiliary hypothesis j is inserted into the auxiliary hypothesis list L ′.
In step S308, if E ₂ is empty (the auxiliary hypothesis has been expanded for all state transitions from s ₂ [g]), the process proceeds to step S309. Otherwise, the process returns to step S305 to check the next state transition.

In step S309, if L [f] is empty (all auxiliary hypotheses have been expanded), the process proceeds to step S310. Otherwise, the process returns to step S304 to develop the next auxiliary hypothesis.
In step S310, all the elements of the newly generated auxiliary hypothesis list L ′ are moved to the already empty auxiliary hypothesis list L ′ [f], and the process proceeds to step S311.
In step S311, W [f] = W [g ″] is set for the auxiliary hypothesis g ″ having the smallest cumulative weight in the auxiliary hypothesis list L [f].
In step S312, the hypothesis correction procedure ends.

のように補正する。この補正は、先に示した仮説の補正手順における（Ａ１）と（Ａ２）の項を加えたものに等しい。そして、
Ｗ[ｈ]＜Ｗ[ｆ] ならば Ｌ[ｈ]＝Ｌ[ｈ]＋Ｌ[ｆ]、ｆを削除
Ｗ[ｈ]＞Ｗ[ｆ] ならば Ｌ[ｆ]＝Ｌ[ｈ]＋Ｌ[ｆ]、ｈを削除
のように補助仮説リストを連結する。 Further, in the symbol string conversion method using these two hypothesis generation units, when a hypothesis is inserted into the hypothesis list, as in the symbol string conversion method using one hypothesis generation unit, the preceding hypothesis generation unit includes If there is a hypothesis that reaches the same state, the hypothesis is narrowed down by leaving the smaller cumulative weight in the hypothesis list. However, since the auxiliary hypotheses may be lost by this hypothesis narrowing down, the auxiliary hypothesis list of the narrowed down hypotheses is connected to the auxiliary hypothesis list of the remaining hypotheses. At that time, it is necessary to correct the cumulative weight of the auxiliary hypotheses in the auxiliary hypothesis list. This means that when hypothesis h and hypothesis f have reached the same state in the previous hypothesis generator, the cumulative weight of the auxiliary hypothesis is

Correct as follows. This correction is equivalent to the addition of the terms (A1) and (A2) in the hypothesis correction procedure described above. And
If W [h] <W [f] L [h] = L [h] + L [f], delete f If W [h]> W [f] L [f] = L [h] + L [ f], concatenate the auxiliary hypothesis lists like h is deleted.

  The symbol string conversion procedure (FIGS. 8 and 9) is substantially the same calculation procedure as the procedure (FIG. 4) for searching using only the preceding hypothesis generation unit except for the hypothesis correction procedure (step S209). Compared with the conversion procedure, the amount of calculation can be reduced. Although there is a calculation load in the hypothesis correction procedure, since the hypothesis correction procedure is executed only when the output symbol is not ε in the state transition of the preceding hypothesis generation unit, the rate at which the output symbol of the previous hypothesis generation unit is ε The greater the amount, the greater the effect of processing volume reduction.

Next, according to the symbol string conversion procedure of FIGS. 8 and 9 using two hypothesis generation units, the hypothesis generation rule of FIG. 5 is T ₁ and the hypothesis generation rule of FIG. 6 is T _2. The process of obtaining the output symbol string when, da, “is given will be explained step by step. However, the hypothesis (current state number _{s 1} hypothesis generation rule _{T 1,} the output symbol sequence O, cumulative weight W, auxiliary hypothesis list L) is expressed as _(s 1, O, W, L), the auxiliary hypothesis (hypothesis The current state number s ₂ , the output symbol string O, and the cumulative weight W) of the generation rule T ₂ are expressed as (s ₂ , O, W). The auxiliary hypothesis list L is expressed as {(s ₂ , O, W), (s ₂ ′, O ′, W ′),.

[Start the symbol conversion procedure using two hypothesis generation rules]
Starting from step S201, empty hypothesis lists H and H ′ are created in step S202.
In step S203, the hypothesis (0, φ, 0, {(0, φ, 0)}) is inserted into the hypothesis list H.

[Processing of input symbol “A”]
In step S204, the symbol “a” is read, and in step S205, a hypothesis (0, φ, 0, {(0, φ, 0)}) is extracted from the hypothesis list H.
State transition <0 → 1, a: rain / 0> from the current state 0 of the hypothesis generation rule T ₁ of this hypothesis with the input symbol “a”
<0 → 2, A: 飴 / 0>
Make a list _{E 1,} including.
Since _{E 1} = not at φ at step S206 proceeds to step S207, the state transition from the state transition process list _{E 1} <0 → 1, Ah: Rain / 0> was removed, a new hypothesis in step S208 f = (1, rain , 0, {(0, φ, 0)}). Since the output symbol of the state transition <0 → 1, a: rain / 0> is not ε in step S209, the hypothesis f = (1, rain, 0, {(0, φ, 0)}) is corrected. Proceed to (FIG. 10).

In step S302, an empty auxiliary hypothesis list L ′ is generated.
In step S303, AW [f] = 0.
Auxiliary hypothesis list in step S304 L = {(0, φ , 0)} from the auxiliary hypotheses (0, phi, 0) was taken out, transition possible state transitions in the input symbol "rain" from the state 0 of hypothesis generation rule _{T 2} <0 → 1, rain: rain / 0>,
<0 → 3, rain: rain / 0>
Make a state transition process list E _2, including.
Since the state transition process list E ₂ = φ is not satisfied in step S305, the process proceeds to step S306. State transition from the state transition process list _{E 2} <0 → 1, rain: Rain / 0> was removed, auxiliary hypotheses in step S307 (1, rain, 0) generates the auxiliary hypotheses (1 in step S308, rain, 0 ) Is inserted into L ′.

Returning to step S305, since the state transition process list E ₂ = φ is not satisfied, the process proceeds to step S306. State transition process list _{E 2} state transition from <0 → 3, rain: Rain / 0> was removed, step S307 in the auxiliary hypothesis (3, rain, 0) to generate a step S308 in the auxiliary hypothesis (3, rain, 0 ) Is inserted into L ′.
Returning to step S305, since the state transition process list E ₂ = φ, the process proceeds to step S309. Since the auxiliary hypothesis list L = φ, the process proceeds to step S310, and the elements (1, rain, 0) and (3, rain, 0) of the auxiliary hypothesis list L ′ are moved to L.
In step S311, the hypothesis cumulative weight becomes 0, and the hypothesis correction is terminated in step S312. As a result, the hypothesis is (1, rain, 0, {(1, rain, 0), (3, rain, 0)}).

Returning to step S210, the hypothesis (1, 0, {(1, rain, 0), (3, rain, 0)}) is inserted into the hypothesis list H ′.
Because it is not a state transition process list _{E 1} = phi returns to step S206 proceeds to step S207, the state transition from the state transition process list _{E 1} <0 → 2, Ah: candy / 0> was removed, new hypotheses in step S208 f = (2, 飴, 0, {0, φ, 0}), and since the output symbol of the state transition <0 → 2, a: ＜ / 0> is not ε in step S209, the hypothesis f = (2 , 飴, 0, {(0, φ, 0)}), the process proceeds to step S301.
In step S302, an empty auxiliary hypothesis list L ′ is generated.
In step S303, the cumulative weight AW [f] = 0.

Auxiliary hypothesis list in step S304 L = {(0, φ , 0)} from the auxiliary hypotheses (0, phi, 0) was taken out, transition possible state transitions in the input symbol "rain" from the state 0 of hypothesis generation rule _{T 2} <0 → 2, 飴: 飴 / 0>,
<0 → 4, 飴: 飴 / 0>
Make a list _{E 2,} including.
Since the list E ₂ = φ is not satisfied in step S305, the process proceeds to step S306. State transition from the list _{E 2} <0 → 2, candy: candy / 0> was removed, auxiliary hypotheses in step S307 (2, candy, 0) generates the auxiliary hypotheses in step S308 (2, candy, 0) auxiliary Insert into hypothesis list L ′.

Returning to step S305, since the state transition process list E ₂ = φ is not satisfied, the process proceeds to step S306. E ₂ state transition from <0 → 4, candy: candy / 0> was removed, step S307 in the auxiliary hypothesis (4, candy, 0) generates the auxiliary hypothesis (4, candy, 0) at step S308 auxiliary hypothesis Insert into list L '.
Returning to step S305, since the state transition process list E ₂ = φ, the process proceeds to step S309. Since the auxiliary hypothesis list L = φ, the process proceeds to step S310, and the elements (2, 飴, 0) and (4, 飴, 0) of the auxiliary hypothesis list L ′ are moved to L.
In step S311, the hypothesis cumulative weight becomes 0, and the hypothesis correction is terminated in step S312. As a result, the hypothesis is (2, 0, {(2, 飴, 0), (4, 飴, 0)}).

Returning to step S210, the hypothesis (2, 飴, 0, {(2, 飴, 0), (4, 飴, 0)}) is inserted into the hypothesis list H ′.
Returning to step S206, since E ₁ = φ, the process proceeds to step S211. Since the hypothesis list H = φ, the process proceeds to step S212 and the hypotheses (1, rain, 0, {(1, rain, 0), (3, rain, 0)}) and (2, 飴, 0) in the hypothesis list H ′. 0, {(2, 飴, 0), (4, 飴, 0)}) are moved to the hypothesis list H, and since the next input symbol exists in step S213, the process returns to step S204.

[Processing of input symbol “me”]
Subsequently, the symbol “me” is read in step S204, and a hypothesis (1, rain, 0, {(1, rain, 0), (1, 1, rain, 0)}) is extracted from the hypothesis list H in step S205. . State transition of the input symbol is "because" from the current state 1 of hypothesis generation rule T ₁ of this hypothesis <1 → 5, because: ε / 1>
Make a state transition process list E _1, including.
Because it is not a state transition process list _{E 1} = phi at step S206 proceeds to step S207, the state transition from the state transition process list _{E 1 <1} → 5, because: epsilon / 1> was removed, a new hypothesis in step S208 (5 , Rain, 1, {(1, rain, 0), (3, rain, 0)}), and the process proceeds to step S209.

Since the output symbol of the state transition <1 → 5, ε / 1> is ε, the process proceeds to the next step S210 and the hypothesis (5, rain, 1, {(1, rain, 0), (3, rain, 0)}) is inserted into the hypothesis list H ′.
Returning to step S206, since the state transition process list E ₁ = φ, the process proceeds to step S211. Since the hypothesis list H ≠ φ, the process returns to step S205, and the hypothesis (2, 飴, 0, {(2, 飴, 0), (4, 飴, 0)}) is extracted from the hypothesis list H. State transition from the current state 2 of this hypothesis to the input symbol “M” (<2 → 5, M: ε / 2>, φ)
Make a state transition process list E _1, including.

Since the state transition process list _{E 1} = not at φ at step S206 proceeds to step S207, the state transition from the state transition process list _{E 1} <2 → 5, because: epsilon / 2> was removed, a new hypothesis in step S208 (5 , 飴, 2, {(2, 飴, 0), (4, 飴, 0)}), and the process proceeds to step S209.
Since the output symbol of the state transition <2 → 5: ε / 2> is ε, the process proceeds to the next step S210 and the hypothesis (5, 飴, 2, {(2, 飴, 0), (4, 飴, 0)}) is inserted into the hypothesis list H ′.

However, a hypothesis (5, rain, 1, {(1, rain, 0), (3, rain, 0)}) already exists in the hypothesis list H ′, and the arrival state in the hypothesis generation rule T ₁ is 5. Therefore, the hypothesis (5, rain, 1, {(1, rain, 0), (3, rain, 0)}) with a small cumulative weight remains, and the hypothesis (5, 飴, 2, {(2, 飴, 0), (4, 飴, 0)}) are deleted, but the auxiliary hypothesis lists are linked and the hypotheses are (5, rain, 1, {(1, rain, 1), (3, rain, 1) , (2, 飴, 2), (4, 飴, 2)}).
Returning to step S206, since the state transition process list E ₁ = φ, the process proceeds to step S211. Since the hypothesis list H = φ, the process proceeds to step S212, and the hypotheses in the hypothesis list H ′ (5, rain, 1, {(1, rain, 1), (1, rain, 1), (2, 飴, 2) , (4, 飴, 2)}) is moved to the hypothesis list H, and since the next input symbol exists in step S213, the process returns to step S204.

[Processing of input symbol “da”]
Subsequently, in step S204, the symbol “da” is read, and in step S205, the hypothesis (5, rain, 1, {(1, rain, 1), (1, rain, 1), (2, hail, 2), (4, 2)}). State transition in which the input symbol is “da” from the current state 5 of the hypothesis generation rule T ₁ of this hypothesis <0 → 4: ball / 1>
Make a state transition process list E _1, including. Here, since the transition from the current state 5 of the hypothesis generation rule T ₁ to the state 0 can be made without an input symbol, the state transition of the hypothesis generation rule T ₁ included in E ₁ is <0 → 4: ball / 1>. It has become.

Because it is not a state transition process list _{E 1} = phi at step S206 proceeds to step S207, the state transition from the state transition process list _{E 1} <0 → 4, but: Ball / 1> was removed, a new hypothesis in step S208 f = (4, rainball, 2, {(1, rain, 1), (3, rain, 1), (2, 飴, 2), (4, 飴, 2)}) are generated, and the state is determined in step S209. Since the output symbol of transition <0 → 4: ball / 1> is not ε, hypothesis f = (4, rainball, 2, {(1, rain, 1), (3, rain, 1), ( 2, 飴, 2), (4, 飴, 2)}) is corrected to step S301.
In step S302, an empty auxiliary hypothesis list L ′ is generated.
In step S303, the cumulative weight AW [f] = 2−min (1, 1, 2, 2) = 1.

In step S304, the auxiliary hypothesis list L = {(1, rain, 1), (3, rain, 1), (2, 飴, 2), (4, 飴, 2)} is used as the auxiliary hypothesis (1, rain, 1). ) And a state transition that can be transitioned from the state 1 of the hypothesis generation rule T _{2 with} the input symbol “ball” <1 → 4, ball: ball / 5>
Make a state transition process list E _2, including.
Since E ₂ = φ is not satisfied in step S305, the process proceeds to step S306. State transition from the state transition process list _{E 2} <1 → 4, Ball: Ball / 5> was removed, the auxiliary hypothesis (4, candy, 7) at step S307 to generate. Here, the cumulative weight of the auxiliary hypothesis is W [g] + AW [f] + w [e ₂ ] = 1 + 1 + 5 = 7
It is calculated as follows. In step S308, the auxiliary hypothesis (4, raindrop, 7) is inserted into L ′.

Returning to step S305, since E ₂ = φ, the process proceeds to step S309.
Is not a supplementary hypothesis list L = phi proceeds to step S304, auxiliary hypotheses (3, rain, 1) was taken out, and the state transition does not exist possible transitions in the input symbol "ball" from the state 3 of hypothesis generation rule T ₂ State transition process list E ₂ = φ.
Returning to step S305, since E ₂ = φ, the process proceeds to step S309. Is not a supplementary hypothesis list L = phi proceeds to step S304, auxiliary hypothesis (2, candy, 2) is removed and transition possible state transitions in the input symbol "ball" from the state 2 of hypothesis generation rule T ₂ <2 → 4 , Jade: Jade / 1>
Make a list _{E 2,} including.

Since E ₂ = φ is not satisfied in step S305, the process proceeds to step S306. State transition from the state transition process list _{E 2} <2 → 4, Ball: Ball / 1> was removed, the auxiliary hypothesis (4, hard candy, 4) in step S307 to generate. Here, the cumulative weight of the auxiliary hypothesis is W [g] + AW [f] + w [e ₂ ] = 2 + 1 + 1 = 4
It is calculated as follows.
In step S308, the auxiliary hypothesis (4, jasper, 4) is inserted into the auxiliary hypothesis list L ′. At this time, there is an auxiliary hypothesis (4, rainball, 7) in the auxiliary hypothesis list L ′, so the auxiliary hypothesis (4, jasper, 4) with a small cumulative weight is left and (4, rainball, 7) is deleted. To do.

Returning to step S305, since E ₂ = φ, the process proceeds to step S309.
Is not a L = phi proceeds to step S404, the auxiliary hypothesis (4, candy, 2) is removed and hypothesis generation rule transitions possible state transitions from the state 4 in the input symbol "ball" of T ₂ are nonexistent and the state transition process Let list E ₂ = φ.
Returning to step S305, since E ₂ = φ, the process proceeds to step S309. Since the auxiliary hypothesis list L = φ, the process proceeds to step S310, and the element (4, jasper, 3) of the auxiliary hypothesis list L ′ is moved to L.

In step S311, the cumulative weight of the hypothesis is 4, and the correction of the hypothesis is ended in step S312, and as a result, the hypothesis is (4, rainball, 4, {(4, jasper, 4)}).
Returning to step S210, the hypothesis (4, rainball, 4, {(4, jasper, 4)}) is inserted into H ′.
Returning to step S206, since E ₁ = φ, the process proceeds to step S211. Since H = φ, the process proceeds to step S212, and the hypothesis (4, rainball, 4, {(4, jasper, 4)}) in H ′ is moved to H, and the next input symbol exists in step S213. The process returns to step S204.

[Processing of input symbol “MA”]
Subsequently, the symbol “ma” is read in step S204, and a hypothesis (4, rainball, 4, {(4, jasper, 4)}) is extracted from the hypothesis list H in step S205. State transition from the current state 1 of the hypothesis generation rule T ₁ of this hypothesis to the input symbol “MA” <4 → 5, MA: ε / 1>
Make a state transition process list E _1, including.
Since _{E 1} = not at φ at step S206 proceeds to step S207, the state transition from the state transition process list _{E 2} <4 → 5, or: epsilon / 1> was removed, a new hypothesis in step S208 f = (5, rain Ball, 5, {(4, jade ball, 4)}), and the process proceeds to step S209.
Since the output symbol of the state transition <4 → 5 or ε / 1> is ε, the process proceeds to the next step S210 and the hypothesis f = (5, rainball, 5, {(4, jasper, 4)}). Is inserted into the hypothesis list H ′.

Returning to step S206, since E ₁ = φ, the process proceeds to step S211. Since the hypothesis list H = φ, the process proceeds to step S212, and the hypothesis (5, rainball, 5, {(4, jasper, 4)}) in the hypothesis list H ′ is moved to the hypothesis list H. Since there is no input symbol, the process proceeds to step S214.
In step S214, the hypothesis h = (5, rainball, 5, {(4, jasper, 4)}) in the hypothesis list H is the only hypothesis that has reached the end state, and the auxiliary hypothesis list L [h] The cumulative weight W [g] = 4 + (5-1) + 0 = 5 is calculated for the auxiliary hypothesis g = (4, jasper, 4) included in the sub-hypothesis, and the output symbol string “Kotama” of the auxiliary hypothesis is converted. And the symbol string conversion process ends in step S215.
E. Roche and Y. Schabes, “Finite-State Language Processing”, MIT Press. 1997 Nagao Makoto "Iwanami Lecture Software 14; Knowledge and Reasoning", Iwanami Shoten, 109-113 E. Roche and Y. Schabes, “Finite-State Language Processing”, MIT Press, 1997, Chapter 15 “Speech Recognition by Composition of Weighted Finite Automata”

Conventionally, there has been a symbol string conversion method for converting a symbol string using two hypothesis generation units. However, in order to perform symbol string conversion that requires a complex conversion process using two hypothesis generation units, 2 is used. The conversion function required for one hypothesis generation unit is sophisticated, and a lot of manpower and cost are required for its creation.
The present invention intends to provide a symbol string conversion method and apparatus capable of obtaining a high-level conversion function by combining three or more hypothesis generation units. That is, the function of each individual hypothesis generation unit is simple, but an advanced conversion function is obtained by combining them in multiple stages.

  The symbol string conversion method according to the present invention is characterized by a symbol string input unit that sequentially reads a symbol string by a computer, and three or more M that convert a symbol string input by a hypothesis generation rule into hypotheses that are candidates for other symbol strings. A symbol string conversion unit including a plurality of hypothesis generation units, and a symbol string output unit that captures a conversion result output by the symbol string conversion unit. The symbol string conversion unit is provided for each of the M hypothesis generation units. A hypothesis generation process for generating a hypothesis composed of symbol strings obtained in a process in which a state transition can be performed with respect to an input symbol string to be input, and a cumulative weight value assigned to the generated hypothesis Hypothesis correction processing that corrects by the singular value among the cumulative weights assigned to the hypotheses generated by the generation unit, and the hypothesis narrowing down that eliminates unnecessary hypotheses using the corrected cumulative weights and narrows down to the high likelihood hypotheses Points to perform processing A.

Still another feature of the symbol string conversion method according to the present invention is the symbol string conversion method,
In the hypothesis correction process, the cumulative weight given to the hypothesis generated by each hypothesis generation unit is calculated from the cumulative weight given to each hypothesis generated in the hypothesis generation unit in the next stage or the corrected cumulative weight. This includes a hypothesis correction process that corrects using cumulative weights that exhibit singular values, and this hypothesis correction process is performed every time a symbol is input from the symbol string input unit to the first hypothesis generation unit. This is applied to all the cumulative weights of hypotheses generated by the first-stage hypothesis generator from the cumulative weights of hypotheses generated by the first section.

Still another feature of the symbol string conversion method according to the present invention is the symbol string conversion method according to any one of the above,
Hypothesis narrowing-down processing is a hypothesis that extracts hypotheses having cumulative weights that exhibit singular values among the corrected cumulative weights of hypotheses generated by the first-stage hypothesis generation unit when all input signal sequences are read. Including an extraction process and an input process for inputting the hypothesis extracted in this hypothesis extraction process as an input symbol string of the hypothesis generation part in the next stage, and causing each hypothesis generation part to execute the hypothesis extraction process and the input process. From the hypotheses generated by the first-stage hypothesis generation unit, a hypothesis specified by the corrected cumulative weight is selected from all hypotheses generated by the M-th hypothesis generation unit, and the M-th hypothesis generation unit The hypothesis extracted by the hypothesis extraction process from the hypotheses generated by is output as a symbol string conversion result.

  The present invention further proposes a speech recognition method using the above-described symbol string conversion method, and the feature includes an acoustic feature symbol string conversion processing step for extracting a symbol string representing the acoustic feature from an input speech signal, The symbol string representing the acoustic feature converted in the acoustic feature symbol string conversion processing step is converted into a symbol string corresponding to the input voice signal by any of the symbol string conversion methods described above.

  The present invention further proposes a speech paraphrasing method using a speech recognition method, the feature of which is to convert an input speech signal into an acoustic feature symbol string representing the acoustic feature, and from the symbol string representing the acoustic feature to the word A speech recognition method for converting into a sequence, a first hypothesis generation process that gives a smaller weight when a sequence of words obtained by the speech recognition method is plausible, and a sequence of words obtained by the speech recognition method. The second hypothesis generation process to be replaced with a sequence of words in another language or style, and the sequence of words in the other language or style give a smaller weight when it is plausible as a sentence in the other language or style. The speech signal is converted into a sentence in another language or style using at least three hypothesis generation processes including the third hypothesis generation process, and is output from the symbol string output unit.

  According to the present invention, since a symbol string can be converted using three or more hypothesis generation units, even if the conversion function of each individual hypothesis generation unit is simple, an advanced conversion function can be obtained as a whole. Incidentally, for example, a hypothesis generator having a conversion function for two languages can be created relatively simply. A first hypothesis generator having a function of converting Japanese to English, a second hypothesis generator having a function of converting English to French, a third hypothesis generator having a function of converting French to German, and the like. By applying the present invention in a situation where these hypothesis generation units exist, a symbol conversion device having a conversion function from Japanese to French is obtained by combining the first hypothesis generation unit and the second hypothesis generation unit. be able to. A major feature of the present invention is that a symbol conversion device having a function of converting from Japanese to German can be easily obtained by combining the first hypothesis generator, the second hypothesis generator, and the third hypothesis generator. is there.

  The symbol string conversion apparatus according to the present invention can be configured by hardware, but the embodiment in which the symbol string conversion program according to the present invention is installed in a computer and causes the computer to function as the symbol string conversion apparatus is the best implementation. It is a form.

  In order for the computer to function as the symbol string conversion device according to the present invention, the computer has a symbol string input unit for sequentially reading the symbol strings, and M stages for generating a hypothesis that becomes a candidate for another symbol string through the state transition process. Hypothesis generation unit and a hypothesis correction unit that corrects the weights assigned to each hypothesis generated by the M-stage hypothesis generation unit using the weights assigned to the hypotheses generated by the hypothesis generation unit on the subsequent stage side. A hypothesis narrowing unit that selects a hypothesis having a high likelihood using the weight corrected by the hypothesis correction unit and narrows down the hypothesis, a symbol string output unit that outputs the hypothesis narrowed down by the hypothesis narrowing unit as an output symbol string, To function as a symbol string converter.

  That is, the symbol input unit reads a symbol string to be converted into another symbol string one by one and passes it to the first-stage hypothesis generation unit. The first-stage hypothesis generator generates a symbol to be converted based on a combination that allows state transition by a weighted finite state conversion function (for example, a function that converts Japanese to English) given to the symbol each time it is input. A hypothesis that is a candidate symbol string of the column is generated. The number of hypotheses is not limited to one and is often generated, and each hypothesis is given a weight.

  All hypotheses generated by the first-stage hypothesis generator are input as input symbol strings to the second-stage hypothesis generator. The second-stage hypothesis generation unit generates a hypothesis that becomes a candidate symbol string of another symbol string by the weighted finite state conversion function given to the second-stage hypothesis generation unit. In the hypothesis generated by the second-stage hypothesis generation unit, the weight given by the first-stage hypothesis generation unit and the weight given by the second-stage hypothesis generation unit are accumulated. The procedure for delivering hypotheses generated in each of the possible state transition processes in the second-stage hypothesis generation section as an input symbol string of the third-stage hypothesis generation section is repeated until the M-th hypothesis generation section.

  The hypothesis correction unit calculates the cumulative weights assigned in the individual hypothesis generation processes of the M-1 stage hypothesis generation unit obtained by the procedure of delivering the hypothesis generation process to the M stage hypothesis generation unit. When the candidate symbol string generated in the hypothesis generation process is used as the input symbol string of the M-th hypothesis generation unit, the minimum or maximum cumulative weight is used in the hypothesis generation process possible in the M-th hypothesis generation unit. To correct. In other words, this correction selects the minimum or maximum cumulative weight given to the hypothesis generated by the M-th level hypothesis generation unit, and the selected weight is used as the cumulative weight of the hypothesis generated by the previous hypothesis generation unit. This is done by adding.

  A singular value that is minimum or maximum is selected from the weights after correction of the weights assigned to the hypothesis generated by the hypothesis generation unit in the M-1 stage, and the M-2 stage is used by using the selected weight. The accumulated weight of each hypothesis generated by the hypothesis generator is corrected. This correction procedure is repeated until the hypothesis generated by the first-stage hypothesis generation unit. The operations of the hypothesis generation unit and the hypothesis correction unit described above are executed every time one symbol is input to the first-stage hypothesis generation unit.

  The hypothesis narrowing-down unit reads the hypothesis having the minimum or maximum corrected cumulative weight among the hypotheses generated by the first-stage hypothesis generation unit at the time when all the input symbol strings are read. When the hypothesis is narrowed down as an input symbol string and the hypothesis is input to the second-stage hypothesis generator, the hypothesis with the smallest or maximum corrected cumulative weight among the hypotheses generated by the second-stage hypothesis generator Are narrowed down as input symbol strings of the hypothesis generation unit. This narrowing-down procedure is repeated up to the M-th level hypothesis generation unit, and a hypothesis having a minimum or maximum cumulative weight among the hypotheses generated by the M-th level hypothesis generation unit is output to the symbol string output unit. Let the sequence be the symbol string conversion result.

を計算し、このＷ（Ｘ）を与えるＯ_１，Ｏ_２，…，Ｏ_Ｍの組のうちＯ_Ｍを記号列変換結果とする。
本発明では、式（５）を次の式（６）

のように考える。 To summarize the above operation, when there are M hypothesis generation units T ₁ , T ₂ ,..., T _M (M is 3 or more) and an input symbol string X is given, T ₁ is used _first . the symbol string X into a symbol string O ₁ as a hypothesis Te, the symbol string O ₁ further converted to generate an output symbol sequence O ₂ as an input symbol string of hypothesis generation portion T ₂ of the second stage, further this was subject to conversion as 3 stage hypotheses input symbol string generating unit T _3, this is repeated until the M-th hypothesis generator means to obtain an output symbol string O _M ultimately T _M. To determine the final output symbol string O _M, as with the hypothesis generator is two, all the state transitions according to the M hypothesis generation portion when an input symbol string of the first hypothesis generating unit X From the combination of processes, a conversion result that minimizes the sum of the cumulative weights of the state transition processes in all hypothesis generation units is obtained. That means

Was _{calculated, O} 1, _O 2 to give the W (X), ..., and symbol string conversion result _{O M} of the set of _{O M.}
In the present invention, the equation (5) is replaced by the following equation (6):

Think like this.

An embodiment of the present invention will be described below with reference to the drawings.
FIG. 11 shows an embodiment of the present invention. The symbol string conversion unit 102 according to the present invention includes hypothesis generation units 201-1 to 201-M cascaded in M stages, and the M-th hypothesis generation unit 201-M from the M hypothesis generation units 201-2. Each of M hypothesis generation rule storage units 101-1 to 101-M, hypothesis correction units 301-1 to 301-M, and hypothesis correction units 301-1 to 301-M provided to each of the hypothesis generation rule storage units 101-1 to 101-M. Are formed by M hypothesis narrowing-down units 401-1 to 401 -M. Each hypothesis correction unit 301-1 to 301 -M has a hypothesis generated by the previous hypothesis generation unit with a hypothesis value that exhibits a singular value with the maximum or minimum cumulative weight among the hypotheses generated by the subsequent hypothesis generation unit. The weight correction means 301A for correcting the cumulative weight of No. 1 and the repetition control means 301B for repeatedly executing this correction operation from the (M−1) -th stage to the first-stage hypothesis correction unit. The hypothesis narrowing-down units 401-1 to 401-M have, for example, the smallest cumulative weight among the corrected hypotheses generated by the respective hypothesis generation units 201-1 to 201-M at the time when all the symbol strings are read. A hypothesis extraction unit 401A that executes a process of inputting the hypothesis to be input to the hypothesis generation unit of the next stage, and an input processing unit 401B that uses the hypothesis extracted by the hypothesis extraction unit 401A as an input of the next stage; Is provided.

One symbol is input from the symbol input unit 103 to the first hypothesis generation unit 201-1. Further, hypotheses narrowed down in accordance with the corrected cumulative weight are output to the symbol string output unit 106 in the final stage hypothesis narrowing-down unit 401 -M.
The operation of each part will be described in detail below.
The initial hypothesis generation unit 201-1 always generates a hypothesis based on a hypothesis generation rule given to itself for one symbol input. On the other hand, the hypotheses (candidate symbol strings) generated by the hypothesis generation means in the previous stage are input to the hypothesis generation units 201-2 to 201-M in the second and subsequent stages. Therefore, here, in order to distinguish the hypotheses output from the first-stage hypothesis generation unit 201-1 from the hypotheses output from the second-stage and later hypothesis generation units 201-2 to 201-M, the second-stage and later hypotheses are distinguished. The hypothesis generated by the generation units 201-2 to 201-M will be referred to as an m-th stage auxiliary hypothesis or simply an auxiliary hypothesis.

  The hypotheses generated by the respective hypothesis generation units 201-1 to 201-M are generated through the weighted finite state transition process as described in the background section, and a plurality of hypotheses are generated based on the possibility of the transition process. Is done. Furthermore, each hypothesis is given a weight according to the conditions attached to the weighted finite state transition process. The weight is accumulated according to the state transition from the first-stage hypothesis generation unit 201-1 to the final-stage hypothesis generation unit 201-M.

  When a hypothesis in which a non-empty content exists in the hypothesis generated by the final hypothesis generation unit 201-M starts to be generated, the hypothesis correction units 301-1 to 301-M start operating. When a hypothesis whose contents are included in the hypothesis generated by the final hypothesis generation unit 201-M starts to be generated, the hypothesis correction unit 301-M, for example, has a cumulative weight among the hypotheses generated by the hypothesis generation unit 201-M. The hypothesis that exhibits the minimum value is extracted, and the cumulative weight of the hypothesis generated by the hypothesis generation unit 201- (M-1) of the previous M-1 stage is corrected using the cumulative weight of the hypothesis. This correction is performed by adding the minimum cumulative weight extracted in the final stage to all the cumulative weights of the hypotheses generated by the hypothesis generation unit 201- (M-1) in the previous stage. This correction procedure is repeated until the hypothesis output by the first-stage hypothesis generation unit 201-1 is corrected.

  At the time when all the symbol strings are read, the hypothesis narrowing-down units 401-1 to 401-M have a high likelihood (likely) hypothesis among all the hypotheses generated by the respective hypothesis generation units 201-1 to 201-M. Select and refine the hypothesis. The narrowing-down operation is performed as follows. That is, the first-stage hypothesis narrowing-down unit 401-1 extracts the hypothesis having the smallest cumulative weight, for example, from the hypotheses generated by the first-stage hypothesis generation unit 201-1, and uses this hypothesis as the second-stage hypothesis generation unit 201-2. As an input symbol string, the second hypothesis generation unit 201-2 generates an auxiliary hypothesis in accordance with a finite state transition process possible with this input symbol string.

  The second-stage hypothesis correction unit 301-2 extracts a hypothesis having the smallest cumulative weight, for example, from the auxiliary hypotheses generated by the second-stage hypothesis generation unit 201-2, and uses this auxiliary hypothesis as the third-stage hypothesis generation unit 201. -3 is repeated until the auxiliary hypothesis output by the final-stage hypothesis generation unit 201-M. The auxiliary hypotheses extracted from the auxiliary hypotheses generated by the final-stage hypothesis narrowing-down section 401-M from the auxiliary hypothesis generation section 201-M are output to the symbol string output section 106, and the conversion of the symbol strings is completed.

12 and 13 show a specific procedure for converting a symbol string according to an embodiment of the present invention. First, in each hypothesis generation unit, one or more second stage auxiliary hypotheses always exist for one hypothesis, and one or more m + 1 stage auxiliary hypotheses always exist for the m stage auxiliary hypothesis. Therefore, in this embodiment, the set of the second stage auxiliary hypothesis for the hypothesis h is the second stage auxiliary hypothesis list L [h], and the set of the m + 1 stage auxiliary hypothesis for the m stage auxiliary hypothesis g _m is the mth stage auxiliary. This is represented by a hypothesis list L [g _m ].

Further, s _m [g _m] the m-th stage auxiliary hypothesis g _m when s ₁ [h] a state that has reached the end in the state transition process of hypothesis h, the m-th auxiliary hypothesis g _m is when there is a hypothesis h The state reached last in the state transition process of W, W [h] is the (corrected) cumulative weight in the state transition process of hypothesis h, and W [g _m ] is the state transition process of the m-th stage auxiliary hypothesis g _m ( Corrected cumulative weight. Further, O [g _m ] is a symbol string output in the state transition process of the m-th stage auxiliary hypothesis g _m .

The procedure shown in FIGS. 12 and 13 will be described below in comparison with the embodiment shown in FIG.
The basic procedure is the same as that shown in FIGS. 8 and 9, but the processes in steps S203 and S403, steps S209 and S409, and steps S214 and S414 are different. Starting from step S401, as an initial setting, empty hypothesis lists H and H ′ are generated in step S402. In step S403, an initial hypothesis h is generated, and s ₁ [h] = 0, W [h] = 0, and L [h] = φ (empty list) are inserted into the hypothesis list H. In addition, an initial m-th stage auxiliary hypothesis g _m is generated for m = 2 to M, and the last state s _m [g _m ] = 0, W [g _m ] = 0, O [g _m ] = φ and (empty symbol string), a second stage auxiliary hypothesis _{g 2} is inserted into the second stage auxiliary hypothesis list L [h], m = m -1 stage and _{g m} against 3~M auxiliary hypothesis list L Insert into [g _m-1 ].

In step S404, the symbol string input unit 103 reads one symbol and substitutes the symbol for x.
The following steps S405 to S408 are executed in the hypothesis generation unit 201.
In step S405, one hypothesis is extracted from the hypothesis list H and substituted into h, and a state transition list E ₁ that can be transitioned with the input symbol x is generated from the state s ₁ [h].
In step S406, if the list E ₁ = φ (empty list), the process proceeds to step S411. Otherwise, the process proceeds to step S407.

In step S407, fetches one state transition from the list _{E 1,} is substituted into _{e 1,} generates a new hypotheses _{f 1} in the step _{S408, s 1 [f 1]} = n [e 1], W [f ₁ ] = W [h] + w [e ₁ ] and L [f ₁ ] = L [h].
Step S409 is a means for correcting the output symbol string o [e ₁ ] ≠ ε (if not empty) hypothesis f ₁ in hypothesis correction units 301-1 to 301-M (from step S501 in FIGS. 14 and 15). Step S512, which will be described separately after the description of this procedure) is executed.
Step S410, in the hypothesis refining unit 401-1 to 401-M, refine the hypothesis by inserting the hypothesis _{f 1} to the hypothesis list H '.

In step S411, if H = φ (all hypotheses have been expanded), the process proceeds to step S412. Otherwise, the process returns to step S405 to develop the next hypothesis.
In step S412, all the elements of the newly generated hypothesis list H ′ are moved to the already empty hypothesis list H, and the process proceeds to step S413.
In step S413, if there is a next input symbol in the symbol string input unit 103, the process returns to step S404. Otherwise, it is determined that all the input symbol strings have been read, and the process proceeds to step S414.

をすべてのｈ_ｍについて求め、ｓ_Ｍ[ｈ_Ｍ]∈Ｆ_ＭかつＷ[ｈ_Ｍ]が最小のｈ_Ｍを選び、Ｏ[ｈ_Ｍ]を変換結果として記号列出力部１０６において出力する。 In step S414, the hypothesis (s ₁ [h ₁ ] εF ₁ ) that has reached the end state from the hypothesis list H and the cumulative weight (W [h ₁ ]) of the state s ₁ [h ₁ ] Add end weight ρ (s ₁ [h ₁ ]), and for m = 2 to M, when s _m [h _m ] ∈F _m and h _m ∈L [h _m−1 ]

Asking for all _{h m,} select _{s M [h M]} ∈F _M and W _{[h M]} is smallest _{h M,} and outputs the symbol sequence output unit 106 O _{[h M]} as a conversion result.

In step S415, the symbol string conversion procedure in one embodiment of the present invention is terminated.
A hypothesis correction procedure in the hypothesis correction units 301-1 to 301 -M will be described.
In the present invention, the hypothesis correction is performed when the cumulative weight for the state transition process of the hypothesis in the m-th stage hypothesis generation unit is set as the input symbol string of the m + 1-stage hypothesis generation unit as the hypothesis in the state transition process of the hypothesis. This means that a state transition process having a minimum cumulative weight among the possible state transition processes in the m + 1 stage hypothesis generation unit is obtained and the cumulative weight is added to the hypothetical cumulative weight.

となり、その補正累積重みが最小である仮説の補正累積重みは

となり、先に述べた式（４）でＭ＝３の場合における仮説生成部の累積重みの和の最小値と一致する。従って、補正された累積重みを用いて、１段目の仮説生成部の出力記号列Ｙを求めて、そのＹを２段目の仮説生成部の入力記号列としたときの可能な状態遷移過程の中で累積重みが最小となる出力記号列Ｚを求めて、そのＺを３段目の仮説生成部の入力記号列としたときの可能な状態遷移過程の中で累積重みが最小となる状態遷移過程に対応する出力記号列が記号列変換結果となる。 For example, when M = 3, the corrected cumulative weight of the hypothesis that accepts the symbol string X and outputs the symbol string Y when the input symbol string X is read is:

The corrected cumulative weight of the hypothesis with the minimum corrected cumulative weight is

Thus, the value coincides with the minimum value of the sum of the cumulative weights of the hypothesis generation unit in the case of M = 3 in Equation (4) described above. Accordingly, a possible state transition process when the output symbol string Y of the first-stage hypothesis generation unit is obtained using the corrected cumulative weight and the Y is used as the input symbol string of the second-stage hypothesis generation unit. The output symbol string Z having the smallest cumulative weight is obtained, and the state in which the cumulative weight is the smallest in the possible state transition process when Z is used as the input symbol string of the third-stage hypothesis generator The output symbol string corresponding to the transition process becomes the symbol string conversion result.

In the present invention, correction of hypotheses is efficiently performed using an m-th stage auxiliary hypothesis list associated with each hypothesis. The hypothesis to be corrected is f _m−1 (m−1 stage auxiliary hypothesis), and the m−1 stage auxiliary hypothesis list including the m stage auxiliary hypothesis derived from the input symbol string of f _m−1 is L [f. _m−1 ].
From the state s _m [g _m ] of the _m-th hypothesis generation unit (g _m represents one m-th auxiliary hypothesis included in L [f _m-1 ]), the input symbol o [e _m-1 ]. via the state transition _{e m} by generating a m-th stage auxiliary hypothesis _{f m} to reach the next state n _{[e m],} and its arrival state _{s m [f m] = n} [e m]. Also, in the state transition process of the m−1 stage auxiliary hypothesis f _m−1 , the latest state transition before the last state transition (e _m−1 ) of f _{m−1 and} whose output symbol is not ε. Is t _m−1 . In this case, the cumulative weight W for m-th auxiliary hypothesis f _m [f _m] can be calculated by adding the following three elements.

(A1) Cumulative weight of the m-th stage auxiliary hypothesis f _m generated in the state n [t _m-1 ] in the state transition process of the m-1 stage auxiliary hypothesis f _m-1 :
This value is equal to W _{[f m].} That is, in the state transition process in which the output symbol is ε, the auxiliary hypothesis list is directly inherited by the hypothesis of the subsequent state transition process, and therefore the m−1 stage auxiliary hypothesis generated in the state n [t _m−1 ]. The list itself is in the current L [f _m-1 ] and is equal to the cumulative weight W [f _m ] of each _m-th stage auxiliary hypothesis in the m-1 stage auxiliary hypothesis list L [f _m-1 ].

に等しく、これを現在のｍ−１段目補助仮説ｆ_ｍ−１の累積重みＷ[ｆ_ｍ−１]から差し引くことにより、状態ｎ[ｔ_ｍ−１]から状態ｓ_ｍ−１[ｆ_ｍ−１]までの間の累積重みが求まる。すなわち、

である。 (A2) Weights accumulated from state n [t _m-1 ] to state s _m-1 [f _m-1 ]:
However, in the state transition process in which the output symbol is ε, the m-th stage auxiliary hypothesis list is directly inherited by the newly generated hypothesis, so that the corrected accumulation of the hypothesis that has reached the state n [t _m−1 ] is performed. The weight is the smallest cumulative weight in the current m-1 stage auxiliary hypothesis list L [f _m-1 ].

Equally, by subtracting this from the current m-1 stage auxiliary hypothesis _{f m-1} of the cumulative weight W _{[f m-1],} from one state _{n [t m-1] s} m-1 [f m _-1 ] until the cumulative weight is obtained. That is,

It is.

を代入することにより行われる。 (A3) the weight in the state transition _{e m} of m from a state _{s m} th stage of hypothesis generation unit _{[g m]} to _{s m [f m]: w} [e m]
Therefore, the cumulative weight of the m-th stage auxiliary hypothesis _{f m} is calculated as _{W [f m] = W [} g m] + AW [f m-1] + w [e m]. However, when m <M, W [f _m ] is further corrected using the hypothesis generation unit in the (m + 1) th stage. After that, it is inserted into the m−1 stage auxiliary hypothesis list L [f _m−1 ].
Then, the m-1 stage auxiliary hypothesis _{f m-1} correction, W of _{[f m-1]} is the newly generated the m-th auxiliary hypothesis g _{(however, g∈L [f m-1]} ) Minimum value in cumulative weight W [g]

This is done by substituting

The hypothesis correction procedure will be described below with reference to FIGS. This hypothesis correction procedure is executed in step S409 shown in FIGS. Assume that the hypothesis or auxiliary hypothesis corrected in this procedure is f _m−1 , and the output symbol from the m−1 stage hypothesis generation unit is o [e _m−1 ].
In step S501, a procedure for correcting the hypothesis or auxiliary hypothesis f _m−1 is started.
In step S502, an empty auxiliary hypothesis list _Lm is generated.
In step S503, the weight correction term AW [fm _-1 ] is obtained as AW [fm _-1 ] = W [fm _-1 ] -W [g ']. However, g 'is the cumulative weight in f _m-1 of the auxiliary hypothesis list L [f _m-1] refers to the minimum of the auxiliary hypothesis.

In step S504, one auxiliary hypothesis is extracted from L [f _m−1 ] and substituted into g _m . A list E _{m of} state transitions that can be transitioned from the state s _m [g _m ] with the input symbol o [e _m−1 ] is generated, and the process proceeds to step S405.
In step S505, if E _m is empty (auxiliary hypotheses have been expanded for all state transitions from s _m [g _m ]), the process proceeds to step S510. Otherwise, the process proceeds to step S506.
In step S506, extract a single state transition from the list _{E m,} it is substituted into _{e m.}
At step S507, the generated auxiliary hypothesis _{f m, s m [f m} ] = n [e m], W [f m] = W [g m] + AW [f m-1] + w [e m], O _{[f m]} = and _{O [g m] · o [} e m].

In step S508, o _{[e m]} is not the epsilon, and if m <M, to correct the _{f m.} Here, m is incremented by one, and the same procedure is performed by executing steps S501 to S513. After the correction, m is decremented by one and the process proceeds to the next step S509.
In step S509, the auxiliary hypothesis f _m is inserted into the auxiliary hypothesis list L _m , and the process proceeds to step S505.
In step S510, if L [f _m-1 ] is φ (all auxiliary hypotheses have been expanded), the process proceeds to step S511. Otherwise, the process returns to step S504 to develop the next auxiliary hypothesis.
In step S511, new elements of the list L _m of generated auxiliary hypothesis, already transferred all vacant and became auxiliary hypothesis list L [f _m-1], the process proceeds to step S512.

In step S512, W [f _m−1 ] = W [g ″] is set for the auxiliary hypothesis g ″ having the smallest cumulative weight in the auxiliary hypothesis list L [f _m−1 ].
In step S513, the hypothesis correction procedure in the embodiment of the present invention is terminated.
The symbol string conversion procedure (FIGS. 12 and 13) according to the present invention is almost the same calculation procedure as the procedure (FIG. 4) for searching using only the first-stage hypothesis generator except for the hypothesis correction step S409. Also in the hypothesis correction procedure (FIGS. 14 and 15), except for step S508, the procedure is almost the same as the conventional hypothesis correction procedure (FIG. 10), and the amount of calculation can be reduced. Since the hypothesis correction procedure is executed only when the output symbol is not ε in the state transition of the hypothesis generation unit, the larger the ratio that the output symbol of the hypothesis generation unit is ε, the greater the effect of reducing the processing amount.

As an example, consider the conversion (in the case of M = 3) based on the three hypothesis generation rules shown in FIG. 16 and FIGS.
FIG. 16 shows a hypothesis generation rule for converting a Roman character string into a kana. Consider a case where a symbol string input in Roman characters is converted into a corresponding Kanji character string by these three hypothesis generation rules.

Next, according to the symbol string conversion procedure of the present invention, the input symbol string “a, m, e, FIG. 16 is assumed as the hypothesis generation rule T ₁ , FIG. 5 as the hypothesis generation rule T ₂ , and FIG. 6 as the hypothesis generation rule T ₃ . A process for obtaining an output symbol string when d, a, m, a ″ is given will be described step by step. However, the hypothesis (hypothesis generation rule current state number _{s 1} of _{T 1,} the output symbol sequence O, cumulative weight W, auxiliary hypothesis list L) is expressed as _(s 1, O, W, L), m-th auxiliary hypothesis (current state number _{s m} of hypothesis generation rule _{T m,} the output symbol string O, cumulative weight W, m-th auxiliary hypothesis list _{L m)} is expressed as _{(s m, O, W,} L m). The m-th stage auxiliary hypothesis list is represented as {(s _m , O, W, L _m ), (s _m ′, O ′, W ′, L _m ′),.

[Start the symbol string conversion procedure according to the present invention]
Starting from step S401, empty hypothesis lists H and H ′ are created in step S402.
In step S403, a hypothesis h = (0, φ, 0, {(0, φ, 0, {(0, φ, 0, φ)})}) is inserted into the hypothesis list H.

[Processing of input symbol “a”]
In step S404, the symbol “a” is read, and in step S405, a hypothesis (0, φ, 0, {(0, φ, 0, {(0, φ, 0, φ)})}) is extracted from the hypothesis list H. State transition of input symbol “a” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <0 → 0, a: a / 0>
Make a list _{E 1,} including.
Because it is not in the list _{E 1} = _φ in step S406 proceeds to the step S407, the state transition from the list _{E 1 <0 → 0, a} : Oh / 0> and list _{e 1} taken out, a new hypothesis in step S408 _f 1 = ( 0, a, 0, {(0, φ, 0, {(0, φ, 0, φ)})}). In step S409, since o [e ₁ ] is not ε, the hypothesis f ₁ = (0, a, 0, {(0, φ, 0, {(0, φ, 0, φ)})}) is corrected. The process proceeds to step S501. At this time, m = 2 is set.

To generate an empty auxiliary hypothesis list _{L 2} in step S502.
In step 503, AW [f ₁ ] = 0.
In step S504, from the auxiliary hypothesis list L [f ₁ ] = {(0, φ, 0, {(0, φ, 0, φ)})}, the auxiliary hypothesis g ₂ = (0, φ, 0, {(0, φ, 0, φ)}), and the state transition <0 → 1, a: rain / 0>, which can be transitioned from the state 0 of the hypothesis generation rule T _{2 with} the input symbol “a”
<0 → 2, A: 飴 / 0>
Make a list _{E 2,} including.
Since E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from the list _{E 2} <0 → 1, Ah: Rain / 0> and List _{e 2} removed, auxiliary hypothesis _f 2 = (1 in step S507, the rain, 0, {(0, φ , 0, φ)} ) Is generated.

In o _{[e 2]} rather than ε step S508, the and, since at m = 2 is M = 3 Since m <M, the process proceeds to step S501 to correct the auxiliary hypothesis _{f 2.} At this time, m is increased by 1, and m = 3 is set.
To generate an empty auxiliary hypothesis list _{L 3} in step S502.
In step S503, AW [f ₂ ] = 0.
In step S504, the auxiliary hypothesis g ₃ = (0, φ, 0, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, φ, 0, φ)}, and from the state 0 of the hypothesis generation rule T _3. State transitions that can be transitioned with the input symbol “rain” <0 → 1, rain: rain / 0>,
<0 → 3, rain: rain / 0>
Make a list _{E 3,} including.

Since E ₃ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E ₃ <0 → 1, rain: Rain / 0> and _{e 3} removed, auxiliary hypothesis _f 3 = in Step S507 (1, rain, 0, phi) to produce a.
Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting auxiliary hypothesis _{f 3} Listing _{L 3.}
Returning to step S505, since E ₃ = φ is not satisfied, the process proceeds to step S506. State transition from E ₃ <0 → 3, rain: Rain / 0> and _{e 3} removed, auxiliary in step S507 hypothesis _f 3 = (3, rain, 0, phi) to produce a.
Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting the hypothesis _{f 3} Listing _{L 3.}
Returning to step S505, since the list E ₃ = φ, the process proceeds to step S510.

In step S510, since auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511.
In step S511, the elements of the auxiliary hypothesis list _{L 3} {(1, rain, 0, φ), (3 , rain, 0, φ)} transferred to all the auxiliary hypothesis list L _{[f 2].}
In step S512, since the cumulative weights of the auxiliary hypotheses of L [f ₂ ] are both 0, W [f ₂ ] = 0 is set.
Step S513 In Exit correction auxiliary hypothesis _{f 2,} and m = 2 in the m reduced by one, the flow returns to step S508.
After the correction procedure auxiliary hypothesis _{f 2} at step S508, the process proceeds to step S509.
At step S509, the inserted auxiliary hypothesis _{f 2} to the auxiliary hypothesis list _{L 2,} the flow returns to step S505.

Since the list E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition <0 → 2, a: 飴 / 0> is extracted from list E ₂ and set as e _2, and auxiliary hypothesis f ₂ = (2, 飴, 0, {(0, φ, 0, φ)}) in step S507. Is generated.
In o _{[e 2]} rather than ε step S508, the and, since at m = 2 is M = 3 Since m <M, the process proceeds to step S501 to correct the auxiliary hypothesis _{f 2.} At this time, m is increased by 1, and m = 3 is set.
To generate an empty auxiliary hypothesis list _{L 3} in step S502.
In step S503, AW [f ₂ ] = 0.

In step S504, an auxiliary hypothesis g ₃ = (0, φ, 0, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(0, φ, 0, φ)}, and from the state 0 of the hypothesis generation rule T _3. Possible state transitions with input symbol “飴” <0 → 2, ＜: 飴 / 0>,
<0 → 4, 飴: 飴 / 0>
Make a list _{E 3,} including.
Since list E ₃ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from the list _{E 3} <0 → 2, candy: candy / 0> and List _{e 3} removed, auxiliary hypothesis _f 3 = in Step S507 (2, candy, 0, phi) to produce a.

Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting auxiliary hypothesis _{f 3} to the auxiliary hypothesis list _{L 3.}
Returning to step S505, since E ₃ = φ is not satisfied, the process proceeds to step S506. State transition from E ₃ <0 → 4, candy: candy / 0> and _{e 3} removed, auxiliary in step S507 hypothesis _f 3 = (4, candy, 0, phi) to produce a.
Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting auxiliary hypothesis _{f 3} to the auxiliary hypothesis list _{L 3.}
Returning to step S505, since E ₃ = φ, the process proceeds to step S510.
In step S510, since auxiliary hypothesis list L [f ₂ ] = φ, the process proceeds to step S511.

In step S511, elements of the list _{L 3} {(2, candy, 0, φ), (4 , candy, 0, φ)} transferred to all the auxiliary hypothesis list L _{[f 2].}
In step S512, since the cumulative weights of the auxiliary hypotheses of L [f ₂ ] are both 0, W [f ₂ ] = 0 is set.
Step S513 In Exit correction auxiliary hypothesis _{f 2,} and m = 2 in the m reduced by one, the flow returns to step S508.
After the correction procedure _{f 2} at step S508, the process proceeds to step S509.
At step S509, the inserting the _{f 2} to _{L 2,} the flow returns to step S505.
Since E ₂ = φ in step S505, the process proceeds to step S510.

In step S510, since L [f ₁ ] = φ, the process proceeds to step S511.
In step S511, transferred all the elements of the _{L 2} to L _{[f 1].}
In step S512, L [f ₁ ] = {(1, rain, 0, {(1, rain, 0, φ), (3, rain, 0, φ)}), (2, 飴, 0, {( 2, 重 み, 0, φ), (4, 飴, 0, φ)})} are both 0, so W [f ₁ ] = 0.
Exit correction auxiliary hypothesis _{f 1} at step S513, the flow returns to step S409.
After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410.

In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406.
Since the state transition list E ₁ = φ in step S406, the process proceeds to step S411.
Since the hypothesis list H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(0, ah, 0, {(1, rain, 0, {(1, rain, 0, φ), (3, rain) , 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})})}
Is moved to the hypothesis list H.
Proceeding to step S413, since the next input symbol exists, the process returns to step S404.

[Processing of input symbol “m”]
In step S404, the symbol “m” is read, and in step S405, the hypothesis h = {(0, ah, 0, {(1, rain, 0, {(1, rain, 0, φ), (3, Rain, 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})})}
Take out. State transition of input symbol “m” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <0 → 1, m: ε / 0>
Make a list _{E 1,} including.

Since _{E 1} = not at φ at step S406 proceeds to step S407, the state transition list state transition from _{E 1 <0 → 0, m} : ε / 0> and _{e 1} removed, new hypotheses in step S408 f ₁ = ( 1, a, 0, {(1, rain, 0, {(1, rain, 0, φ), (3, rain, 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})})
Is generated. Since o [e ₁ ] = ε in step S409, the process proceeds to step S410.
In step S410 inserts the hypothesis _{f 1} to the hypothesis list H ', processing proceeds to step S406.

In step S406, since the state transition list E ₁ = φ, the process proceeds to step S411, and since the hypothesis list H = φ, the process proceeds to step S412.
In step S412, the elements of the hypothesis list H ′ {(1, A, 0, {(1, Rain, 0, {(1, Rain, 0, φ), (3, Rain, 0, φ)}), (2 , 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})})}
Are all moved to the hypothesis list H.
Proceeding to step S413, since the next input symbol exists, the process returns to step S404.

[Processing of input symbol “e”]
In step S404, the symbol “e” is read, and in step S405, a hypothesis h = (1, ah, 0, {(1, rain, 0, {(1, rain, 0, φ), (3, rain) , 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})})
Take out. State transition of input symbol “e” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <1 → 0, e: me / 0>
Make a state transition list E _1, including.

Since _{E 1} = not a φ in step S406 proceeds to the step S407, the state transition from _{E 1:} the _{e 1} Remove the <1 → 0, e because / 0>, a new hypothesis in step S408 f ₁ = (0, rain , 0, {(1, rain, 0, {(1, rain, 0, φ), (3, rain, 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ ), (4, 飴, 0, φ)})})
Is generated. Since step S409 o _{[e 1]} is not a ε, the process proceeds to step S501 to correct the hypothesis _{f 1.} At this time, m = 2 is set.
To generate an empty auxiliary hypothesis list _{L 2} in step S502.

In step S503, AW [f ₁ ] = 0.
In step S504, the auxiliary hypothesis list L [f ₁ ] = {(1, rain, 0, {(1, rain, 0, φ), (3, rain, 0, φ)}), (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})}
From the auxiliary hypothesis g ₂ = (1, rain, 0, {(1, rain, 0, φ), (3, rain, 0, φ)})
State transition that can be transitioned from the state 1 of the hypothesis generation rule T _{2 by} the input symbol “M” <1 → 5, M: ε / 1>
Make a state transition list E _2, including.

Since the state transition E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition list _{E 2} state transition from <1 → 5, Oh: epsilon / 1> and _{e 2} removed, auxiliary hypothesis _f 2 = (5, rain in step S507, 1, {(1, rain, 0, phi) , (3, rain, 0, φ)}).
Since o [e ₂ ] is ε in step S508, the process proceeds to step S509.
In step S509 the _{f 2} is inserted into _{L 2,} the flow returns to step S505.
Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, L [f ₁ ] is not φ, and the process returns to step S504.

In step S504, auxiliary hypothesis list L [f ₁ ] = {(2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})}
From the auxiliary hypothesis g ₂ = (2, 飴, 0, {(2, 飴, 0, φ), (4, 飴, 0, φ)})
State transition that can be transitioned from the state 1 of the hypothesis generation rule T _{2 with} the input symbol “M” <2 → 5, M: ε / 2>
Make a state transition list E _2, including.

Since E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E ₂ <2 → 5, because: epsilon / 2> and _{e 2} removed, auxiliary hypothesis _f 2 = (5 in step S507, the candy, 2, {(2, candy, 0, phi), (4 , 飴, 0, φ)}).
Since o [e ₂ ] is ε in step S508, the process proceeds to step S509.
In step S509 the _{f 2} is inserted into _{L 2,} the flow returns to step S505.
However, an auxiliary hypothesis (5, rain, 1, {(1, rain, 0, φ), (3, rain, 0, φ)}) already exists in the auxiliary hypothesis list L ₂ , and the hypothesis generation rule T ₂ Since the arrival state at 5 is 5, an auxiliary hypothesis with a small cumulative weight (5, rain, 1, {(1, rain, 0, φ), (3, rain, 0, φ)}) remains, and (5 , 飴, 2, {(2, 飴, 0, φ), (4, 飴, 0, φ)}) are deleted. At this time, the auxiliary hypothesis list of each auxiliary hypothesis is connected by correcting the cumulative weight, and (5, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), ( 2, 飴, 2, φ), (4, 飴, 2, φ)}). In addition, auxiliary hypothesis that has reached the state 5, state transition <5 → 0, ε: ε / 0> through, because to reach the state 0 with no input symbol, the auxiliary hypothesis list L ₂ auxiliary hypothesis,
(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, hail, 2, φ), (4, hail, 2, φ)})
Remains.

Since E ₂ = φ in step S505, the process proceeds to step S510. In step S510, L [f ₁ ] is φ, and therefore, the process proceeds to step S511.
Step S511 In all elements of the _{L 2} transferred to L _{[f 1].}
In step S512, L [f ₁ ] = {(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, hail, 2, φ), ( In the auxiliary hypothesis of 4, １, 2, φ)})}, the minimum value of the cumulative weight is 1, so W [f ₁ ] = 1.
Exit correction auxiliary hypothesis _{f 1} at step S513, the flow returns to step S409.

After correction procedure hypothesis _{f 1} at step S409, the process proceeds to step S410.
In step S410, insert the hypothesis _{f 1} to the hypothesis list H ', the flow returns to step S406.
Since the hypothesis transition list E ₁ = φ in step S406, the process proceeds to step S411.
Since H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(0, rain, 1, {(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1 , Φ), (0, 飴, 2, {(2, 飴, 2, φ), (4, 飴, 2, φ)})}
Is moved to the hypothesis list H.

Proceeding to step S413, since the next input symbol exists, the process returns to step S404.
[Processing of input symbol “d”]
In step S404, the symbol “d” is read, and in step S405, the hypothesis h = {(0, rain, 1, {(0, rain, 1, {(1, rain, 1, φ), (3, Rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)})}
Take out. State transition of input symbol “d” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <0 → 3, d: ε / 0>
Make a hypothesis transition list E _1, including.

Since _{E 1} = not at φ at step S406 proceeds to step S407, the state transition from _{E 1:} a _{e 1} removed <0 → 3, d ε / 0>, a new hypothesis in step S408 f ₁ = (3, rain , 1, {(0, rain, 1, {(1, rain, 0, φ), (3, rain, 0, φ)}), (0, 飴, 2, {(2, 飴, 0, φ ), (4, 飴, 0, φ)})})
Is generated. Since o [e ₁ ] = ε in step S409, the process proceeds to step S410.
In step S410 the _{f 1} is inserted into H ', processing proceeds to step S406.
In step S406, since E ₁ = φ, the process proceeds to step S411, and since H = φ, the process proceeds to step S412.

In step S412, the elements of the hypothesis list H ′ {(3, candy, 1, {(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)})})}
Are all moved to the hypothesis list H.
Proceeding to step S413, since the next input symbol exists, the process returns to step S404.
[Processing of input symbol “a”]
In step S404, the symbol “a” is read, and in step S405, the hypothesis h = (3, candy, 1, {(0, rain, 1, {(1, rain, 0, φ), (3, rain) , 0, φ), (2, 飴, 0, φ), (4, 飴, 0, φ)})})
Take out. State transition of input symbol “e” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <3 → 0, a: da / 0>
Make a list _{E 1,} including.

Since _{E 1} = not a φ in step S406 proceeds to the step S407, the state transition from _{E 1:} the _{e 1} Remove the <1 → 0, e because / 0>, a new hypothesis in step S408 f ₁ = (0, rain 1, {(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)})})
Is generated. Since step S409 o _{[e 1]} is not a ε, the process proceeds to step S501 to correct the hypothesis _{f 1.} At this time, m = 2 is set.
To generate an empty auxiliary hypothesis list _{L 2} in step S502.

In step S503, AW [f ₁ ] = 0.
In step S504, auxiliary hypothesis list L [f ₁ ] = {(0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, 飴, 2, φ) , (4, 飴, 2, φ)})}
From the auxiliary hypothesis g ₂ = (0, rain, 1, {(1, rain, 1, φ), (3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2 , Φ)})
State transition that can be transitioned from the state 0 of the hypothesis generation rule T _{2 with} the input symbol “da” <0 → 4, ball: ball / 1>
Make a list _{E 2,} including.

Since E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E _2: and _{e 2} removed <0 → 4, but the ball / 1>, auxiliary hypothesis _f 2 = (4 in step S507, the rain ball, 2, {(1, rain, 1, phi), ( 3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)}).
In o _{[e 2]} rather than ε step S508, the and, since at m = 2 is M = 3 Since m <M, the process proceeds to step S501 of correcting the _{f 2.} At this time, m is increased by 1, and m = 3 is set.

To generate an empty auxiliary hypothesis list _{L 3} in step S502.
In step S503, AW [f ₂ ] = 2−1 = 1.
In step S504, auxiliary hypothesis list L [f ₂ ] = {(1, rain, 1, φ), (3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)} is extracted from auxiliary hypothesis g ₃ = (1, rain, 1, φ), and state transition that can be transitioned from input 1 “ball” from state 1 of hypothesis generation rule T ₃ <1 → 4, ball: ball / 5>
Make a list _{E 3,} including.

Since E ₃ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E ₃ <1 → 4, Ball: Ball / 5> and _{e 3} removed, auxiliary hypothesis _f 3 = in Step S507 (4, candy, 7, phi) to produce a. Here, it is calculated as W [f ₃ ] = W [g ₃ ] + AW [f ₂ ] + w [e ₃ ] = 1 + 1 + 5 = 7.
Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting the _{f 3} to _{L 3.}
Returning to step S505, since E ₃ = φ, the process proceeds to step S510.

In step S510, since L [f ₂ ] is not φ, the process returns to step S504.
In step S504, the auxiliary hypothesis g ₃ = (3, rain) from L [f ₂ ] = {(3, rain, 1, φ), (2, 飴, 2, φ), (4, 飴, 2, φ)}. , 1, φ), and since there is no state transition that can be transitioned from the state ₃ of the hypothesis generation rule T _{3 with} the input symbol “ball”, E ₃ = φ.
The process proceeds to step S505, because the _{E 3} is φ, the process proceeds to step S510.
In step S510, since L [f ₂ ] is not φ, the process returns to step S504.
In step S504, the auxiliary hypothesis g ₃ = (2, 飴, 2, φ) is extracted from the auxiliary hypothesis list L [f ₂ ] = {(2, 飴, 2, φ), (4, 飴, 2, φ)}. State transition that can be transitioned from the state 2 of the hypothesis generation rule T _{3 with} the input symbol “ball” <2 → 4, ball: ball / 1>
Make a list _{E 3,} including.

Since E ₃ = φ is not satisfied in step S505, the process proceeds to step S506. State transition from E ₃ <2 → 4, Ball: Ball / 1> and _{e 3} removed, auxiliary in step S507 hypothesis _f 3 = (4, hard candy, 4, phi) to produce a. Here, it is calculated as W [f ₃ ] = W [g ₃ ] + AW [f ₂ ] + w [e ₃ ] = 2 + 1 + 1 = 4.
Not a m = 3 and M = 3 Since m <M at step S508, the process proceeds to step S509, the inserting the _{f 3} to _{L 3.}
Here, the auxiliary hypothesis has already reached the same state 4 of hypothesis generation rule T ₃ to L ₃ (4, candy, 7, phi) because there, small cumulative weight (4, hard candy, 4 , Φ) only, and delete (4, rainball, 7, φ).

Returning to step S505, since E ₃ = φ, the process proceeds to step S510.
In step S510, since L [f ₂ ] is not φ, the process returns to step S504.
In step S504, the auxiliary hypothesis g ₃ = (4, 飴, 2, φ) is extracted from L [f ₂ ] = {(4, 飴, 2, φ)}, and the input symbol “ ₃ ” is input from state 4 of the hypothesis generation rule T _3. Since there is no state transition that can be transitioned by the “ball”, E ₃ = φ.
The process proceeds to step S505, because the _{E 3} is φ, the process proceeds to step S510.
In step S510, since L [f ₂ ] is φ, the process proceeds to step S511.

Step S511 In all the elements of _{L 3} were transferred to L _{[f 2],} the process proceeds to step S512.
In step S512, the minimum cumulative weight among L [f ₂ ] = {(4, jade ball, 4, φ)} is 4, so W [f ₂ ] = 4.
Step S513 In Exit correction auxiliary hypothesis _{f 2,} and m = 2 in the m reduced by one, the flow returns to step S508.
After correction of _{f 2} at step S508, the process proceeds to step S509.
At step S509, the inserting the _{f 2} to _{L 2,} the process proceeds to step S505.

In step S505 _{E 2} is so φ, the process proceeds to step S510.
Since L [f ₂ ] is φ in step S510, the process proceeds to step S511.
Step S511 In _{L 2} elements {(4, candy, 4, {(4, hard candy, 4, phi)})}
Is moved to L [f ₁ ], and the process proceeds to step S512.
In step S512, the minimum cumulative weight in L [f ₁ ] is 4, so W [f ₁ ] = 4.

Exit correction auxiliary hypothesis _{f 1} at step S513, the flow returns to step S409.
After the correction procedure _{f 1} at step S409, the process proceeds to step S410.
In step S410, insert the _{f 1} to H ', the flow returns to step S406.
Since E ₁ = φ in step S406, the process proceeds to step S411.
Since H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(3, candy, 4, {(4, rainball, 4, {(4, jade ball, 4, φ)})}) }
To H.

Proceeding to step S413, since the next input symbol exists, the process returns to step S404.
[Processing of input symbol “m”]
In step S404, the symbol “m” is read, and in step S405, the hypothesis h = {(0, candy, 4, {(4, rainball, 4, {(4, jade ball, 4, φ)}. )})}
Take out. State transition of input symbol “m” from current state 3 of hypothesis generation rule T ₁ of this hypothesis <0 → 1, m: ε / 0>
Make a list _{E 1,} including.

Since _{E 1} = not at φ at step S406 proceeds to step S407, the state transition from _{E 1 <0 → 1, m} : ε / 0> is a list _{e 1} removed, new hypothesis f ₁ = (0 at step S408, Ameda, 4, {(4, rainball, 4, {(4, jade ball, 4, φ)})})
Is generated. Since o [e ₁ ] = ε in step S409, the process proceeds to step S410.
In step S410 the _{f 1} is inserted into H ', processing proceeds to step S406.
In step S406, since E ₁ = φ, the process proceeds to step S411, and since H = φ, the process proceeds to step S412.

In step S412, the element of H ′ {(1, Ameda, 1, {(4, rainball, 4, {(4, jade ball, 4, φ)})})}
Move all to H.
Proceeding to step S413, since the next input symbol exists, the process returns to step S404.
[Processing of input symbol “a”]
In step S404, the symbol “a” is read, and in step S405, a hypothesis h = (1, Ameda, 4, {(4, rainball, 4, {(4, amatama, 4, φ)}) })
Take out. State transition of input symbol “a” from current state 0 of hypothesis generation rule T ₁ of this hypothesis <1 → 0, a: ma / 0>
Make a list _{E 1,} including.

Step S406 because _{E 1} = not a φ in the procedure proceeds to step S407, the state transition from the list _{E 1 <1 → 0, a} : or / 0> and _{e 1} taken out, new hypothesis f ₁ = (0 in step S408, Amedama, 4, {(4, rainball, 4, {(4, jade ball, 4, φ)})})
Is generated. Since step S409 o _{[e 1]} is not a ε, the process proceeds to step S501 to correct the hypothesis _{f 1.} At this time, m = 2 is set.
To generate an empty auxiliary hypothesis list _{L 2} in step S502.
In step S503, AW [f ₁ ] = 0.

In step S504, auxiliary hypothesis list L [f ₁ ] = {(4, rainball, 4, {(4, jade ball, 4, φ)})}
From the auxiliary hypothesis g ₂ = (4, rainball, 4, {(4, jade ball, 4, φ)})
, And state transition that can be transitioned from the state 4 of the hypothesis generation rule T _{2 with} the input symbol “MA” <4 → 5, MA: ε / 1>
Make a list _{E 2,} including.

Since E ₂ = φ is not satisfied in step S505, the process proceeds to step S506. State transition <4 → 5 or ε / 1> is extracted from list E ₂ and set as e _2, and auxiliary hypothesis f ₂ = (5, rainball, 5, {(4, jasper, 4, φ) in step S507. }).
Since o [e ₂ ] is ε in step S508, the process proceeds to step S509.
At step S509, the inserting the _{f 2} to _{L 2,} the process proceeds to step S505.
In step S505 _{E 2} is so φ, the process proceeds to step S510.
Since L [f ₂ ] is φ in step S510, the process proceeds to step S511.
Step S511 In _{L 2} elements {(5, rain ball, 5, {(4, hard candy, 4, phi)})}
Is moved to L [f ₁ ], and the process proceeds to step S512.

In step S512, since the minimum cumulative weight in L [f ₁ ] is 5, W [f ₁ ] = 5.
Exit correction auxiliary hypothesis _{f 1} at step S513, the flow returns to step S409.
After the correction procedure _{f 1} at step S409, the process proceeds to step S410.
In step S410, insert the _{f 1} to H ', the flow returns to step S406.
Since E ₁ = φ in step S406, the process proceeds to step S411.
Since H = φ in step S411, the process proceeds to step S412 and the element of H ′ {(0, candy, 5, {(5, rainball, 5, {(4, jade ball, 4, φ)})}) }
To H.

となり、その補助仮説ｈ_３の出力記号列“飴玉”を変換結果として出力し、ステップＳ４１５で記号列変換処理を終了する。 Proceeding to step S413, since there is no next input symbol, the process proceeds to step S414.
In step S414, the hypothesis h _{1 in} H is equal to (0, candy, 5, {(5, rainball, 5, {(4, jade ball, 4, φ)})}) has reached the end state. And its auxiliary hypothesis h ₂ = (5, rainball, 5, {(4, jasper, 4, φ)}) and its auxiliary hypothesis h ₃ = (4, jasper, 4, φ) The cumulative weight

Next, and it outputs the output symbol string "candy" auxiliary hypothesis h ₃ as the conversion result, and terminates the symbol string conversion processing in step S415.

  This conversion result is the same as the result when the input symbol string is “A, A, D, D” in the conventional symbol string conversion procedure using the two hypothesis generation units. In the present invention, a new hypothesis generation rule for converting a symbol string “a, m, e, d, a, m, a” input in Roman letters into a symbol string “A, M, D, M” is added to the result. Realize the process of converting Roman input to Kanji. That is, according to the present invention, by selecting whether to add or remove a hypothesis generation rule for converting a Roman character input into a Hiragana character, a symbol string conversion device that converts a Roman character input into a Kanji character, and a symbol that converts a Hiragana input into a Kanji character The column conversion device can be easily obtained, and the advantage that the conversion destination can be easily changed is obtained.

FIG. 17 shows a second embodiment of the present invention. This embodiment shows a case where speech recognition is performed using three hypothesis generation units 201-1, 201-2, and 201-3.
A voice feature symbol string extraction unit 506 extracts a time series of a short time acoustic pattern of a voice as an acoustic feature symbol string from the voice signal input unit 505 that inputs a voice signal. The extracted acoustic feature symbol string is input to the first-stage hypothesis generation unit 201-1 constituting the symbol string conversion unit 102 according to the present invention, and the first-stage hypothesis generation unit 201-1 converts the acoustic feature symbol string into a word-based symbol. Convert a column to a target hypothesis. For this purpose, the first-stage hypothesis generation unit 201-1 is stored in the acoustic model hypothesis generation rule storage unit 105-1, and the standard feature of the sound fixed unit (for example, element) is used for a short time (for example, 10 milliseconds). ) Stored in the acoustic model hypothesis generation rule having the acoustic pattern obtained by analysis and the word dictionary hypothesis generation rule storage unit 105-2 and corresponding to the pronunciation of various words A rule for generating a hypothesis for a word pronunciation dictionary to be read is read, and a hypothesis with an acoustic feature symbol string as a candidate for a word-based symbol string is generated. In this example, the hypothesis generated by the first-stage hypothesis generation unit 201-1 is that the second and third-stage hypothesis generation units 201-2 and 201-3 read the rules for generating the hypothesis generation for the language model 1 and 2, and the language Hypotheses are generated through a finite state transition process for increasing the likelihood of each of the hypotheses, the accumulated weight of each hypothesis group is corrected by the hypothesis correction units 301-1 to 301-M, and the hypothesis narrowing units 401-1 to 401- A case in which hypotheses are narrowed down by M and a speech recognition result is output to the symbol string output unit 106 is shown.

  Here, the language model 1 can be, for example, a language model that gives ease of general word connection, and the language model 2 shows easy word connection of sentences used in a specific field. It can be a language model to give. In this way, by combining the language model 1 that gives ease of general word connection and the language model 2 that gives ease of word connection specialized in sentences used in a specific field, In addition to providing high speech recognition accuracy for text in a specific field, speech recognition that maintains a certain level of speech recognition accuracy for text outside the specific field is also possible.

A method of describing a word pronunciation dictionary and a language model for speech recognition using weighted finite state transition rules is described in, for example, “Weighted finite-state transducers in speech” by M. Mohri, F. Pereira, M. Riley at the international conference ASR2000. recognition ", Proceeding of ASR2000, pp.97-106,2000.
As an acoustic model that represents a set of standard acoustic pattern sequences of various speech fixed units (for example, phonemes), for example, a hidden Markov model method that models a set of these acoustic pattern sequences based on probability / statistical theory ( Hidden Markov Model (hereinafter referred to as HMM) is the mainstream. Details of the HMM method are disclosed in, for example, “Recognition of Speech by Stochastic Model” by Seiichi Nakagawa, edited by the Institute of Electronics, Information and Communication Engineers.

In the case of speech recognition, the score of the acoustic feature symbol (acoustic pattern) calculated by the acoustic model is used as the weight of the first-stage hypothesis generation unit. However, the larger the score, the closer the input acoustic pattern is to the sound fixed unit represented by the acoustic model, so a negative acoustic score is used as the weight. In the calculation of the acoustic score by the hidden Markov model, for example, a probability value based on a Gaussian distribution is used.
Acoustic patterns used for speech recognition include mel-frequency cepstral coefficients (referred to as MFCC), delta MFCC, LPC cepstrum, logarithmic power obtained by analyzing speech signals every short time (eg, 10 milliseconds). and so on.

As described in the verification destination of the speech recognition rate, according to the present invention, in the procedure for converting the symbol string using the three hypothesis generation processes, a search process using only the first hypothesis generation process except for the hypothesis correction unit ( The calculation procedure is almost the same as in FIG. 4), and efficient speech recognition can be performed.

Table 1 shows the processing time and speech recognition accuracy required for speech recognition processing when speech for 10 academic lectures by 10 men is input using the speech recognition method according to the present invention. However, the processing time is a value (actual time ratio) obtained by dividing the actual utterance time by the processing time required for the speech recognition processing. The vocabulary size of the word pronunciation dictionary is 30,000. In the experiment, a computer corresponding to a clock number of 2.8 GHz was used. In the conventional method, only two hypothesis generators (a word dictionary hypothesis generator and a general sentence hypothesis generator) are used. In the present invention, in addition to the above two hypothesis generators, the technology of the academic lecture is used. It uses a hypothesis generator that operates with reference to a language model 2 specialized in the field text.

  Compared to the conventional symbol string conversion method using two hypothesis generation units, the symbol string conversion method using three hypothesis generation units according to the present invention has a higher value in word recognition rate while maintaining substantially the same real-time ratio. showed that. Such a result was obtained by using the language model 2 specialized in the technical field of academic lectures. Since the present invention can use three or more hypothesis generators at the same time as described above, by combining various hypothesis generators with almost no increase in processing time compared to the case of using two hypothesis generators. The accuracy of voice recognition can be improved.

  On the other hand, the present invention can be applied to speech paraphrase processing, and efficient speech paraphrase processing can be performed efficiently using four hypothesis generation units. Here, the voice paraphrasing process represents converting the utterance content into another sentence style while performing voice recognition. For example, there is a process of rewriting a spoken sentence into a style like written language.

  FIG. 18 shows an embodiment of the speech paraphrasing device 600. The voice signal sent from the voice signal input unit 505 that inputs voice is converted into an acoustic feature symbol string by a voice feature symbol string extraction unit 506 that extracts the time series of the short-time acoustic pattern of the voice as a symbol string, and the sound The feature symbol string is input and sent to the symbol string converter 102 which performs symbol string conversion according to the present invention. Subsequently, the symbol string conversion unit 102 analyzes the standard features of the speech fixed unit (for example, phonemes) from the acoustic model hypothesis generation rule storage unit 105-1 for each short time (for example, 10 milliseconds). A word pronunciation dictionary hypothesis generation rule that gives an acoustic model given by collating a series of acoustic patterns obtained from the word utterance dictionary hypothesis generation rule storage unit 105-2 with the pronunciation of various words as a series of fixed speech units. From the language model hypothesis generation rule storage unit 105-3, a language model hypothesis generation rule that gives ease of concatenation of spoken words, and from the paraphrase hypothesis generation rule storage unit 105-5, A paraphrase hypothesis generation rule that rewrites a word or a word string in a sentence is changed from a paraphrase destination language model hypothesis generation rule storage unit 105-6 to a simple phrase in the rephrased sentence. Read the paraphrase language model hypothesis generation rule that gives the ease of concatenation, read the acoustic feature symbol string sent from the acoustic feature symbol string extraction unit 506, obtain the output symbol string with the minimum cumulative weight, and obtain the symbol string output unit 106. The symbol string output unit 106 outputs the received output symbol string as a speech recognition result.

  For example, Japanese Unexamined Patent Application Publication No. 2004-110673 “Sentence Style Conversion Method, Sentence Style Conversion Device, Sentence Style Conversion” describes an individual hypothesis generation unit used to simultaneously convert and output a sentence style while recognizing speech. A storage medium storing a program and a sentence style conversion program ".

  The above-described symbol string conversion apparatus according to the present invention can be made to function as a symbol string conversion apparatus by a computer by installing the symbol string conversion program according to the present invention in the computer. The symbol string conversion program according to the present invention is written in a computer-readable program language, and is recorded on a recording medium such as a magnetic disk or CD-ROM that can be read by the computer. The computer is installed through these recording media or communication lines.

The figure showing an example of the weighted finite state conversion rule used for hypothesis generation. The figure which defined the weighted finite state conversion rule of FIG. 1 with the table | surface. The block diagram showing embodiment of the hypothesis generation method by a weighted finite state conversion rule. The flowchart explaining the procedure of the hypothesis generation method by a weighted finite state conversion rule. The figure showing an example of the weighted finite state conversion rule which converts the symbol string showing the series of kana into the kanji series corresponding to the kana. The figure showing an example of the weighted finite state conversion rule which gives the possibility of the connection of the kanji series in Japanese by weight. The block diagram showing embodiment of the conventional symbol sequence conversion method using two hypothesis generation parts. The flowchart explaining the procedure of the conventional symbol sequence conversion method using two hypothesis generation parts. 9 is a flowchart for explaining the continuation of FIG. The flowchart explaining the correction method of the hypothesis in the conventional symbol sequence conversion method using two hypothesis generation parts. The block diagram showing embodiment of the symbol sequence conversion method using three or more hypothesis generation parts by this invention. The flowchart explaining the procedure of the symbol sequence conversion method using the 3 or more hypothesis generation part in this invention. The flowchart for demonstrating the continuation of FIG. The figure explaining the correction | amendment method of the hypothesis or auxiliary hypothesis in the symbol sequence conversion method by the 3 or more weighted finite state converter in this invention. The figure showing an example of the weighted finite state conversion rule which converts the symbol string represented by a Roman character into the symbol string of corresponding Hiragana. The flowchart for demonstrating the continuation of FIG. The block diagram showing one Embodiment of the speech recognition method by this invention. The block diagram showing one Embodiment of the voice paraphrasing method by this invention.

Explanation of symbols

10 states
11 End state
12 State transition
101 Hypothesis generation rule storage
102 Symbol string converter
103 Symbol string input section
101-1. First stage hypothesis generation rule storage unit
101-2 latter-stage hypothesis generation rule storage unit
104-1 First-stage hypothesis generation rule storage unit
104-M M-th stage hypothesis generation rule storage unit
105-1 Hypothesis generation rule storage unit for acoustic model
105-2 Hypothesis generation rule storage unit for word utterance dictionary
105-3 Hypothesis Generation Rule Storage for Language Model 1
105-4 Hypothesis Generation Rule Storage for Language Model 2
105-5 Paraphrase Hypothesis Generation Rule Storage
105-6 Hypothesis Generation Rule Storage Unit for Paraphrase Destination Language Model
106 Symbol string output unit 201-1 to 201-M Hypothesis generation unit 301-1 to 301-M Hypothesis correction unit 401-1 to 401-M Hypothesis narrowing unit

Claims (10)

A symbol string input unit that sequentially reads a symbol string by a computer, and a symbol string that includes three or more M hypothesis generation units that convert a symbol string input according to a state transition rule into a hypothesis that is a candidate for another symbol string A conversion unit and a symbol string output unit that captures a conversion result output by the symbol string conversion unit are configured.
A hypothesis generation process in which the symbol string conversion unit generates a hypothesis including symbol strings obtained in a process in which state transition is possible with respect to an input symbol string input to each of the M hypothesis generation units;
A hypothesis correction process for correcting the cumulative weight value assigned to the generated hypothesis by a singular value in the cumulative weight assigned to the hypothesis generated by each hypothesis generation unit;
A hypothesis narrowing process that removes unnecessary hypotheses using the corrected cumulative weights and narrows down to a high likelihood hypothesis,
The symbol string conversion method characterized by performing this. The symbol string conversion method according to claim 1,
In the hypothesis correction process, the cumulative weight given to the hypothesis generated by the hypothesis generation unit is selected from the cumulative weight given to each hypothesis generated in the hypothesis generator in the next stage or the corrected cumulative weight. Each time a symbol is input from the symbol string input unit to the first-stage hypothesis generation unit, a hypothesis correction process for correcting using a cumulative weight exhibiting a singular value is generated by the M-1th-stage hypothesis generation unit. A symbol string conversion method, which is applied to all hypothetical cumulative weights generated by a first-stage hypothesis generating unit from hypothetical cumulative weights. In the symbol string conversion method according to claim 1 or 2,
The hypothesis narrowing-down process extracts hypotheses having cumulative weights exhibiting singular values among the corrected cumulative weights of hypotheses generated by the first-stage hypothesis generation unit when all input signal sequences are read. It includes a hypothesis extraction process and an input process that inputs the hypothesis extracted by this hypothesis extraction process as an input symbol string of the hypothesis generation section in the next stage, and executes these hypothesis extraction processes and input processes for each hypothesis generation section The hypothesis designated by the singular value of the accumulated weight corrected from the hypotheses generated by the M-th hypothesis generation unit is selected from the hypotheses generated by the first-stage hypothesis generation unit. A symbol string conversion method characterized in that a hypothesis extracted by the hypothesis extraction process is output as a symbol string conversion result from the hypotheses generated by the eye hypothesis generation means. A symbol having a symbol string input unit for sequentially reading symbol strings, and three or more cascaded hypothesis generation units that convert symbol strings input according to state transition rules into hypotheses that are candidates for other symbol strings A symbol string conversion device configured to include a string conversion unit and a symbol string output unit that captures a conversion result output from the symbol string conversion unit,
Corresponding to each of the M cascaded hypothesis generators, the cumulative weight assigned to each hypothesis generated by each hypothesis generator is assigned to each hypothesis generated by each subsequent hypothesis generator. A hypothesis correction unit that corrects by a singular value in the accumulated weight, and a hypothesis narrowing unit that removes unnecessary hypotheses using the accumulated weight corrected by the hypothesis correction unit and narrows down to a high likelihood hypothesis. Characteristic symbol string conversion device. In the symbol string conversion device according to claim 4,
The hypothesis correction unit includes the cumulative weight assigned to the hypothesis generated by each hypothesis generation unit, among the cumulative weights assigned to the hypotheses generated by the hypothesis generation unit in the next stage or the corrected cumulative weights. A weight correction unit that corrects using a cumulative weight that exhibits a singular value, and the weight correction unit performs the M-1 stage at each time a symbol is input from the symbol string input unit to the first stage hypothesis generation unit. A symbol string conversion device comprising: an iterative control unit that corrects over all of the cumulative weights of hypotheses generated by the first hypothesis generator from the cumulative weights of hypotheses generated by the hypothesis generator. In any of the symbol string conversion devices according to claim 4 or 5,
The hypothesis narrowing down unit is a hypothesis extracting unit that extracts a hypothesis having a cumulative weight that exhibits a singular value among the corrected cumulative weights of the hypotheses generated by each hypothesis generation unit when all of the input symbol strings are read. Input processing means for inputting the hypothesis extracted by the hypothesis extraction means as an input symbol string of the hypothesis generation section of the next stage, and these hypothesis extraction means and input processing means are connected to M stages from the hypothesis generation section of the first stage. It is specified by the singular value of the accumulated weight corrected from all hypotheses generated by the M-th hypothesis generation unit from the hypotheses generated by the first-stage hypothesis generation unit by operating in the hypothesis generation unit. A symbol string conversion device, wherein a hypothesis is selected, and a hypothesis extracted by the hypothesis extraction means is output as a symbol string conversion result from hypotheses generated by an M-th stage hypothesis generation unit.   4. An acoustic feature symbol string conversion processing step for extracting a symbol string representing the acoustic feature from the input speech signal, and the symbol string representing the acoustic feature converted in the acoustic feature symbol string conversion processing step. A speech recognition method, wherein the symbol sequence corresponding to the input speech signal is converted by the symbol sequence conversion method according to claim 1.   8. A speech recognition method according to claim 7, wherein an input speech signal is converted into an acoustic feature symbol string representing the acoustic feature, and the symbol string representing the acoustic feature is converted into a sequence of words, and a word obtained by the speech recognition method. A first hypothesis generation process that gives a smaller weight when the sequence of words is plausible as a sentence, a second hypothesis generation process that replaces the word sequence obtained by the speech recognition method with a sequence of words in another language or style, The speech signal is generated using three or more hypothesis generation processes, in which the sequence of words in the other language or style is composed of a third hypothesis generation process that gives a smaller weight to a case that is likely to be a sentence of the other language or style. A speech paraphrasing method characterized by converting to a sentence in another language or style and outputting it from a symbol string output unit.   A program that is described in a program language that can be read by a computer, and that causes the computer to function as the device according to any one of claims 4 to 6.   A recording medium comprising a computer-readable recording medium, wherein the program according to claim 9 is recorded on the recording medium.

Images