Information processing device, information processing method, and recording medium

Abstract

An information processing device according to the present invention: performs an in-mask token search for each token, with regard to an in-mask token sequence that includes not less than one in-mask token which is masked and that is included in a temporary token sequence which is generated by processing a signal and in which not less than one token is masked; when a result of the in-mask token search is identical to an immediately subsequent token that immediately follows the in-mask token sequence and that is not masked, generates an in-mask / out-of-mask token sequence in which a token up to the in-mask token is the same; performing out-of-mask token search for each token in parallel with the in-mask token search for each token, for a token subsequent to the in-mask token of the in-mask / out-of-mask token sequence; and, when the in-mask token search for each token ends, selects an appropriate search result token sequence from one or more search result token sequences.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to the technical fields of an information processing device, an information processing method, and a recording medium.

[0002] As a speech recognition technology, a technology has been proposed in which, among the recognition results obtained by NAR (non-autoregressive) decoding, tokens with low reliability are recognized by AR (autoregressive) decoding (see Patent Document 1).

[0003] M. Someki, N. Eng, Y. Higuchi and S. Watanabe, "Segment-Level Vectorized Beam Search Based on Partially Autoregressive Inference," IEEE Automatic Speech Recognition and Understanding Workshop 2023

[0004] An object of this disclosure is to provide an information processing device, an information processing method, and a recording medium that aim to improve the technology related to the prior art documents mentioned above.

[0005] One aspect of the information processing device includes an in-mask token search means for performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens, wherein the one or more tokens are masked and included in the provisional token sequence; a generation means for generating an in-mask / out-of-mask token sequence in which, when the search result of the relevant in-mask token is the same as the unmasked token immediately following the in-mask token sequence, the tokens up to the relevant in-mask token are the same; an out-of-mask token search means for performing an out-of-mask token search for each token in the in-mask / out-of-mask token sequence in parallel with the in-mask token search for each token by the in-mask token search means; and a selection means for selecting an appropriate search result token sequence from one or more search result token sequences when the in-mask token search for each token by the in-mask token search means has finished.

[0006] Another aspect of the information processing device includes a second processing means that outputs a second processed token sequence when a first processed token sequence generated by processing a signal is input, and an output means that outputs a provisional token sequence including a masked token sequence in which tokens that differ between the first processed token sequence and the second processed token sequence are masked.

[0007] One aspect of an information processing method is an information processing method executed by a computer, the information processing method including: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens, wherein the one or more tokens are masked and included in the provisional token sequence; and, if the search result for the in-mask token is identical to an unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are identical; performing an out-of-mask token search for each token in the in-mask / out-of-mask token sequence that is subsequent to the in-mask token in the in-mask token sequence, in parallel with the in-mask token search for each token by the in-mask token search means; and, when the in-mask token search for each token by the in-mask token search means is completed, selecting an appropriate search result token sequence from one or more search result token sequences.

[0008] Another aspect of the information processing method is an information processing method executed by a computer, which includes, when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence, and outputting a provisional token sequence including a masked token sequence in which tokens that differ between the first processed token sequence and the second processed token sequence are masked.

[0009] One aspect of the recording medium is an information processing method executed by a computer, the information processing method including: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal and that includes one or more masked in-mask tokens, and if the search result for the in-mask token is identical to the unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are identical; performing an out-of-mask token search for each token in the in-mask / out-of-mask token sequence after the in-mask token in the in-mask / out-of-mask token sequence in parallel with the in-mask token search for each token by the in-mask token search means; and, when the in-mask token search for each token by the in-mask token search means is completed, selecting an appropriate search result token sequence from one or more search result token sequences.

[0010] Another aspect of the recording medium is one in which a computer program is recorded to cause a computer to execute an information processing method, which includes, when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence, and outputting a provisional token sequence including a masked token sequence in which tokens that differ between the first processed token sequence and the second processed token sequence are masked.

[0011] FIG. 1 is a block diagram showing an example of a configuration of an information processing device according to an embodiment. FIG. 2 is a flowchart showing an example of an operation of the information processing device according to an embodiment. FIG. 3 is a block diagram showing an example of a configuration of an information processing device according to an embodiment. FIG. 4 is a flowchart showing an example of an operation of the information processing device according to an embodiment. FIG. 5 is a conceptual diagram showing a concept of an operation of the information processing device according to an embodiment. FIG. 6 is a block diagram showing an example of a configuration of an information processing device according to an embodiment. FIG. 7 is a flowchart showing an example of an operation of the information processing device according to an embodiment.

[0012] Hereinafter, an information processing device, an information processing method, and a recording medium according to an embodiment will be described with reference to the drawings. [1: First Embodiment]

[0013] A first embodiment of an information processing device, an information processing method, and a recording medium will be described with reference to Fig. 1 and Fig. 2. In the following, the first embodiment of the information processing device, the information processing method, and the recording medium will be described using an information processing device 10.

[0014] As shown in FIG. 1 , the information processing device 10 includes an in-mask token search unit 11, an in-mask / out-mask token sequence generation unit 12, an out-mask token search unit 13, and a selection unit 14. The information processing device 10 processes a provisional token sequence including an in-mask token sequence including one or more masked tokens. The provisional token sequence is a token sequence generated by processing a signal. The one or more masked tokens may be tokens with low reliability. The reliability may be an index calculated when generating the provisional token sequence. Masked tokens are referred to as in-mask tokens. Unmasked tokens are referred to as out-mask tokens. Consecutive in-mask tokens may be merged into one in-mask token sequence. In an in-mask token sequence, in-mask tokens may be adjacent to each other. The information processing device 10 searches for in-mask tokens included in the in-mask token sequence.

[0015] The in-mask token search unit 11 performs an in-mask token search for each token in the in-mask token sequence. The in-mask token search unit 11 performs an in-mask token search for each token in the in-mask token sequence that includes one or more masked tokens and is included in the provisional token sequence. The in-mask token search may be an operation of predicting a token corresponding to the in-mask token. The in-mask token search unit 11 may refer to the search results of tokens preceding the token to be searched, and adopt two or more candidates with high probabilities as the search result for the token to be searched. The in-mask token search may be, for example, a beam search. The in-mask token search unit 11 may perform the search using a decoder.

[0016] The operation of the information processing device 10 will be described with reference to the flowchart in Fig. 2. As shown in Fig. 2, the information processing device 10 determines whether the token to be searched is an in-mask token or an un-mask token (step S11). If the token to be searched is an in-mask token (step S11: in-mask token), the in-mask token search unit 11 performs an in-mask token search for each token (step S12).

[0017] When an in-mask token search is repeated for each token in the in-mask token sequence, an out-of-mask token (referred to as the "immediately following token") immediately following the in-mask token sequence appears in the search result. The number of in-mask tokens included in the in-mask token sequence is obtainable information. For example, if the in-mask token sequence includes four in-mask tokens, and the in-mask token search is repeated for each token five times, the result of the fifth in-mask token search may be the same as the immediately following token immediately following the in-mask token sequence.

[0018] However, the number of in-mask tokens included in the in-mask token sequence may contain an error. For this reason, even if information is acquired that the in-mask token sequence includes four in-mask tokens, there may be cases where the in-mask token search is repeated four times for each token, and the result of the fourth in-mask token search is the same as the token immediately following the in-mask token sequence, or where the in-mask token search is repeated six times for each token, and the result of the sixth in-mask token search is the same as the token immediately following the in-mask token sequence.

[0019] Furthermore, the in-mask token sequence may include an in-mask token that is identical to the immediately following token immediately after the in-mask token sequence. For example, if the in-mask token sequence includes four in-mask tokens and the second in-mask token from the beginning is identical to the immediately following token immediately after the in-mask token sequence, when the in-mask token search is repeated five times for each token, the result of the second in-mask token search and the result of the fifth in-mask token search may be identical to the immediately following token immediately after the in-mask token sequence. In other words, even if the result of the in-mask token search is identical to the immediately following token immediately after the in-mask token sequence, it does not necessarily mean that the searched token is an unmasked token. Therefore, if the result of the in-mask token search is identical to the immediately following token immediately after the in-mask token sequence, both an in-mask hypothesis that the token identical to the immediately following token is an in-mask token and an unmasked hypothesis that the token identical to the immediately following token is an unmasked token can be made.

[0020] Therefore, if the search result for the relevant in-mask token is the same as the token immediately following the in-mask token sequence (step S13: Yes), the in-mask / out-of-mask token sequence generation unit 12 generates a new token sequence with the same tokens up to the relevant in-mask token (step S14). The in-mask / out-of-mask token sequence generation unit 12 may generate a new token sequence based on the hypothesis that the relevant in-mask token is an out-of-mask token. The token sequence generated by the in-mask / out-of-mask token sequence generation unit 12 is a token sequence that includes both in-mask tokens and out-of-mask tokens, and is therefore referred to as an in-mask / out-of-mask token sequence. In other words, the in-mask / out-of-mask token sequence generation unit 12 generates an in-mask / out-of-mask token sequence in order to separately maintain the hypothesis that the relevant in-mask token is an out-of-mask token.

[0021] In the in-mask hypothesis, the token to be searched and the token next to the token to be searched are also treated as in-mask tokens. Therefore, for an in-mask token sequence in which an in-mask token search has been performed token-by-token, the next search is an in-mask token search. In the out-of-mask hypothesis, the token to be searched and the token next to the token to be searched are also treated as out-of-mask tokens. Therefore, for a generated in-mask / out-of-mask token sequence, the next search is an out-of-mask token search. That is, in this embodiment, if the search result of an in-mask token is the same as the token immediately following the in-mask token sequence, an in-mask token search and an out-of-mask token search are performed token-by-token in parallel for tokens following the in-mask token. If there are multiple token sequences, the operations from step S11 to step S15 are performed for each token sequence.

[0022] When the token to be searched is an unmasked token (step S11: unmasked token), the unmasked token search unit 13 performs an unmasked token search for each token (step S15). The unmasked token search unit 13 performs an unmasked token search for each token in parallel with the in-mask token search for each token by the in-mask token search unit 11. The unmasked token search may be an operation of selecting a corresponding token in the provisional token sequence. The unmasked token search is referred to as a "token-specified search."

[0023] When the in-mask token search unit 11 has finished the in-mask token search (step S16: Yes), the selection unit 14 selects an appropriate search result token sequence from one or more search result token sequences (step S17). For example, the in-mask token search unit 11 may finish the in-mask token search when it has searched all in-mask tokens included in the in-mask token sequence. For example, the selection unit 14 may select a search result token sequence with a high degree of confidence as an appropriate search result token sequence from one or more search result token sequences.

[0024] In this way, the information processing device 10 performs an information processing method that includes performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal and includes one or more masked in-mask tokens, where one or more tokens are included in the masked provisional token sequence; if the search result for the in-mask token is identical to the unmasked token immediately following the in-mask token sequence, generating an in-mask / out-mask token sequence in which the tokens up to the in-mask token are identical; performing a token-designated search for each token in the in-mask / out-mask token sequence for tokens subsequent to the in-mask token in the in-mask / out-mask token sequence in parallel with the in-mask token search for each token; and, when the in-mask search for each token is completed, selecting an appropriate search result token sequence from the one or more search result token sequences.

[0025] The information processing device 10 described above may be realized by a computer reading a computer program recorded on a recording medium. In this case, the computer program may cause the computer to execute an information processing method including: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal and includes one or more masked in-mask tokens, one or more of which are included in the provisional token sequence in which the tokens are masked; generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are identical if the search result for the in-mask token is identical to the unmasked token immediately following the in-mask token sequence; performing a token-designated search for each token in the in-mask / out-of-mask token sequence after the in-mask token in parallel with the in-mask token search for each token; and selecting an appropriate search result token sequence from the one or more search result token sequences when the in-mask search for each token is completed. [Technical Effect]

[0026] In the information processing device 10 disclosed herein, when a search for an in-mask token appears that is identical to the unmasked token immediately following the in-mask token string, an in-mask hypothesis and an out-of-mask hypothesis are generated, thereby preventing the in-mask token search from being erroneously terminated and causing an omission error. Also, the information processing device 10 performs an in-mask token search for each token with low reliability, thereby reducing the processing load. [2: Second embodiment]

[0027] A second embodiment of the information processing device, information processing method, and recording medium will be described with reference to FIGS. 3 to 6. Hereinafter, the second embodiment of the information processing device, information processing method, and recording medium will be described using an information processing device 20. Note that, for the second embodiment, descriptions that overlap with the description of the first embodiment will be omitted as appropriate. [2-1: Configuration of the information processing device 20]

[0028] The configuration of the information processing device 20 will be described with reference to Fig. 3. Fig. 3 is a block diagram showing the configuration of the information processing device 20.

[0029] 3 , the information processing device 20 includes a calculation device 21 and a storage device 22. The information processing device 20 may further include a communication device 23, an input device 24, and an output device 25. However, the information processing device 20 does not necessarily include at least one of the communication device 23, the input device 24, and the output device 25. The calculation device 21, the storage device 22, the communication device 23, the input device 24, and the output device 25 may be connected via a data bus 26.

[0030] The arithmetic device 21 includes at least one processor (i.e., one processor or multiple processors) as hardware. The processor may include, for example, a processor conforming to a von Neumann computer architecture. The processor conforming to the von Neumann computer architecture may include at least one of a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor may include, for example, a processor conforming to a non-von Neumann computer architecture. The processor conforming to the non-von Neumann computer architecture may include at least one of an FPGA (Field Programmable Gate Array) and an ASIC (Application Specific Circuit).

[0031] The arithmetic device 21 reads a computer program 221 including at least one of computer program code and computer program instructions. For example, the arithmetic device 21 may read the computer program 221 stored in the storage device 22. For example, the arithmetic device 21 may read the computer program 221 stored in a computer-readable, non-transitory recording medium using a recording medium reading device (not shown) included in the information processing device 20. The computer program 221 read from the recording medium may be stored in the storage device 22. The arithmetic device 21 may acquire (i.e., download or read) the computer program 221 from a device (not shown) located outside the information processing device 20 via the communication device 23 (or another communication device). The downloaded computer program 221 may be stored in the storage device 22.

[0032] The arithmetic device 21 executes the loaded computer program 221. As a result, logical functional blocks for executing the information processing method to be performed by the information processing device 20 are realized within the arithmetic device 21. In other words, the arithmetic device 21, together with the storage device 22 or the like in which the computer program 221 is recorded (in other words, together with the storage device 22 and the computer program 221 recorded in the storage device 22 or the like), can function as a controller or computer for realizing logical functional blocks for executing the processing to be performed by the information processing device 20. In other words, the at least one processor included in the arithmetic device 21, the memory (recording medium) included in the storage device 22 or the like, and the computer program 221 are configured so that the information processing device 20 performs the information processing to be performed by the information processing device 20.

[0033] A computational model that can be constructed by machine learning may be implemented in the computational device 21 by the computational device executing the computer program 221. An example of a computational model that can be constructed by machine learning is a computational model including a neural network (so-called artificial intelligence (AI)). In this case, learning of the computational model may include learning of parameters of the neural network (e.g., at least one of a weight and a bias). The computational device 21 may execute an information processing method using the computational model. In other words, the operation of executing the information processing method may include the operation of executing the information processing method using the computational model. Note that a computational model that has been constructed by offline machine learning using training data may be implemented in the computational device 21. Furthermore, the computational model implemented in the computational device 21 may be updated by online machine learning on the computational device 21. Alternatively, the calculation device 21 may execute the information processing method using a calculation model implemented in a device external to the calculation device 21 (i.e., a device provided outside the information processing device 20) in addition to or instead of the calculation model implemented in the calculation device 21.

[0034] The recording medium for recording the computer program 221 executed by the arithmetic device 21 may be at least one of a CD-ROM, CD-R, CD-RW, flexible disk, MO, DVD-ROM, DVD-RAM, DVD-R, DVD+R, DVD-RW, DVD+RW, Blu-ray (registered trademark), or other optical disk, a magnetic medium such as a magnetic tape, a magneto-optical disk, a semiconductor memory such as a USB memory, or any other medium capable of storing a program. The recording medium may include a device capable of recording a computer program (for example, a general-purpose device or a dedicated device in which the computer program 221 is implemented in a state in which it can be executed in at least one of the forms of software and firmware). Furthermore, each process or function included in the computer program 221 may be realized by a logical processing block realized within the arithmetic device 21 when the arithmetic device 21 (i.e., processor) executes the computer program 221, or may be realized by hardware such as a predetermined gate array (FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit)) provided in the arithmetic device 21, or may be realized in a form that mixes logical processing blocks and partial hardware modules that realize some elements of the hardware.

[0035] The storage device 22 includes at least one memory capable of storing desired data. In other words, the storage device 22 includes at least one memory containing desired data. For example, the storage device 22 may store a computer program 221 executed by the arithmetic device 21. In this case, the storage device 22 (memory) may be used as the above-mentioned recording medium for recording the computer program 221 executed by the arithmetic device 21. The storage device 22 may temporarily store data used by the arithmetic device 21 when the arithmetic device 21 is executing the computer program 221. The storage device 22 may also store data to be stored long-term by the information processing device 20. The storage device 22 may include at least one of a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device. In other words, the storage device 22 may include a non-temporary recording medium.

[0036] The communication device 23 may be capable of communicating with devices external to the information processing device 20. The communication device 23 may perform wired communication or wireless communication.

[0037] The input device 24 is a device capable of accepting information input to the information processing device 20 from outside. The input device 24 may include an operation device (e.g., a keyboard, a mouse, a touch panel, etc.) that can be operated by a user of the information processing device 20. The input device 24 may include a recording medium reading device that can read information recorded on a recording medium that is detachable from the information processing device 20, such as a USB (Universal Serial Bus) memory. Note that when information is input to the information processing device 20 via the communication device 23 (in other words, when the information processing device 20 acquires information via the communication device 23), the communication device 23 may function as an input device.

[0038] The output device 25 is a device capable of outputting information to the outside of the information processing device 20. The output device 25 may output visual information such as text or images, auditory information such as sound, or tactile information such as vibration, as the information. The output device 25 may include, for example, at least one of a display, a speaker, a printer, and a vibration motor. The output device 25 may be capable of outputting information to a recording medium detachable from the information processing device 20, such as a USB memory. Note that when the information processing device 20 outputs information via the communication device 23, the communication device 23 may function as the output device.

[0039] 3 shows an example of logical functional blocks realized in the arithmetic device 21 to execute the information processing method. As shown in FIG. 3, an in-mask token search unit 211, an in-mask / out-mask token sequence generation unit 212, an out-mask token search unit 213, and a selection unit 214 are realized in the arithmetic device 21. The "in-mask token search unit 211" is a component corresponding to the "in-mask token search unit 11" in the first embodiment described above, the "in-mask / out-mask token sequence generation unit 212" is a component corresponding to the "in-mask / out-mask token sequence generation unit 12" in the first embodiment described above, the "out-mask token search unit 213" is a component corresponding to the "out-mask token search unit 13" in the first embodiment described above, and the "selection unit 214" is a component corresponding to the "selection unit 14" in the first embodiment described above. [2-2: Information Processing Method Executed by Information Processing Device 20]

[0040] The information processing performed by the information processing device 20 can be applied to generation processes such as speech recognition processing, sentence generation processing, and translation processing. The information processing device 20 may be configured as a device for searching each token for low-reliability parts in the results of a collective search of the entire token sequence. The information processing device 20 may generate multiple search result token sequences and select an appropriate search result token sequence from the multiple search result token sequences. The number of search result token sequences to be generated may be appropriately determined based on the amount of calculation and accuracy.

[0041] An information processing method executed by the information processing device 20 will be described with reference to Fig. 4 to Fig. 6. Fig. 4 is a flowchart showing an example of the flow of the information processing method executed by the information processing device 20. Fig. 5 is a conceptual diagram showing an example of a provisional token sequence acquired by the information processing device 20. Fig. 6 is a conceptual diagram showing an example of the flow of the information processing method executed by the information processing device 20.

[0042] 4, the provisional token sequence acquisition unit 215 acquires a provisional token sequence (step S21). The provisional token sequence may be a token sequence generated by processing a signal. For example, if the information processing performed by the information processing device 20 is speech recognition processing, the provisional token sequence may be a token sequence generated by processing a speech signal. The provisional token sequence may be a token sequence in which one or more tokens are masked.

[0043] The provisional token sequence may be a token sequence generated by masking tokens with low reliability in a token sequence generated by processing a signal. The provisional token sequence may be a token sequence output by a provisional token generation mechanism P. The provisional token generation mechanism P may process an input signal collectively to output a token sequence, and may also output a provisional token sequence by masking tokens with low reliability. The provisional token generation mechanism P may mask tokens with reliability below a threshold.

[0044] As illustrated in FIG. 5, when a signal is input, the provisional token generation mechanism P generates a token string along with the reliability of each token. FIG. 5 illustrates a case where the next token is arranged on the right. FIG. 5 illustrates a case where the reliability of the third to fifth tokens is lower than the desired reliability. The provisional token generation mechanism P outputs a provisional token string in which tokens with low reliability are masked. FIG. 5 illustrates a case where the three tokens with low reliability, the third to fifth, are masked, and the provisional token generation mechanism P outputs a provisional token string including the tokens "so," "no," "in-mask token string including three tokens," "sha," "cho," "ni," "ju," "nin," "su," and "ru."

[0045] The search target token determination unit 216 determines whether the token to be searched for is the first token in the token string within the mask (step S22). The token to be searched for may be the token following a token that has already been searched for. If there is no token that has already been searched for, the token to be searched for may be the first token in the provisional token string.

[0046] If the token to be searched is not the first token in the in-mask token string (step S22: No), that is, if the token to be searched is an unmasked token, the unmasked token search unit 213 performs a token-specified search (step S23). The unmasked token search unit 213 performs a token-specified search for each unmasked token included in the provisional token string. The search target token determination unit 216 determines whether the searched token is the last token in the provisional token string (step S24). If the token is not the last token in the provisional token string (step S24: No), the process returns to step S22.

[0047] Fig. 6 illustrates an example of processing a provisional token string including the tokens "so", "no", "sha", "cho", "ni", "ju", "nin", "su", and "ru" in a masked token string including three tokens illustrated in Fig. 5. As illustrated in Fig. 6, consecutive masked tokens may be merged into one masked token string.

[0048] In the example shown in FIG. 6, the first token is an unmasked token, and a token-specified search is performed. The token-specified search may be an operation of searching for a token contained in a provisional token sequence that corresponds to the token to be searched for. Since the first token contained in the provisional token sequence is "so", the first token may be "so". Therefore, the token sequence resulting from the first search may be "so".

[0049] Furthermore, the second token is an unmasked token, and a token-specified search is performed. Since the second token contained in the provisional token sequence is "の", the second token may be "の". Therefore, the token sequence resulting from the second search may be "そ".

[0050] Furthermore, in step S23, the score acquiring unit 217 may acquire a token sequence score. The token sequence score may be a score indicating the likelihood of the search result token sequence. The token sequence score may be updated by searching each token. In this embodiment, the token sequence score may be, for example, a Connectionist Temporal Classification (CTC) score disclosed in the document "Shinji Watanabe, Takaaki Hori, Suyoun Kim, John R Hershey, and Tomoki Hayashi, "Hybrid CTC / attention architecture for end-to-end speech recognition," IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 8, pp. 1240-1253, 2017."

[0051] If the token to be searched is the first token in the in-mask token sequence (step S22: Yes) and the token to be searched is an in-mask token (step S11: in-mask token), the in-mask token search unit 211 performs an in-mask token search (step S12). Note that if the token to be searched is the first token in the in-mask token sequence (step S22: Yes), the token to be searched is an in-mask token (step S11: in-mask token). The in-mask token search may be an operation to search for the token to be searched for, regardless of the provisional token sequence. The in-mask token search may be an operation to search for a likely token that follows the search result token sequence. The in-mask token search in this embodiment may be a beam search. That is, in this embodiment, multiple search result token sequences may be generated. This embodiment illustrates, as an example, a case where a beam search with a beam width of 2 is performed.

[0052] Furthermore, in step S12, the score acquiring unit 217 may acquire a token score and a token sequence score. The token score is a score different from the token sequence score. The token sequence score may be the same as the score acquired in the above-described step S23, or may be the above-described CTC score. The token score may be a score indicating the likelihood of a token continuing from the immediately preceding token sequence. The token score may be a decoder score. The score acquiring unit 217 may calculate a score for the search result token sequence based on the token score and the token sequence score. For example, the score acquiring unit 217 may calculate the score for the search result token sequence by adding the token score and the token sequence score. Furthermore, the score acquiring unit 217 may calculate the score for the search result token sequence by weighting and adding the token score and the token sequence score. The token search within the mask may be joint decoding using CTC and a decoder. In other words, the token search within the mask may be joint decoding using CTC / attention.

[0053] In the example shown in Figure 6, the third token is the head of the in-mask token string and is an in-mask token. Therefore, an in-mask token search is performed on the third token. Figure 6 illustrates a case where "kai" and "gai" are searched for as two tokens that are highly likely to follow "sono". In other words, the token string resulting from the third search may be "sono kai" and "sono gai".

[0054] 6, the token immediately following the in-mask token sequence is "sha" (company), which is different from both "kai" (group) and "gai" (outside). In other words, the in-mask token search result is not the same as the immediately following token (step S13: No), and the search target token determination unit 216 determines whether to end the search of the in-mask token sequence (step S16).

[0055] As described above, the number of tokens included in the in-mask token sequence is obtainable information. Therefore, the in-mask token search unit 211 may end the in-mask token search when it has performed an in-mask token search for the number of tokens included in the in-mask token sequence. Also, as described above, the number of tokens included in the in-mask token sequence may include an error. Therefore, the in-mask token search unit 211 may end the in-mask token search when it has performed a predetermined number of in-mask token searches that are greater than the number of tokens included in the in-mask token sequence. In other words, the in-mask token search unit 211 may end the in-mask token search when it has performed a predetermined number of in-mask token searches. The predetermined number may be a constant such as "2". Also, the predetermined number may be a number corresponding to the number of tokens included in the in-mask token sequence, such as half the number of tokens included in the in-mask token sequence.

[0056] In this embodiment, an example is shown in which it is determined that the in-mask token search is to end when "two" more in-mask token searches have been performed than the number of tokens included in the in-mask token sequence, that is, when "five" in-mask token searches have been performed. Since the third search is the first in-mask token search, the determination in step S16 is No, and the process proceeds to step S11. Note that, as described above, when there are multiple token sequences, the operations from step S11 to step S15 are performed for each token sequence.

[0057] In the example shown in Figure 6, the tokens to be searched following "sono kai" (the meeting) and "sono gaitachi" (so outside) are tokens within the mask. Therefore, a token search within the mask is performed for the fourth token. Figure 6 illustrates a case where a search is performed for tokens following "sono kai" (the meeting) or "sono gaitachi" (so outside), and the two search result token sequences with the highest probability are "sono kaisha" (the company) and "sono gaitachi" (eating out).

[0058] In the example shown in FIG. 6 , the immediately following token is “company”, which is the same as the “company” in “the company”. Therefore, the inside-mask / outside-mask token sequence generation unit 212 generates an inside-mask / outside-mask token sequence called “the company” (step S14). As a result, “the company”, which is an inside-mask token sequence based on the hypothesis that “company” is an inside-mask token, and “the company”, which is an inside-mask / outside-mask token sequence based on the hypothesis that “company” is an outside-mask token, exist. In this way, the token sequence resulting from the fourth search may be “the company” and “the eating out”, which are inside-mask token sequences, and “the company”, which is an inside-mask / outside-mask token sequence. Note that bold italics in FIG. 6 indicate that a token is an outside-mask token. The score of the inside-mask / outside-mask token sequence generated in step S14 may be the same as the score of the inside-mask token sequence that includes the same token sequence as the inside-mask / outside-mask token sequence. In other words, the score of the inside-mask / outside-mask token sequence generated in step S14 may be the same as the score calculated in step S12.

[0059] In the example shown in Figure 6, the tokens to be searched following the in-mask token sequences "the company" and "eating out" are in-mask tokens. Therefore, an in-mask token search is performed on the fifth token following "the company" or "eating out". Figure 6 illustrates a case where tokens following "the company" or "eating out" are searched for, and the two search result token sequences with the highest probability are "the company's" and "the eating out's".

[0060] In the example shown in Figure 6, the token to be searched following "the company" in the inside-outside-mask token sequence is an outside-mask token (step S11: outside-mask token). Therefore, a token-specified search is performed on the fifth token following "the company" (step S15). As shown in the example in Figure 6, the token following the immediately following token "company" in the provisional token sequence is "chief", so the fifth token may be "chief". In this way, the fifth search result token sequence may be the inside-mask token sequences "the company" and "the eating out company", and the inside-out-mask token sequence "the company president".

[0061] Furthermore, in step S15, the score acquiring unit 217 may acquire at least one of a token score and a token sequence score. The token score may be the same as the score acquired in the above-mentioned step S12, or may be the above-mentioned decoder score. Furthermore, the token sequence score may be the same as the score acquired in the above-mentioned step S23, or may be the above-mentioned CTC score. The score acquiring unit 217 may calculate a score based on at least one of the token score and the token sequence score.

[0062] In the search for tokens outside the mask in step S15, only the token sequence score (e.g., CTC score) may be used. That is, in the search for token sequences within the mask, the token sequence score and the token score (e.g., decoder score) are basically used, but for some tokens to be searched, only the token sequence score may be used without using the token score. This allows the entire section of the token sequence within the mask to be processed more quickly than when both the token score and the token sequence score are used.

[0063] When acquiring both the token score and the token sequence score, the score acquiring unit 217 may calculate the score in the same manner as in step S12. For example, when a search result token sequence after a token-specified search for each token is identical to a search result token sequence after an in-mask token search for each token is performed in parallel, the score acquiring unit 217 may acquire the token score in step S15. For example, the token sequence of the seventh search result, described below, "the president of the company," corresponds to a case where a search result token sequence after a token-specified search for each token is identical to a search result token sequence after an in-mask token search for each token is performed in parallel.

[0064] In the example shown in Figure 6, the tokens to be searched following the in-mask token sequences "sono kiyan no" and "sono gaishoku no" are in-mask tokens. Therefore, an in-mask token search is performed for the sixth token following "sono kiyan no" or "sono gaishoku no". Figure 6 illustrates a case where a search is performed for tokens following "sono kiyan no" or "sono gaishoku no" and the two search result token sequences with the highest probability are "sono kiyan no sha" and "sono gaishoku no shi".

[0065] The "sha" in "the company's company" is the same as the immediately following token "sha." Therefore, the inside-mask / outside-mask token sequence generation unit 212 generates the inside-mask / outside-mask token sequence "the company's company." This creates two token sequences: "the company's company," which is an inside-mask token sequence based on the hypothesis that "sha" is an inside-mask token, and "the company's company," which is an inside-mask / outside-mask token sequence based on the hypothesis that "sha" is an outside-mask token.

[0066] In the example shown in Figure 6, the token to be searched following "the company president", which is an inside-mask / outside-mask token sequence, is an outside-mask token. Therefore, a token-specified search is performed on the sixth token following "the company president". As shown in the example in Figure 6, the token following "long" in the provisional token sequence is "ni", so the sixth token may be "ni". In this way, the token sequence resulting from the sixth search may be "the company's company" and "the eating out company's company", which are inside-mask token sequences, and "the company's company" and "the eating out company's company", which are inside-mask token sequences.

[0067] In the example shown in Figure 6, the tokens to be searched following the in-mask token sequences "the company" and "the eating out sign" are in-mask tokens. Therefore, an in-mask token search is performed on the seventh token following "the company" or "the eating out sign". Figure 6 illustrates a case where tokens following "the company" or "the eating out sign" are searched for, and the two search result token sequences with the highest probability are "the company president" and "the eating out sign".

[0068] In the example shown in FIG. 6 , the tokens to be searched following the token sequence inside and outside the mask, "To the president of that company" and "The company of that company", are outside the mask. Therefore, a token-specified search is performed on the seventh token following "To the president of that company" and "The company of that company". As shown in the example shown in FIG. 6 , the token following "Ni" in the provisional token sequence is "Ju", so the seventh token following "To the president of that company" may be "Ju". As shown in the example shown in FIG. 6 , the token following "Sha" in the provisional token sequence is "Naga", so the seventh token following "The company of that company" may be "Naga". In this way, the seventh search result token sequence may be the token sequence inside the mask, "The president of that company" and "The restaurant owner", and the token sequence inside and outside the mask, "To the president of that company" and "The president of that company".

[0069] The seventh search is the case where the in-mask token search has been performed "5" times. Therefore, the in-mask token search is terminated as described above (step S16: Yes), and the selection unit 214 selects an appropriate search result token sequence (step S17). The selection unit 214 may select an appropriate search result token sequence from a search result token sequence with a high score.

[0070] For example, the score of the search result up to "the company" in "become the president of that company" in the seventh search result may be a score based on the token score and the token sequence score. The score of the token sequence after "the company" may be a score based on the token sequence score. Furthermore, the score of "the president of that company", which is a token sequence inside and outside the mask in the seventh search result, may be the same as the score of "the president of that company", which is a token sequence inside the mask.

[0071] Since more in-mask token searches have been performed than the number of tokens included in the in-mask token sequence, the selection unit 214 may determine that search result token sequences that do not include an immediately following token are inappropriate. In other words, the selection unit 214 may first exclude search result token sequences that do not include an immediately following token from the options, and then select a search result token sequence with a high score as an appropriate search result token sequence. In the example shown in Figure 6, the selection unit 214 may select "the president of that company" with the higher score from "become the president of that company" and "the president of that company," and determine that the in-mask token sequence is "of that company."

[0072] In the above-described embodiment, whether to terminate the in-mask token search is determined based on the number of searches, but this is not limiting. For example, the in-mask token search may be terminated when the two search result token sequences with the highest scores are in-mask and out-of-mask token sequences.

[0073] The search target token determination unit 216 determines whether the searched token is the last token of the provisional token sequence (step S24). If the searched token is not the last token of the provisional token sequence (step S24: No), the process returns to step S22. In other words, the information processing device 20 generates a sequence of tokens. [2-4: Technical Effects]

[0074] The information processing device 20 according to this disclosure selects an appropriate search result token sequence from one or more search result token sequences based on the score, thereby enabling highly accurate information processing. [3: Third Embodiment]

[0075] A third embodiment of an information processing device, an information processing method, and a recording medium will be described with reference to FIGS. 7 to 9. Hereinafter, the third embodiment of an information processing device, an information processing method, and a recording medium will be described using an information processing device 30. The third embodiment differs from the first and second embodiments in the method of compressing past performance information. Note that, for the third embodiment, descriptions that overlap with the descriptions of the first and second embodiments will be omitted as appropriate. Note that, in the drawings, parts common to the first and second embodiments are designated by the same reference numerals.

[0076] 7, the arithmetic unit 21 included in the information processing device 30 includes a temporary token sequence generation unit 318 as a logical functional block. The temporary token sequence generation unit 318 may include a decoder unit 3182 and a mask unit 3183. [3-1: Information Processing Method Executed by the Information Processing Device 30]

[0077] The information processing device 30 is configured as a device that generates a provisional token sequence. An information processing method executed by the information processing device 30 will be described with reference to Fig. 8. Fig. 8 is a flowchart showing an example of the flow of the information processing method executed by the information processing device 30.

[0078] 8, the provisional token sequence generator 318 acquires a first processed token sequence (step S31). The first processed token sequence is a token sequence generated by processing a signal. The signal may be, for example, an audio signal. The first processed token sequence may be a token sequence output by a CTC decoder to which the signal is input.

[0079] When the decoder unit 3182 receives a first processed token sequence, it outputs a second processed token sequence (step S32). The decoder unit 3182 outputs the token sequence all at once, rather than outputting the processing results one token at a time. When a token sequence is received, the decoder unit 3182 may have a mechanism that is capable of outputting the processing results one token at a time, or outputting the token sequence all at once. In step S32, the decoder unit 3182 outputs the token sequence all at once.

[0080] The masking unit 3183 generates a provisional token sequence by masking tokens that differ between the first processed token sequence and the second processed token sequence (step S33). Tokens that differ between the first processed token sequence and the second processed token sequence are estimated to be tokens with low reliability. In other words, the provisional token sequence generation unit 318 generates a provisional token sequence by masking tokens with low reliability that are included in the token sequence generated by processing the signal.

[0081] Furthermore, the provisional token sequence generation unit 318 may mask tokens included in the provisional token sequence that have low reliability in at least one of the first processed token sequence and the second processed token sequence. Alternatively, the decoder unit 4182 may output the second processed token sequence by masking tokens that have low reliability in the processing result.

[0082] The information processing device 30 may include a CTC decoder unit as a mechanism for outputting a first processed token sequence when a signal is input. That is, the information processing device 30 may include both a mechanism for outputting a first processed token sequence and a mechanism for outputting a second processed token sequence. [3-2: Technical Effects]

[0083] The information processing device 30 according to this disclosure can accurately identify low-reliability tokens and acquire a masked provisional token sequence. Since low-reliability tokens are accurately masked, it becomes possible to perform processing appropriate to the low-reliability tokens. [4: Fourth Embodiment]

[0084] A fourth embodiment of an information processing device, an information processing method, and a recording medium will be described with reference to FIGS. 9 and 10. Hereinafter, the fourth embodiment of an information processing device, an information processing method, and a recording medium will be described using an information processing device 40. The fourth embodiment differs from the first to third embodiments in the method of compressing past performance information. Note that, for the fourth embodiment, descriptions that overlap with the descriptions of the first to third embodiments will be omitted as appropriate. Note that, in the drawings, parts common to the first to third embodiments are designated by the same reference numerals.

[0085] 9, the arithmetic unit 21 included in the information processing device 40 includes, as logical functional blocks, an in-mask token search unit 211, an in-mask / out-mask token sequence generation unit 212, an out-mask token search unit 213, a selection unit 214, a provisional token sequence acquisition unit 215, a search target token determination unit 216, a score acquisition unit 217, and a provisional token sequence generation unit 418. The provisional token sequence generation unit 418 may include a CTC decoder unit 4181, a decoder unit 4182, and a mask unit 4183. [4-1: Information Processing Method Executed by the Information Processing Device 40]

[0086] The information processing device 40 is configured as a device that generates a provisional token sequence and searches for tokens in a mask of the generated provisional token sequence. An information processing method executed by the information processing device 40 will be described with reference to Fig. 10. Fig. 10 is a flowchart showing an example of the flow of the information processing method executed by the information processing device 40.

[0087] As shown in FIG. 10 , when a signal is input (step S41: Yes), the CTC decoder unit 4181 processes the signal and outputs a first processed token sequence (step S42). The CTC decoder unit 4181 outputs the token sequence all at once, rather than outputting the processing results one token at a time. The CTC decoder unit 4181 may output the first processed token sequence by masking tokens whose processing results have a low reliability. The decoder unit 4182 receives the first processed token sequence and outputs a second processed token sequence (step S43). The decoder unit 4182 outputs the token sequence all at once, rather than outputting the processing results one token at a time. The decoder unit 4182 may output the second processed token sequence by masking tokens whose processing results have a low reliability.

[0088] The masking unit 4183 generates a provisional token sequence by masking tokens that differ between the first processed token sequence and the second processed token sequence (step S44). Tokens that differ between the first processed token sequence and the second processed token sequence are estimated to be tokens with low reliability. In other words, the provisional token sequence generation unit 418 generates a provisional token sequence by masking tokens with low reliability included in the token sequence generated by processing the signal. The provisional token sequence may include in-mask tokens based on the reliability determined during the first processing and the second processing, and in-mask tokens based on the difference between the token sequences.

[0089] The operation after step S44 may be an operation of re-searching for tokens with low reliability among the tokens searched for by the tentative token sequence generation unit 418. The search target token determination unit 216 determines whether the token to be searched for is the first token in the in-mask token sequence (step S22). If the token to be searched for is not the first token in the in-mask token sequence (step S22: No), that is, if the token to be searched for is an out-of-mask token, the out-of-mask token search unit 213 performs a token-specified search (step S23).

[0090] If the token to be searched is the first token in the in-mask token sequence (step S22: Yes) and the token to be searched is an in-mask token (step S11: in-mask token), the in-mask token search unit 211 performs an in-mask token search (step S12). If the result of the in-mask token search is the same as the immediately following token (step S13: Yes), the in-mask / out-of-mask token sequence generation unit 212 generates a new token sequence with the same tokens up to the in-mask token (step S14).

[0091] The operations from step S11 to step S14 are performed for each of a plurality of token sequences. If the token to be searched is an unmasked token (step S11: unmasked token), the unmasked token search unit 213 performs a token-specified search (step S15). After performing the operations from step S11 to step S14 for all token sequences, the search target token determination unit 216 determines whether or not to end the search for the in-mask token sequence (step S16). If the search for the in-mask token sequence is not to end (step S16: No), the process returns to step S11. If the search for the in-mask token sequence is to end (step S16: Yes), the selection unit 214 selects an appropriate search result token sequence (step S17). The search target token determination unit 216 determines whether the searched token is the last token of the provisional token sequence (step S24). If the token is not the last token of the provisional token sequence (step S24: No), the process returns to step S22. [4-2: Technical Effects]

[0092] The information processing device 40 according to this disclosure accurately masks tokens with low reliability, and therefore can perform processing appropriate to the tokens with low reliability. [5: Supplementary Note]

[0093] Some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes: [Supplementary Note 1] An information processing apparatus comprising: in-mask token search means for performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens that are included in the provisional token sequence in which one or more tokens are masked; generation means for generating an in-mask / out-of-mask token sequence in which, when a search result for the in-mask token is the same as an unmasked token immediately following the in-mask token sequence, the in-mask token search means for performing an out-of-mask token search for each token in the in-mask / out-of-mask token sequence in parallel with the in-mask token search for each token by the in-mask token search means; and selection means for selecting an appropriate search result token sequence from one or more search result token sequences when the in-mask token search for each token by the in-mask token search means is completed. [Supplementary Note 2] The information processing device according to Supplementary Note 1, wherein the out-of-mask token search means performs an out-of-mask token search for each of the unmasked out-of-mask tokens included in the provisional token sequence. [Supplementary Note 3] The information processing device according to Supplementary Note 1, wherein the in-mask token search means terminates the in-mask token search when a predetermined number of in-mask token searches have been performed. [Supplementary Note 4] The information processing device according to Supplementary Note 1, wherein the selection means selects the appropriate search result token sequence based on the score of each of the one or more search result token sequences.[Supplementary Note 5] The information processing device according to Supplementary Note 4, further comprising: score acquisition means for acquiring a token score used to select a selected token when performing the in-mask token search for the in-mask tokens; and acquiring a token sequence score of a token sequence including a selected token when performing the in-mask token search for the in-mask tokens and the out-mask token search for the unmasked out-mask tokens; wherein the selection means selects the appropriate search result token sequence based on a score of the search result token sequence calculated based on at least one of the token score and the token sequence score. [Supplementary Note 6] The information processing device according to Supplementary Note 1, wherein the in-mask token search is an operation of selecting a token having a relatively high score from a plurality of candidate tokens. [Supplementary Note 7] The information processing device according to Supplementary Note 1, wherein the out-mask token search is an operation of selecting a corresponding token from unmasked out-mask tokens included in the provisional token sequence. [Supplementary Note 8] The information processing device according to Supplementary Note 1, further comprising provisional token sequence generation means for generating the provisional token sequence by masking tokens with low confidence included in a token sequence generated by processing a signal. [Supplementary Note 9] The information processing device according to Supplementary Note 8, wherein the provisional token sequence generation means comprises: first processing means for outputting a first processed token sequence when a signal is input; and second processing means for outputting a second processed token sequence when the first processed token sequence is input, and generates the provisional token sequence by masking tokens where corresponding tokens in the first processed token sequence and the second processed token sequence are different. [Supplementary Note 10] The information processing device according to Supplementary Note 9, which masks tokens with low reliability in at least one of the first processed token sequence and the second processed token sequence from among unmasked unmasked tokens included in the provisional token sequence. [Supplementary Note 11] The information processing device comprises: second processing means for outputting a second processed token sequence when the first processed token sequence generated by processing a signal is input; and output means for outputting a provisional token sequence including an in-mask token sequence where tokens different between the first processed token sequence and the second processed token sequence are masked.[Supplementary Note 12] The information processing device according to Supplementary Note 11, comprising first processing means for generating the first processed token sequence, wherein the second processing means performs an in-mask token search for tokens included in the in-mask token sequence. [Supplementary Note 13] An information processing method executed by a computer, comprising: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens that are included in the provisional token sequence in which one or more tokens are masked; if a search result for the in-mask token is the same as an unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are the same; performing an out-of-mask token search for each token after the in-mask token in the in-mask / out-of-mask token sequence in parallel with the in-mask token search for each token; and when the in-mask token search for each token is completed, selecting an appropriate search result token sequence from one or more search result token sequences. [Supplementary Note 14] A recording medium having recorded thereon a computer program for causing a computer to execute an information processing method, the method including: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens, wherein the provisional token sequence is generated by masking one or more tokens; if the search result for the in-mask token is the same as the unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are the same; performing an out-of-mask token search for each token in parallel with the in-mask token search for each token in the in-mask / out-of-mask token sequence for tokens subsequent to the in-mask token in the in-mask / out-of-mask token sequence; and, when the in-mask token search for each token is completed, selecting an appropriate search result token sequence from one or more search result token sequences.[Supplementary Note 15] An information processing method executed by a computer, comprising: when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence, and outputting a provisional token sequence including an in-mask token sequence in which tokens different from the first processed token sequence and the second processed token sequence are masked. [Supplementary Note 16] A recording medium having recorded thereon a computer program for causing a computer to execute an information processing method, comprising: when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence, and outputting a provisional token sequence including an in-mask token sequence in which tokens different from the first processed token sequence and the second processed token sequence are masked.

[0094] Furthermore, the present invention may be subordinate to some or all of the configurations described in Supplementary Notes 2 to 10, which are subordinate to Supplementary Notes 1, and to each of Supplementary Notes 13 and 14, in a similar subordinate relationship to Supplementary Notes 2 to 10. Furthermore, the present invention may be subordinate to some or all of the configurations described in Supplementary Notes 12, which are subordinate to Supplementary Notes 11, and to each of Supplementary Notes 15 and 16, in a similar subordinate relationship to Supplementary Notes 12. Furthermore, not limited to Supplementary Notes 1, 11, 13, 14, 15, and 16, but also to various hardware, software, various recording means for recording software, or systems, in a similar manner, within the scope of each of the above-mentioned embodiments.

[0095] This disclosure may be modified as appropriate within the scope that does not contradict the gist or idea of ​​the invention that can be read from the claims and the entire specification, and information processing devices, information processing methods, and programs that involve such modifications are also included in the technical idea of ​​this disclosure.

[0096] 10, 20, 30, 40 Information processing device 11, 211, 411 In-mask token search unit 12, 212 In-mask / out-mask token sequence generation unit 13, 213 Out-mask token search unit 14, 214 Selection unit 215 Provisional token sequence acquisition unit 216 Search target token determination unit 217 Score acquisition unit 318, 418 Provisional token sequence generation unit 3182, 4182 Decoder unit 4181 CTC decoder unit 3183, 4183 Mask unit

Claims

1. An information processing device comprising: an in-mask token search means for performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens that are included in the provisional token sequence in which one or more tokens are masked; a generation means for generating an in-mask / out-of-mask token sequence in which, when the search result for the relevant in-mask token is the same as the unmasked token immediately following the in-mask token sequence, the tokens up to the relevant in-mask token are the same; an out-mask token search means for performing an out-mask token search for each token in the in-mask / out-of-mask token sequence after the relevant in-mask token in parallel with the in-mask token search for each token by the in-mask token search means; and a selection means for selecting an appropriate search result token sequence from one or more search result token sequences when the in-mask token search for each token by the in-mask token search means has ended.

2. The information processing apparatus according to claim 1, wherein the unmasked token search means performs an unmasked token search for each unmasked token included in the provisional token sequence.

3. The information processing apparatus according to claim 1, wherein said in-mask token searching means terminates said in-mask token searching when a predetermined number of in-mask token searches have been performed.

4. The information processing device according to claim 1, wherein said selection means selects the appropriate search result token sequence based on the score of each of said one or more search result token sequences.

5. The information processing device according to claim 4, further comprising a score acquisition means for acquiring a token score used to select a selected token when searching for the in-mask token among the in-mask tokens, and acquiring a token sequence score of a token sequence including a selected token when searching for the in-mask token among the in-mask tokens and when searching for the unmasked out-of-mask tokens, wherein the selection means selects the appropriate search result token sequence based on a score of the search result token sequence calculated based on at least one of the token score and the token sequence score.

6. The information processing device according to claim 1, wherein the token search within the mask is an operation of selecting a token with a relatively high score from a plurality of candidate tokens.

7. The information processing device according to claim 1, wherein the unmasked token search is an operation of selecting a corresponding token from unmasked unmasked tokens included in the provisional token sequence.

8. An information processing apparatus according to claim 1, further comprising provisional token sequence generating means for generating the provisional token sequence by masking tokens with low reliability included in the token sequence generated by processing a signal.

9. The information processing device according to claim 8, wherein the provisional token sequence generation means comprises: first processing means for outputting a first processed token sequence when a signal is input; and second processing means for outputting a second processed token sequence when the first processed token sequence is input; and generates the provisional token sequence by masking tokens that are different in corresponding tokens between the first processed token sequence and the second processed token sequence.

10. An information processing device according to claim 9, wherein, of the unmasked tokens included in the provisional token sequence, tokens with low reliability in at least one of the first processed token sequence and the second processed token sequence are masked.

11. An information processing device comprising: second processing means for outputting a second processed token sequence when a first processed token sequence generated by processing a signal is input; and output means for outputting a provisional token sequence including an in-mask token sequence in which tokens that differ from the first processed token sequence and the second processed token sequence are masked.

12. An information processing apparatus according to claim 11, further comprising: a first processing means for generating the first processed token sequence; and said second processing means for searching for tokens included in the in-mask token sequence.

13. An information processing method executed by a computer, comprising: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens, wherein the one or more tokens are masked; if the search result for the in-mask token is the same as the unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are the same; performing an out-of-mask token search for each token in the in-mask / out-of-mask token sequence after the in-mask token in the in-mask token sequence, in parallel with the in-mask token search for each token; and when the in-mask token search for each token is completed, selecting an appropriate search result token sequence from one or more search result token sequences.

14. A recording medium having recorded thereon a computer program for causing a computer to execute an information processing method, the method comprising: performing an in-mask token search for each token in an in-mask token sequence that is a provisional token sequence generated by processing a signal, the in-mask token sequence including one or more masked in-mask tokens, wherein the provisional token sequence is generated by masking one or more tokens; if the search result for the in-mask token is the same as the unmasked token immediately following the in-mask token sequence, generating an in-mask / out-of-mask token sequence in which the tokens up to the in-mask token are identical; performing an out-of-mask token search for each token in parallel with the in-mask token search for each token; and, when the in-mask token search for each token is completed, selecting an appropriate search result token sequence from one or more search result token sequences.

15. An information processing method executed by a computer, comprising: when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence; and outputting a provisional token sequence including an in-mask token sequence in which tokens that differ between the first processed token sequence and the second processed token sequence are masked.

16. A recording medium having a computer program recorded thereon for causing a computer to execute an information processing method including: when a first processed token sequence generated by processing a signal is input, outputting a second processed token sequence; and outputting a provisional token sequence including an in-mask token sequence in which tokens that differ between the first processed token sequence and the second processed token sequence are masked.

Images