Information processing device, generation method, and generation program

Abstract

The information processing device (100) includes: an acquisition unit (120) that acquires a plurality of character strings; a morphological analysis execution unit (130) that performs morphological analysis on the plurality of character strings; and a generation unit (150) that generates a learned model based on a plurality of words that are predicates among a plurality of part-of-speech words obtained by performing morphological analysis on the plurality of character strings and a plurality of words among the plurality of part-of-speech words, wherein the learned model infers the meaning of a first word among the plurality of words among the plurality of part-of-speech words.

Description

Technical Field

[0001] The present invention relates to an information processing device, a generation method, and a recording medium having a generation program recorded thereon. Background Art

[0002] For many years, research has been conducted on full-text retrieval technology for retrieving desired text from a large amount of text, document classification technology for classifying text according to its description content, and question-answering technology for computers to answer people's questions. Among these technologies, meaning processing, which analyzes the meaning of text content described in natural language, is important. Technologies related to meaning processing have been proposed (see Patent Document 1). In addition, technologies related to meaning processing are described in Non-Patent Document 1. For example, learning based on variance representation is exemplified in Non-Patent Document 1.

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application Laid-Open No. 62-221776

[0006] Non-patent literature

[0007] Non-patent document 1: Naokan Okazaki "Language Processing Distributed Expression Learning", Artificial Intelligence, Vol.31, No2, p189-201, 2016.3 Summary of the Invention

[0008] Problems to be solved by the invention

[0009] In the technique described in Non-Patent Document 1, a single word vector is assigned to each word. Therefore, only a single meaning is inferred for each word assigned a word vector. However, the meaning of a word changes depending on the context. Therefore, the technique described in Non-Patent Document 1 fails to address word ambiguity. Furthermore, the vectors in the technique described in Non-Patent Document 1 are dimensionally compressed, requiring significant processing time for learning.

[0010] An object of the present invention is to generate a learned model corresponding to ambiguity in a shorter processing time.

[0011] Means for solving problems

[0012] An information processing device according to one embodiment of the present invention is provided. The information processing device includes: an acquisition unit that acquires a plurality of character strings; a morphological analysis execution unit that performs morphological analysis on the plurality of character strings; and a generation unit that generates a learned model based on a plurality of words serving as predicates among a plurality of part-of-speech words obtained by performing morphological analysis on the plurality of character strings and a plurality of words among the plurality of part-of-speech words, the learned model inferring the meaning of a first word among the plurality of words among the plurality of part-of-speech words.

[0013] Effects of the Invention

[0014] According to the present invention, a learned model corresponding to ambiguity can be generated in a shorter processing time. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a diagram showing the hardware configuration of the information processing device according to the first embodiment.

[0016] Figure 2 This is a diagram showing functional blocks included in the information processing device in the learning phase of the first embodiment.

[0017] Figure 3 This is a flowchart (part 1) showing an example of processing executed by the information processing device in the learning phase of the first embodiment.

[0018] Figure 4 This is a flowchart showing an example of the learning process in the first embodiment.

[0019] Figure 5 This is a diagram showing an example (part 1) of a subword context matrix according to the first embodiment.

[0020] Figure 6 This is a diagram showing an example (part 2) of a subword context matrix in embodiment 1.

[0021] Figure 7 This is a flowchart (part 2) showing an example of processing executed by the information processing device in the learning phase of the first embodiment.

[0022] Figure 8 This is a diagram showing a specific example of generating a learned model from a question sentence and a response sentence according to the first embodiment.

[0023] Figure 9 This is a diagram showing functional blocks included in the information processing device in the operation phase of the first embodiment.

[0024] Figure 10 This is a flowchart showing an example of processing executed by the information processing device in the operation phase of the first embodiment.

[0025] Figure 11 It is a diagram showing a comparative example.

[0026] Figure 12 This is a diagram showing functional blocks included in the information processing device in the operation phase of the second embodiment.

[0027] Figure 13 This is a flowchart showing an example of processing executed by the information processing device in the operation phase of the second embodiment. DETAILED DESCRIPTION

[0028] Hereinafter, an embodiment will be described with reference to the drawings. The following embodiment is merely an example, and various modifications can be made within the scope of the present invention.

[0029] Implementation Method 1

[0030] <Learning Stage>

[0031] Figure 1 This figure illustrates the hardware configuration of the information processing device of Embodiment 1. Information processing device 100 is a device that executes the generation method. Information processing device 100 can also be referred to as a learning device. For example, information processing device 100 can be considered a cloud server. Alternatively, information processing device 100 can be part of a system.

[0032] The information processing device 100 includes a processor 101 , a volatile storage device 102 , and a nonvolatile storage device 103 .

[0033] Processor 101 controls the entire information processing device 100. For example, processor 101 may be a CPU (Central Processing Unit), an FPGA (Field Programmable Gate Array), or the like. Processor 101 may also be a multiprocessor. Information processing device 100 may also be implemented using a processing circuit, software, firmware, or a combination thereof. Furthermore, the processing circuit may be a single circuit or a composite circuit.

[0034] The volatile storage device 102 is the main storage device of the information processing device 100. For example, the volatile storage device 102 is a RAM (Random Access Memory). The non-volatile storage device 103 is an auxiliary storage device of the information processing device 100. For example, the non-volatile storage device 103 is an HDD (Hard Disk Drive) or an SSD (Solid State Drive).

[0035] Next, the functions of the information processing device 100 will be described.

[0036] Figure 2 This diagram shows functional blocks included in the information processing device in the learning phase of Embodiment 1. The information processing device 100 includes a storage unit 110 , an acquisition unit 120 , a morphological analysis execution unit 130 , a predicate segmentation unit 140 , and a generation unit 150 .

[0037] The storage unit 110 may be realized as a storage area reserved in the volatile storage device 102 or the nonvolatile storage device 103 .

[0038] Part or all of the acquisition unit 120, morphological analysis execution unit 130, predicate segmentation unit 140, and generation unit 150 may also be implemented as processing circuitry. Part or all of the acquisition unit 120, morphological analysis execution unit 130, predicate segmentation unit 140, and generation unit 150 may also be implemented as modules of a program executed by the processor 101. For example, the program executed by the processor 101 is also referred to as a generation program. For example, the generation program is recorded on a recording medium.

[0039] The acquisition unit 120 acquires learning data. Specifically, the acquisition unit 120 acquires multiple character strings. For example, the acquisition unit 120 acquires learning data containing multiple character strings. The learning data is text data. Furthermore, the learning data may also be referred to as a learning example sentence.

[0040] The morphological analysis execution unit 130 performs morphological analysis on a plurality of character strings, thereby obtaining a plurality of part-of-speech words.

[0041] Here, the acquisition unit 120 may sequentially acquire character strings as learning data, thereby acquiring multiple character strings. Furthermore, the morphological analysis execution unit 130 may perform morphological analysis on the sequentially acquired multiple character strings, thereby obtaining multiple part-of-speech words.

[0042] The functions of the predicate segmentation unit 140 will be described later. Furthermore, the functions of the predicate segmentation unit 140 may be included in the generation unit 150. Specifically, the information processing device 100 may include the storage unit 110, the acquisition unit 120, the morphological analysis execution unit 130, and the generation unit 150 including the functions of the predicate segmentation unit 140.

[0043] The generation unit 150 generates a learned model based on multiple words that are multiple predicates among multiple words with word types obtained by performing morpheme analysis on multiple strings, and multiple words among the multiple words with word types. In other words, the generation unit 150 generates a learned model based on multiple words corresponding to multiple predicates among the multiple words with word types and multiple words among the multiple words with word types. Here, the multiple words among the multiple words with word types are words with word types such as verbs, nouns, and adjectives. That is, the multiple words among the multiple words with word types can be words with any word type. In addition, the words that are predicates are words that are verbs, adjectives, descriptive verbs, or サ变 nouns.

[0044] In addition, the learned model can also be referred to as inference information. The inference information is information indicating the correspondence between each of the multiple words that are multiple predicates and each of the multiple relationship information. The relationship information indicates the relationality of each of the multiple words. The relationship information can be considered to be represented by a two-dimensional table. That is, the relationship information can be considered to be two-dimensional information. In the inference information, the words that are predicates are associated with the relationship information. Therefore, the inference information can be considered to be three-dimensional information. Thus, the learned model can be considered to be three-dimensional information.

[0045] In addition, as shown later, the generation unit 150 generates a learned model based on words such as "運転する", "作業する", "飛ばす", etc. among the multiple words with word types and words such as "ドライバー", "車", etc. among the multiple words with word types.

[0046] In addition, the generation unit 150 can also generate a learned model based on multiple words that are multiple predicates among the multiple words with word types and multiple words that are multiple nouns among the multiple words with word types.

[0047] In the following description, it is assumed that the generation unit 150 generates a learned model based on multiple words that are multiple predicates among the multiple words with word types and multiple words that are multiple nouns among the multiple words with word types.

[0048] In addition, as described in the subsequent operation phase, the information processing device can use the learned model to infer the meaning of the first word among multiple words in the multiple words with word types. For example, the information processing device can use the learned model to infer the meaning of "ドライバー". In addition, for example, "ドライバー" is the first word.

[0049] Next, the processing performed by the information processing device 100 will be described using a flowchart.

[0050] Figure 3 It is a flowchart (part one) showing an example of the processing performed by the information processing device in the learning phase of Embodiment 1.

[0051] (Step S11) The acquisition unit 120 acquires a plurality of strings.

[0052] (Step S12) The morpheme analysis execution unit 130, the predicate segmentation unit 140, and the generation unit 150 perform learning processing. In the learning processing, a learned model is generated.

[0053] (Step S13) The generation unit 150 stores the learned model in the storage unit 110. In addition, the generation unit 150 may also store the learned model in an external device that can be connected to the information processing device 100.

[0054] Figure 4 It is a flowchart showing an example of the learning processing of Embodiment 1. Figure 4 The processing of [corresponding part] corresponds to Step S12.

[0055] (Step S21) The morpheme analysis execution unit 130 performs morpheme analysis on a plurality of strings. Specifically, the morpheme analysis execution unit 130 divides a plurality of strings into a plurality of words with词性. Using a specific example, the morpheme analysis is described. For example, the morpheme analysis execution unit 130 divides the string "a driver who drives a car" in a plurality of strings into "car (noun)", "を (particle)", "drive (verb)", and "driver (noun)".

[0056] (Step S22) The predicate segmentation unit 140 generates a file for each predicate according to a plurality of words with词性. The predicate segmentation unit 140 registers a plurality of nouns in the string containing the predicate into the file for each predicate. For example, the predicate segmentation unit 140 registers "car (noun)" and "driver (noun)" in the string "a driver who drives a car" containing the predicate "drive (verb)" into the file of "drive (verb)". In the above example, a verb is used as the predicate, but other predicates can also be used.

[0057] In addition, a plurality of nouns in the string containing the predicate are called sub-learning example sentences. For example, the sub-learning example sentences are "car (noun)" and "driver (noun)".

[0058] (Step S23) The generation unit 150 generates a learned model. The learned model is generated according to reinforcement learning, supervised learning, transfer learning, or semi-supervised learning, etc. As an example, unsupervised learning using a sub-word context matrix with an increased number of layers for the predicate is described. Here, unsupervised learning refers to the following method: providing learning data that does not contain results (i.e., labels) to the learning device, thereby learning the characteristics of the learning data.

[0059] When setting the M sets of words that appear in the learning data as w1,..., wM In the case of, the sub-word context matrix obtained from the sub-learning example texts is represented by a model obtained by assigning each predicate label to matrix M, where the matrix M sets the sum #(i, j) of the number of times the words wi and wj in the learning data co-occur in the sub-learning example texts as m i，j of the elements. m i，j is represented by the following formula (1).

[0060] [Formula 1]

[0061] m i，j =#(i, j)…(1)

[0062] The finally generated model is represented by a tensor of layer 3. Specifically, the generated model is represented by a sub-word context matrix obtained by assigning each predicate label V K Moreover, the K types of predicates in the learning data are V1,..., V K . Further, in the application stage, the information processing device uses multiple sub-word context matrices and outputs the associated predicate labels as the meanings associated with the co-occurrence of words.

[0063] The generation unit 15**0** can also generate a learned model using the scale of self-mutual information. When the number of occurrences of the word w i is m i , the number of occurrences of the word w j is m j , and the total number of words in the learning data is N, the self-mutual information PMI i，j is represented by the following formula (2).

[0064] [Formula 2]

[0065]

[0066] Next, the processing will be specifically described. The generation unit 15**0** uses the files generated for each predicate. The generation unit 15**0** counts the sub-learning example texts of the same predicate, thereby generating a sub-word context matrix labeled with predicate labels. A sub-word context matrix is specifically shown.

[0067] Figure 5 is a diagram showing an example (one) of the sub-word context matrix of Embodiment 1. For example, the generation unit 15**0** counts m i，i ("driver (noun)", "driver (noun)"), m j，j ("car (noun)", "car (noun)"), m i，j ("driver (noun)", "car (noun)") and mj，i Increment both (“vehicle (noun)”, “driver (noun)”). Additionally, the initial state of the sub-word context matrix is 0.

[0068] This way, the generation unit 150 generates a sub-word context matrix labeled with a predicate label. The generation unit 150 repeatedly counts the sub-learning example sentences with the same predicate, thereby enabling the generation of multiple sub-word context matrices labeled with multiple predicate labels. That is, the generation unit 150 generates a learned model using multiple sub-word context matrices labeled with multiple predicate labels through learning based on multiple sub-learning example sentences. Here, a specific example of multiple sub-word context matrices labeled with multiple predicate labels is shown.

[0069] Figure 6 This is a diagram (part 2) showing an example of the sub-word context matrix of Embodiment 1. Figure 6 A sub-word context matrix showing the predicate label “drive (verb)”. Figure 6 A sub-word context matrix showing the predicate label “work (verb)”. Figure 6 A sub-word context matrix showing the predicate label “fly (verb)”.

[0070] After step S23 ends, the process proceeds to step S13.

[0071] This way, the information processing device 100 generates a learned model using multiple sub-word context matrices. Additionally, the information processing device 100 can also re-learn the learned model.

[0072] In the above text, the case where the acquisition unit 120 acquires multiple strings has been described. These multiple strings can also be the strings of a question sentence and the response sentence for that question sentence. Therefore, the case where the acquisition unit 120 acquires the strings of a question sentence and a response sentence is described.

[0073] Figure 7 This is a flowchart (part 2) showing an example of the processing executed by the information processing device in the learning phase of Embodiment 1.

[0074] In Figure 7 the processing, the difference from Figure 3 the processing is that steps S11a, 11b, and 11c are executed. Therefore, in Figure 7 this, steps S11a, 11b, and 11c are described. Regarding Figure 7 the other steps in Figure 3 this, by labeling the same step numbers as those in

[0075] (Step S11a) The acquisition unit 120 acquires the string of the interrogation sentence. In addition, the acquisition unit 120 may acquire multiple interrogation sentences.

[0076] (Step S11b) The acquisition unit 120 acquires the string of the response sentence. In addition, the acquisition unit 120 may acquire multiple response sentences.

[0077] (Step S11c) The morpheme analysis execution unit 130 combines the string of the interrogation sentence and the string of the response sentence.

[0078] In addition, in Step S12, the morpheme analysis execution unit 130 performs morpheme analysis on the combined string.

[0079] In addition, the acquisition unit 120 may acquire the string of the response sentence before acquiring the string of the interrogation sentence. Also, the morpheme analysis execution unit 130 may not combine the string of the interrogation sentence and the string of the response sentence. In the case where the combination is not performed, the morpheme analysis execution unit 130 in Step S21 performs morpheme analysis on multiple strings (i.e., the string of the interrogation sentence and the string of the response sentence).

[0080] Figure 8 It is a diagram showing a specific example in the case of generating a learned model based on an interrogation sentence and a response sentence in Embodiment 1.

[0081] In Figure 8 specific examples of the string of the interrogation sentence and the string of the response sentence are shown. For example, it is assumed that through morpheme analysis, "to fly (verb)", "driver (noun)", "flight distance (noun)", and "golf (noun)" are obtained. The generation unit 150 adds 1 to the information indicating the relationship between "driver (noun)" and "flight distance (noun)" in the sub-word context matrix corresponding to "to fly (verb)". The generation unit 150 adds 1 to the information indicating the relationship between "driver (noun)" and "golf (noun)" in the sub-word context matrix corresponding to "to fly (verb)".

[0082] In this way, the information processing apparatus 100 can generate a learned model based on the interrogation sentence and the response sentence.

[0083] <Operation stage>

[0084] Figure 9 It is a diagram showing the functional blocks of the information processing apparatus in the operation stage of Embodiment 1. The same as the Figure 2 structure shown is Figure 9 labeled with the same as Figure 2The information processing device 100 a includes a storage unit 110 , an acquisition unit 120 a , a morphological analysis execution unit 130 a , a designation acceptance unit 160 , an inference unit 170 , and an output unit 180 .

[0085] Here, information processing device 100 and information processing device 100a may be the same device or different devices. For example, if information processing device 100 and information processing device 100a are the same device, information processing device 100a includes predicate segmentation unit 140 and generation unit 150. Furthermore, if information processing device 100 and information processing device 100a are the same device, acquisition unit 120a performs the functions of acquisition unit 120, and morphological analysis execution unit 130a performs the functions of morphological analysis execution unit 130.

[0086] The acquisition unit 120a, morphological analysis execution unit 130a, designation acceptance unit 160, inference unit 170, and output unit 180 may be partially or entirely implemented by a processing circuit included in the information processing device 100a. The acquisition unit 120a, morphological analysis execution unit 130a, designation acceptance unit 160, inference unit 170, and output unit 180 may also be partially or entirely implemented as a module of a program executed by a processor included in the information processing device 100a. For example, the program may be stored on a recording medium.

[0087] The acquisition unit 120a acquires the character string data. Furthermore, the acquisition unit 120a acquires the learned model. For example, the acquisition unit 120a acquires the learned model from the storage unit 110. If the learned model is stored in an external device, the acquisition unit 120a acquires the learned model from the external device.

[0088] The morphological analysis execution unit 130 a executes morphological analysis on the character string data acquired by the acquisition unit 120 a .

[0089] Designation accepting unit 160 causes a display device to display a plurality of words obtained by performing morphological analysis on the character string data acquired by acquiring unit 120a. Here, the plurality of words are words with a part of speech such as verb, noun, or adjective. In other words, the plurality of words can be words of any part of speech. Furthermore, the word serving as the predicate is a verb, adjective, adjective-verb, or a word with a noun-like form.

[0090] The display device is, for example, a display that can be connected to the information processing device 100a. The user specifies a target word from among the multiple words displayed on the display device. The designation accepting unit 160 accepts the designation of the target word. The target word is also referred to as the first word.

[0091] The inference unit 170 infers the meaning of the target word using the learned model.

[0092] The output unit 180 outputs the inference result. In addition, the output unit 180 outputs the likelihood of the meaning of the target word.

[0093] Next, the processing performed by the information processing apparatus 100a will be described using a flowchart.

[0094] Figure 10 It is a flowchart showing an example of the processing performed by the information processing apparatus in the operation stage of Embodiment 1.

[0095] (Step S31) The acquisition unit 120a acquires string data. For example, the string data represents "車のドライバー".

[0096] (Step S32) The morpheme analysis execution unit 130a performs morpheme analysis on the string data. For example, the morpheme analysis execution unit 130a divides the string data "車のドライバー" into "車(名词)", "の(助词)", and "ドライバー(名词)".

[0097] (Step S33) The designation reception unit 160 causes the display device to display the words S ("車(名词)", "ドライバー(名词)") obtained by removing particles, auxiliary verbs, and symbols from the divided words.

[0098] The user designates a target word from the words displayed on the display device. Assume w i "ドライバー(名词)" is designated as the target word.

[0099] (Step S34) The designation reception unit 160 receives the designated target word.

[0100] (Step S35) The inference unit 170 performs an inference process for inferring the meaning of the target word using the learned model. The process will be described in detail. In the description, Figure 6 . For example, the inference unit 170 refers to the row of w i "ドライバー(名词)" in the learned model. The inference unit 170 determines whether the word S ("車(名词)", "ドライバー(名词)") exists in the row of "ドライバー(名词)". When the word S ("車(名词)", "ドライバー(名词)") exists in the row of "ドライバー(名词)", the inference unit 170 calculates the count value C using Equation (3) wi . In addition, the count value C wi is the sum of m i，j .

[0101] [Equation 3]

[0102]

[0103] For example, when the predicate label is "運転する(verb)", the inference unit 170 calculates the result of Equation (4).

[0104] [Equation 4]

[0105] C wi = 159 = w ドライバ一，車 …(4)

[0106] The inference unit 170 also calculates the count value C for all predicate labels. wi . The inference unit 170 calculates the maximum value from the count values C of each predicate label. wi The maximum value is the likelihood Score wi of the meaning. The likelihood Score is calculated using Equation (5). wi .

[0107] [Equation 5]

[0108]

[0109] Thus, the inference unit 170 calculates the result of Equation (6).

[0110] [Equation 6]

[0111] <00**********301><00**********302>The predicate label with the maximum value is the meaning Label wi . The meaning Label wi is represented by Equation (7) below. Additionally, for example, the meaning Label wi is "運転する(verb)".

[0113] [Equation 7]

[0114]

[0115] In this way, the inference unit 17**0 infers the meaning of the target word based on the multiple words obtained by performing morphological analysis on the string data acquired by the acquisition unit 120a and the learned model. For example, the inference unit 170 infers the meaning of "ドライバー(noun)" based on "車(noun)", "ドライバー(noun)" obtained by performing morphological analysis on "車のドライバー" and the learned model.

[0116] (Step S36) The output unit 180 outputs the inference result. For example, the output unit 180 outputs "運転する(verb)" as the meaning Label wi . In other words, the output unit 180 outputs the meaning label "運転する(verb)". Additionally, the output unit 180 outputs the likelihood Score wi .

[0117] In this way, the information processing device 100a can infer that the meaning of "driver" in the string data "車のドライバー" is "運転するドライバー" (driver who drives). In addition, even when the string data does not contain a predicate, the information processing device 100a can still infer the meaning.

[0118] In addition, for example, the information processing device 100a obtains the string data "工具のドライバー" (tool driver). When the target word in the string data "工具のドライバー" is "ドライバー", the information processing device 100a infers that the meaning of "ドライバー" is "作業するドライバー" (driver who works). In addition, the information processing device 100a outputs information indicating that the likelihood Score wi is 9.

[0119] In addition, for example, the information processing device 100a obtains the string data "ゴルフのドライバー" (golf driver). When the target word in the string data "ゴルフのドライバー" is "ドライバー", the information processing device 100a infers that the meaning of "ドライバー" is "飛ばすドライバー" (driver that hits). In addition, the information processing device 100a outputs information indicating that the likelihood Score wi is 6.

[0120] In this way, the information processing device 100a can use the learned model to infer different meanings for the same word according to the context. In addition, the information processing device 100a outputs a meaning label. As a result, the user can intuitively understand the meaning label. And the information processing device 100a outputs a likelihood. As a result, the user can know the likelihood of the meaning of the target word.

[0121] Here, a comparative example is shown.

[0122] Figure 11 is a diagram showing a comparative example. Figure 11 Shows 1 sub-word context matrix. When a certain device uses Figure 11 the sub-word context matrix as the learned model to infer the meaning, since the sub-word context matrix of this device is 1, only 1 meaning is inferred.

[0123] On the other hand, the learned model generated by the information processing device 100 includes a plurality of sub-word context matrices. Therefore, the learned model generated by the information processing device 100 can correspond to polysemy. For example, as described above, the information processing device 100a can use the learned model to infer different meanings for the same word according to the context. In addition, when the information processing device 100 generates a learned model, it does not perform dimensional compression. Moreover, the information processing device 100 performs low-cost calculations such as repeatedly counting the sub-learning examples. Therefore, the information processing device 100 can generate a learned model in a short processing time. Therefore, according to Embodiment 1, the information processing device 100 can generate a learned model corresponding to polysemy in a short processing time.

[0124] Embodiment 2

[0125] <Operation phase>

[0126] Next, Embodiment 2 will be described. In Embodiment 2, the matters different from Embodiment 1 will be mainly described. Moreover, in Embodiment 2, the description of the same matters as those in Embodiment 1 will be omitted. In the description of Embodiment 2, reference is made to Figures 1 to 10 .

[0127] Figure 12 is a diagram showing the functional blocks of the information processing device in the operation phase of Embodiment 2. The same Figure 9 structure as that shown in Figure 12 is labeled with the same reference numerals as those shown in Figure 9 .

[0128] The information processing device 100a further includes a determination unit 190. The acquisition unit 120a further acquires semantic sentence data. The semantic sentence data is a plurality of strings representing semantics, that is, a plurality of semantic strings. In other words, the semantic sentence data is a plurality of strings representing the semantics corresponding to the target word, that is, a plurality of semantic strings. For example, the semantic sentence data includes "the driver of the car", "the driver of the tool", and "the driver of golf". In other words, the semantic sentence data includes "the driver of the car", "the driver of the tool", and "the driver of golf" corresponding to "driver".

[0129] The determination unit 190 determines, from the plurality of semantic strings, the semantic string having the same meaning as the meaning of the target word. The determined semantic string is called the correct semantic string. The output unit 180 outputs the correct semantic string.

[0130] Next, the processing performed by the information processing device 100a will be described using a flowchart. <**********349**********> Figure 13It is a flowchart showing an example of the processing executed by the information processing apparatus in the operation stage of Embodiment 2.

[0132] In Figure 13 the processing of, the difference from the processing of Figure 10 is that steps S31a and 35a are executed, and step S36a is executed instead of step S36. Therefore, in Figure 13 , steps S31a, 35a, and 36a are described. Regarding Figure 13 the other steps in, by labeling the same numbers as the step numbers in Figure 10 , the description of the processing is omitted.

[0133] (Step S31a) Acquisition unit 120a acquires semantic sentence data. For example, the semantic sentence data includes "車のドライバー" (driver of a car), "工具のドライバー" (driver of a tool), and "ゴルフのドライバー" (golf driver). Acquisition unit 120a stores the semantic sentence data in storage unit 110.

[0134] Here, in step S31, acquisition unit 120a acquires the string data "タクシードライバー" (taxi driver). In step S35, inference unit 170 infers that the meaning of "タクシードライバー" is "運転するドライバー" (driving driver). The likelihood is 15.

[0135] (Step S35a) Determination unit 190 acquires semantic sentence data via acquisition unit 120a. Here, inference unit 170 infers that the meaning of "タクシードライバー" is "運転するドライバー". Determination unit 190 determines the semantic having the same meaning as "運転するドライバー" from the semantic sentence data. For example, "車" (car) has a relationality with "運転する" (driving). Determination unit 190 determines that "車のドライバー" has the same meaning as "運転するドライバー" based on the information indicating the relationship between "車" and "運転する". Then, determination unit 190 determines "車のドライバー" from the semantic sentence data.

[0136] (Step S36a) Output unit 180 outputs the correct semantic string. For example, output unit 180 outputs the case where the correct semantic string of "ドライバー" in "タクシードライバー" is "車のドライバー".

[0137] In addition, for example, when acquisition unit 120a acquires the string data "精密ドライバー" (precision driver), determination unit 190 determines "工具のドライバー" from the semantic sentence data. Then, output unit 180 outputs the case where the correct semantic string of "ドライバー" in "精密ドライバー" is "工具のドライバー".

[0138] According to Embodiment 2, information processing apparatus 100a uses semantic sentence data, thereby being able to improve the likelihood of meaning.

[0139] In each embodiment described above, the case of inferring the meaning of a noun is described as an example, but it is also possible to infer parts of speech other than nouns in the same manner. In addition, the features of each embodiment can be combined with each other as appropriate.

[0140] Description of labels

[0141] 100, 100a: information processing device; 101: processor; 102: volatile storage device; 103: non-volatile storage device; 110: storage unit; 120, 120a: acquisition unit; 130, 130a: morpheme analysis execution unit; 140: predicate segmentation unit; 150: generation unit; 160: designation acceptance unit; 170: inference unit; 180: output unit; 190: determination unit.

Claims

1. An information processing device, wherein: The information processing device has: an acquisition unit for acquiring a plurality of character strings; a morpheme analysis execution unit that performs morpheme analysis on the plurality of character strings; as well as a generating unit configured to generate inference information for inferring the meaning of a first word among the plurality of words with parts of speech, based on a plurality of words serving as a plurality of predicates among the plurality of words with parts of speech obtained by performing morphological analysis on the plurality of character strings and a plurality of words among the plurality of words with parts of speech, The inference information is information indicating a correspondence relationship between each of the plurality of words serving as the plurality of predicates and each of the plurality of relationship information. The relationship information is information indicating the relationship between two nouns, and is composed of a subword context matrix indicating the relationship between each of the plurality of words with part of speech. The generating unit counts sub-learning example sentences with the same predicate, generates the sum of the number of times two nouns appear together in the sub-learning example sentences or the self-mutual information of the two nouns, and uses the multiple relationship information with the sum or the self-mutual information as elements of the matrix and marked with multiple predicate labels representing multiple predicates to generate the inference information, wherein the sub-learning example sentences are multiple nouns in a character string containing the multiple predicates.

2. The information processing device according to claim 1, wherein The generating unit generates the inference information based on a plurality of words that are a plurality of verbs among the plurality of part-of-speech words and a plurality of words among the plurality of part-of-speech words.

3. The information processing device according to claim 1, wherein The plurality of character strings are character strings of question sentences and character strings of response sentences to the question sentences.

4. The information processing device according to claim 2, wherein: The plurality of character strings are character strings of question sentences and character strings of response sentences to the question sentences.

5. The information processing device according to any one of claims 1 to 4, wherein: The information processing device further includes a designation receiving unit and an inference unit. The acquiring unit acquires character string data that is a character string including the first word, The morphological analysis execution unit performs morphological analysis on the character string data. The designation accepting unit accepts designation of the first word among a plurality of words obtained by executing morphological analysis on the character string data, The inference unit infers the meaning of the first word based on a plurality of words obtained by performing morphological analysis on the character string data and the inference information. The information processing apparatus according to claim 5 , wherein: The information processing device further includes an output unit that outputs the likelihood of the meaning of the first word.

7. The information processing apparatus according to claim 6, wherein: The information processing device further includes a determination unit and an output unit. The acquiring unit acquires a plurality of semantic character strings as a plurality of character strings representing semantics, The specifying unit specifies a semantic character string having the same meaning as the inferred first word from among the plurality of semantic character strings. The output unit outputs the determined semantic character string.

8. A method of generating, wherein: The information processing device obtains a plurality of character strings, The information processing device performs morpheme analysis on the plurality of character strings, The information processing device generates inference information for inferring the meaning of a first word among the plurality of words with parts of speech, based on a plurality of words serving as a plurality of predicates among the plurality of words with parts of speech obtained by performing morphological analysis on the plurality of character strings and a plurality of words among the plurality of words with parts of speech, wherein the inference information is used to infer the meaning of a first word among the plurality of words with parts of speech. The inference information is information indicating a correspondence relationship between each of the plurality of words serving as the plurality of predicates and each of the plurality of relationship information. The relationship information is information indicating the relationship between two nouns, and is composed of a subword context matrix indicating the relationship between each of the plurality of words with part of speech. The information processing device counts sub-learning example sentences with the same predicate, generates the sum of the number of times two nouns appear together in the sub-learning example sentences or the self-mutual information of the two nouns, and uses multiple relationship information with the sum or the self-mutual information as elements of the matrix and marked with multiple predicate labels representing multiple predicates to generate the inference information, wherein the sub-learning example sentences are multiple nouns in a character string containing the multiple predicates.

9. A recording medium having a generation program recorded thereon, wherein: This generation program causes the information processing device to execute the following processing: Get multiple strings; performing morpheme analysis on the plurality of character strings; as well as generating, based on a plurality of words serving as predicates among a plurality of part-of-speech words obtained by performing morphological analysis on the plurality of character strings and a plurality of words among the plurality of part-of-speech words, inference information for inferring the meaning of a first word among the plurality of words among the plurality of part-of-speech words; The inference information is information indicating a correspondence relationship between each of the plurality of words serving as the plurality of predicates and each of the plurality of relationship information. The relationship information is information indicating the relationship between two nouns, and is composed of a subword context matrix indicating the relationship between each of the plurality of words with part of speech. In which, by counting sub-learning example sentences with the same predicate, the sum of the number of times two nouns appear together in the sub-learning example sentences or the self-mutual information of the two nouns is generated, and the inference information is generated using multiple relationship information with the sum or the self-mutual information as elements of the matrix and marked with multiple predicate labels representing multiple predicates, wherein the sub-learning example sentences are multiple nouns in a character string containing the multiple predicates.

Images