Information processing device, information processing method, and program

Abstract

The present invention solves the problem of generating, from a logical equation representing an inference result, a more diverse natural language sentence as compared to a document used to generate inference rules. In order to achieve the foregoing, this information processing device (1) includes: an acquisition unit (11) that acquires a logical equation representing an inference result based on inference rules for an observation event; a generation unit (12) that generates a natural language sentence from text elements that are text from among the constituent elements of the logical equation, by using a language model generated on the basis of a corpus that does not contain documents used to generate the inference rules; and an output unit (13) that outputs the natural language sentence.

Description

[Technical field]

[0001] This relates to a technology for generating natural language sentences from logical expressions that represent inference results based on inference rules. [Background technology]

[0002] There is a known technology for generating natural language sentences from logical formulas that represent inference results based on inference rules. The logical formulas that represent inference results contain information that indicates the applied rules in addition to information that indicates the hypotheses derived by inference. This makes it easy to understand how the hypotheses were derived, but it is difficult to understand intuitively because they are expressed as logical formulas. For this reason, there are cases in which natural language sentences are generated from logical formulas that represent inference results and the generated natural language sentences are output.

[0003] Related techniques applicable to such cases include those described in Patent Documents 1 and 2. Patent Document 1 describes a technique for extracting the meaning of an input natural language sentence in the form of a logical expression, and converting the extraction result back into a natural language sentence and displaying it. This technique refers to natural language expression data associated with a predicate in order to generate a natural language sentence from the extraction result.

[0004] Patent Document 2 describes a technique for converting a text input sentence into an output sentence based on a symbolic logic level sentence using a synonymous conditional statement or conditional statement corresponding to the text input sentence, and outputting a text sentence obtained by inversely converting the output sentence. In this technique, the synonymous conditional statement or conditional statement used when converting the text input sentence into the output sentence is used to inversely convert the output sentence. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 5-81316 [Patent Document 2] Japanese Patent Publication No. 2010-170162 Summary of the Invention [Problem to be solved by the invention]

[0006] Here, when generating natural language sentences from a logical formula that represents an inference result, there are cases where it is desirable to generate natural language sentences with more diversity than the document used to generate the inference rules. A specific example of such a case is a case where consideration is given to the leakage of sensitive information contained in the document. Another specific example of such a case is a case where it is desirable to generate natural language sentences that are simpler than the document.

[0007] However, the technology described in Patent Document 1 cannot necessarily generate natural language sentences with more diversity than the original input natural language sentences even if it refers to natural language expression data associated with a predicate. Also, the technology described in Patent Document 2 cannot necessarily output text sentences with more diversity than the original text input sentences even if it performs reverse conversion using a synonymous conditional statement or conditional statement used for conversion. Therefore, when the technologies described in Patent Documents 1 and 2 are used in a process of generating natural language sentences from a logical formula representing an inference result, there is room for improvement in generating natural language sentences with more diversity than the documents used for generating inference rules.

[0008] One aspect of the present invention has been made in consideration of the above-mentioned problems, and one example of its objective is to provide a technology for generating natural language sentences that are more diverse than the documents used to generate inference rules from a logical formula that represents an inference result. [Means for solving the problem]

[0009] An information processing device according to one aspect of the present invention includes an acquisition means for acquiring a logical formula that represents an inference result based on inference rules for an observed event, a generation means for generating a natural language sentence from a text element, which is text among the components of the logical formula, using a language model generated based on a corpus that does not include documents used to generate the inference rules, and an output means for outputting the natural language sentence.

[0010] An information processing method according to one aspect of the present invention includes obtaining a logical formula that represents an inference result based on inference rules for an observed event, generating a natural language sentence from text elements, which are text among the components of the logical formula, using a language model generated based on a corpus that does not include documents used to generate the inference rules, and outputting the natural language sentence.

[0011] A program according to one aspect of the present invention is a program for causing a computer to function as an information processing device, and causes the computer to function as: an acquisition means for acquiring a logical formula representing an inference result based on inference rules for an observed event; a generation means for generating a natural language sentence from a text element, which is text among the components of the logical formula, using a language model generated based on a corpus that does not include documents used to generate the inference rules; and an output means for outputting the natural language sentence. Effect of the Invention

[0012] According to one aspect of the present invention, it is possible to generate natural language sentences with greater variety from a logical formula representing an inference result, as compared with documents used to generate inference rules. [Brief description of the drawings]

[0013] [Figure 1] 1 is a block diagram showing a configuration of an information processing device according to a first exemplary embodiment of the present invention. [Diagram 2] 1 is a flow chart showing the flow of an information processing method according to an exemplary embodiment 1 of the present invention. [Diagram 3] FIG. 11 is a block diagram showing a configuration of an information processing device according to a second exemplary embodiment of the present invention. [Figure 4] 11 is a flow chart showing the flow of an information processing method according to an exemplary embodiment 2 of the present invention. [Diagram 5] FIG. 11 is a flow chart showing the flow of another information processing method according to the second exemplary embodiment of the present invention. [Figure 6]FIG. 11 is a diagram showing a specific example of a logical expression representing an inference result in exemplary embodiment 2 of the present invention. [Figure 7] FIG. 11 is a flow chart showing a detailed flow of a process for dividing into a plurality of groups in the second exemplary embodiment of the present invention. [Figure 8] FIG. 11 is a flow chart showing a detailed flow of a process for generating candidates for natural language sentences in exemplary embodiment 2 of the present invention. [Figure 9] 11 is a flow chart showing a detailed flow of a process for eliminating candidates of natural language sentences similar to an original sentence in the exemplary embodiment 2 of the present invention. FIG. [Figure 10] 11 is a flow chart showing a detailed flow of a process for eliminating natural language sentence candidates whose contents do not match the original sentence in the second exemplary embodiment of the present invention. FIG. [Figure 11] FIG. 1 is a block diagram showing a configuration of a computer that functions as an information processing device according to each exemplary embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Example Embodiment 1 A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described later.

[0015] <Configuration of information processing device> The configuration of an information processing device 1 according to this exemplary embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1. As shown in Fig. 1, the information processing device 1 includes an acquisition unit 11, a generation unit 12, and an output unit 13.

[0016] The acquisition unit 11 acquires a logical formula representing an inference result based on an inference rule for an observed event. The generation unit 12 generates a natural language sentence from a text element, which is a text among the components of the logical formula, using a language model generated based on a corpus that does not include documents used to generate the inference rules. The output unit 13 outputs the natural language sentence. The acquisition unit 11 is an example of a configuration that realizes the acquisition means described in the claims. The generation unit 12 is an example of a configuration that realizes the generation means described in the claims. The output unit 13 is an example of a configuration that realizes the output means described in the claims.

[0017] (Inference rules) The inference rules are generated from a document including background knowledge. The document includes, for example, contents related to the field to be inferred. The inference rules include one or more rules. The inference results based on the inference rules are generated by an inference engine. The inference engine receives an observed event as input, performs inference using one or more rules among the inference rules that are applicable to the input observed event, generates a hypothesis, and outputs a logical formula representing the inference result. The logical formula representing the inference result is, for example, a first-order predicate logical formula, but is not limited to this. The logical formula representing the inference result includes information indicating the hypothesis and information indicating the rule used to derive the hypothesis. The inference rules and the inference engine may be stored in a memory of the information processing device 1, or may be stored in another device capable of communicating with the information processing device 1.

[0018] (Language Model) A language model is a model that outputs the probability of an input word string as a natural language sentence. In addition, by using such a language model, it is also possible to predict a word that complements an input word string. A word that complements a word string is a word that can become a natural language sentence after complementing a word string with the word. Such a language model is generated by using a corpus. The corpus does not include the above-mentioned document used in generating the inference rule. In other words, all documents included in the corpus used in generating the language model are documents that have not been used in generating the inference rule. In addition, it is preferable that an algorithm used in generating the language model is an algorithm that does not require teacher data. Examples of models that can be generated by such an algorithm include, but are not limited to, an N-gram language model, a hidden Markov model (HMM), a conditional random fields (CRF), and a long short-term memory (LSTM). The language model may be stored in a memory of the information processing device 1, or may be stored in another device that can communicate with the information processing device 1.

[0019] <Information processing method flow> The flow of the information processing method S1 executed by the information processing device 1 configured as above will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of the information processing method S1. As shown in Fig. 2, the information processing method S1 includes steps S101 to S103.

[0020] (Step S101: Acquisition process) In step S101, the acquisition unit 11 acquires a logical formula representing an inference result based on an inference rule for an observed event. Specifically, the acquisition unit 11 inputs an observed event to an inference engine and acquires a logical formula output from the inference engine. For example, the acquisition unit 11 may acquire such a logical formula from another device capable of communicating with the information processing device 1, or may acquire a logical formula stored in a memory of the information processing device 1.

[0021] (Step S102: Generation process) In step S102, the generation unit 12 uses the above-mentioned language model to generate a natural language sentence from text elements, which are text among the components of the logical formula. For example, the generation unit 12 generates a word that supplements a string of multiple text elements by inputting multiple text elements included in the components of the logical formula to the language model. The generation unit 12 also supplements the string of multiple text elements with the generated word to generate a natural language sentence.

[0022] (Step S103: Output process) In step S103, the output unit 13 outputs the generated natural language sentence. For example, the output unit 13 may output the natural language sentence to an output device such as a display or a speaker connected to the information processing device 1. When outputting to a speaker, the output unit 13 generates and outputs voice data from the generated natural language sentence. The output unit 13 may also output the generated natural language sentence by writing it in a memory of the information processing device 1. The output unit 13 may also output the natural language sentence by transmitting it to another device connected to the information processing device 1 so as to be able to communicate with it.

[0023] <Advantages of this exemplary embodiment> As described above, the information processing device 1 and information processing method S1 according to this exemplary embodiment are configured to obtain a logical formula that represents an inference result based on inference rules for an observed event, generate a natural language sentence from a text element, which is text among the components of the logical formula, using a language model generated based on a corpus that does not include documents used to generate the inference rules, and output the generated natural language sentence.

[0024] In this manner, the language model used in this exemplary embodiment is generated based on a corpus that does not include documents used to generate the inference rules. Therefore, natural language sentences generated using such a language model will include expressions contained in such a corpus, and are expected to include expressions different from those contained in the documents used to generate the inference rules. As a result, an effect is obtained in which natural language sentences with more diversity are generated from the logical formula representing the inference result compared to the documents used to generate the inference rules.

[0025] Exemplary embodiment 2 <Configuration of information processing device> The configuration of the information processing device 2 according to this exemplary embodiment will be described with reference to Fig. 3. Fig. 3 is a block diagram showing the configuration of the information processing device 2. As shown in Fig. 3, the information processing device 2 includes an acquisition unit 21, a generation unit 22, an output unit 23, a first determination unit 24, a second determination unit 25, a language model generation unit 26, and a language model L.

[0026] A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first exemplary embodiment are given the same reference numerals, and the description thereof will be omitted as appropriate.

[0027] The acquisition unit 21 acquires a logical formula representing an inference result based on an inference rule R for an observed event. The inference rule R includes rules r1, r2, r3, .... Details of the inference rule R are the same as those of the inference rule described in the exemplary embodiment 1. In this exemplary embodiment, a document used to generate the inference rule R is described as an original text D1. The original text D1 includes documents d11, d12, d13, ... related to a specific field. Note that, although it is described here that the original text D1 includes multiple documents, the original text D1 may include only one document. The specific field is, for example, the medical field or the cybersecurity field, but is not limited thereto.

[0028] The generation unit 22 uses a language model L to generate a natural language sentence from a text element, which is a text among the components of a logical formula. The language model L does not include the original text D1 and is generated using a corpus D2 in the same field as the original text D1. In other words, the corpus D2 includes documents d21, d22, d23, ... related to the same specific field as each document included in the original text D1. Moreover, none of the documents d21, d22, d32, ... are included in the original text D1. Details of the natural language sentence generation process will be described later.

[0029] The language model generation unit 26 uses the corpus D2 to generate the language model L. For example, the language model generation unit 26 may extract N-grams from the corpus D2 to generate an N-gram language model. The process of generating the language model L will be described later in detail.

[0030] The first determination unit 24 refers to a comparison result between the natural language sentence generated by the generation unit 22 and the original text D1, and determines whether or not the natural language sentence is similar to the original text D1. For example, the natural language sentence being similar to the original text D1 may mean that a sentence matching the natural language sentence is included in the original text D1. Furthermore, the original text D1 referred to in order to determine whether the natural language sentence is similar may be the entire original text D1, or may be a part of it. Details of the determination process will be described later.

[0031] The second determination unit 25 refers to a comparison result between the natural language sentence generated by the generation unit 22 and the original text D1 and determines whether the natural language sentence matches the content of the original text D1. For example, the natural language sentence matching the content of the original text D1 may mean that the natural language sentence does not have the opposite content to that of the original text D1. Details of the determination process will be described later.

[0032] The output unit 23 outputs the natural language sentence when the first determination unit 24 determines that the natural language sentence generated by the generation unit 22 is not similar. Moreover, the output unit 23 outputs the natural language sentence when the second determination unit 25 determines that the natural language sentence generated by the generation unit 22 is consistent. In this exemplary embodiment, the output unit 23 outputs the natural language sentence when the first determination unit 24 determines that the natural language sentence generated by the generation unit 22 is not similar and the second determination unit 25 determines that the natural language sentence is consistent. However, this is not limited to the above, and the output unit 23 may output the natural language sentence when the first determination unit 24 determines that the natural language sentence generated by the generation unit 22 is not similar or the second determination unit 25 determines that the natural language sentence is consistent.

[0033] <Information processing method flow: language model generation> First, the flow of the information processing method S2 in which the information processing device 2 generates a language model L will be described with reference to Fig. 4. Note that, although an example in which the language model L is an N-gram language model will be described here, the type of model constituting the language model L is not limited. Fig. 4 is a flow diagram showing the flow of the information processing method S2 in which the language model L is generated. As shown in Fig. 4, the information processing method S2 includes steps S201 to S203.

[0034] (Step S201) In step S201, the language model generation unit 26 acquires a corpus D2.

[0035] (Step S202) In step S202, the language model generation unit 26 extracts N-grams from documents d21, d22, d23, ... included in the corpus D2. The unit of the extracted N-grams is, for example, but is not limited to, a word unit. The value of N is, for example, 3, but is not limited to this.

[0036] (Step S203) In step S203, the language model generation unit 26 learns a language model L using the extracted N-grams. For example, the language model generation unit 26 calculates the consecutive occurrence probability that words included in the extracted N-grams appear consecutively, and generates the language model L based on the statistical result of the consecutive occurrence probability. The language model generation unit 26 may also apply various known methods for improving the accuracy of the N-gram language model. The language model generation unit 26 stores the generated language model L in the memory of the information processing device 2.

[0037] <Information processing method flow: Generation of natural language sentences> Next, the flow of the information processing method S3 in which the information processing device 2 generates a natural language sentence will be described with reference to Fig. 5. Fig. 5 is a flow diagram showing the flow of the information processing method S3 in which a natural language sentence is generated. As shown in Fig. 5, the information processing method S3 includes steps S301 to S306.

[0038] (Step S301) In step S301, the acquisition unit 21 acquires a logical formula representing an inference result. Specifically, the acquisition unit 21 inputs an observed event to an inference engine and acquires a logical formula output from the inference engine. For example, the acquisition unit 21 may acquire such a logical formula from another device capable of communicating with the information processing device 2, or may acquire a logical formula stored in a memory of the information processing device 2.

[0039] (Specific examples of logical expressions) A specific example of a logical expression representing an inference result will be described with reference to Fig. 6. Fig. 6 is a diagram showing a specific example of a logical expression representing an inference result.

[0040] As shown in FIG. 6, the logical formula representing the inference result is expressed in a graph structure and has nodes N1 to N6 and edges E1 to E5. Nodes N1 to N3 indicate observed events input to the inference engine. Edges E1 to E5 indicate rules applied to the observed events indicated by nodes N1 to N3. Here, edges E1 to E5 indicate rules r1 to r5. Nodes N4 to N6 indicate hypotheses inferred by applying rules r1 to r5 to the observed events indicated by nodes N1 to N3. In other words, in this example, the logical formula representing the inference result includes information indicating the input observed events (nodes N1 to N3), information indicating hypotheses derived by inference (nodes N4 to N6), and information indicating rules used to derive the hypotheses (edges E1 to E5).

[0041] In this example, node N1 is an observed event represented by a logical formula "body part (e1, "pre-chapter")" and indicates that the object of observation in event "e1" is the body part "pre-chapter". Here, an event is an attribute of an observed event, such as the observation location, the observer, the timing of the observation, or a document in which the observed event is described, and is assigned to the observed event in advance. Node N2 is an observed event represented by a logical formula "symptom (e1, "bleeding")" = "symptom (e3, "bleeding"), e1 = e3", indicating that the bleeding symptom was observed in events "e1" and "e3", and that event "e1" is equal to event "e3". Node N3 is an observed event represented by a logical formula "symptom (e2, "language disorder")" and indicates that the language disorder symptom was observed in event "e2".

[0042] In this example, node N4 indicates a hypothesis generated by applying rules r1 and r2 indicated by edges E1 and E2 to the observed events indicated by nodes N1 and N2. Node N4 is expressed by the logical formula "body part (e3, "prefrontal cortex")^!function (e3, x)" and indicates the hypothesis that the body part "prefrontal cortex" does not have the function indicated by variable x in event "e3". Node N5 indicates a hypothesis generated by applying rule r3 indicated by edge E3 to the observed event indicated by node N3. Node N5 is expressed by the logical formula "!function (e4, "language function")" and indicates the hypothesis that the body part "prefrontal cortex" does not have the language function in event "e4". Node N6 indicates a hypothesis generated by applying rules r4 and r5 indicated by edges E4 and E5 to the hypotheses indicated by nodes N4 and N5. Node N6 is expressed by the logical formula "x="language function"," and indicates the hypothesis that the variable "x" in the hypothesis indicated by node N4 is a "language function."

[0043] Also, in this example, rule r1 indicated by edge E1 is associated with a portion of original text D1 that is the basis for rule r1, "document d11: p10, L1." In other words, rule r1 was generated by referring to the contents written in that portion. Hereinafter, the portion of original text D1 that is the basis for each rule will also be referred to as basis information. Rule r2 is associated with basis information "document d11: P10, L2." Rule r3 is associated with basis information "document d12: P30, L5."

[0044] In this way, the logical formula expressed in the graph structure shown in Fig. 4 indicates the inference result that "if symptoms of bleeding from the prefrontal cortex and language disorder are observed, the hypothesis that the prefrontal cortex does not have language function is inferred," but it is difficult to understand intuitively. In this exemplary embodiment, by executing the following steps S302 to S306, natural language sentences with more diversity compared to the original sentence D1 are generated and output from such a logical formula.

[0045] (Step S302) In step S302, the generation unit 22 divides a plurality of text elements included in the logical formula expressing the inference result into a plurality of groups. Note that here, it is assumed that the logical formula expressing the inference result includes a plurality of text elements. Details of this step will be described later. Thereafter, the information processing device 2 executes steps S303 to S305 for each divided group. Note that, when the logical formula expressing the inference result includes one text element, it is sufficient to execute steps S303 to S305 assuming that there is one group including that one text element.

[0046] (Step S303) In step S303, the generation unit 22 generates candidates for natural language sentences from one or more text elements included in the group, using the language model L. The number of natural language sentence candidates generated for each group may be one or more. There may also be a group in which the number of natural language sentence candidates is zero. Details of this step will be described later.

[0047] (Step S304) In step S304, the generation unit 22 uses the first determination unit 24 to exclude candidates of the natural language sentence generated in step S303 that are similar to the original sentence D1. The remaining candidates are not similar to the original sentence D1. Details of this step will be described later.

[0048] (Step S305) In step S304, the generation unit 22 uses the second judgment unit 25 to remove candidates of natural language sentences that are not similar to the original text D1 and remain after the process in step S304, and that are not consistent with the content of the original text D1. The remaining candidates are not similar to the original text D1 and are consistent with the content of the original text D1. Details of this step will be described later.

[0049] After executing the processes of steps S303 to S305 for each group, the information processing device 2 executes the next step S306.

[0050] (Step S306) In step S306, the output unit 23 outputs the candidates remaining after the processes in steps S304 and S305 for each group as natural language sentences. As a result, from the logical formulas representing the inference results, those that are not similar to the original text D1 and whose contents are consistent with the original text D1 are output. Specific examples of the output destination are as described in the first exemplary embodiment.

[0051] (Details of the process of dividing into groups) Details of the process of dividing a plurality of text elements into a plurality of groups in step S302 will be described with reference to Fig. 7. Fig. 7 is a flow chart showing a detailed flow of the process of dividing into a plurality of groups (step S302). As shown in Fig. 7, step S302 includes steps S401 to S402.

[0052] (Step S401) In step S401, the generation unit 22 extracts a plurality of text elements from a logical formula representing an inference result. For example, the generation unit 22 may extract a plurality of text elements from a logical formula indicating an observed event. Specifically, in the case of the specific example of the logical formula shown in FIG. 4, the generation unit 22 extracts the text elements "previous chapter", "bleeding", and "language disorder" from the logical formula indicated by the nodes N1 to N3 indicating the observed event.

[0053] (Step S402) In step S402, the generating unit 22 divides the multiple text elements into multiple groups based on the attributes of the observed events related to each text element. For example, each group includes one or multiple text elements with the same attribute of the observed event. Here, the attribute of the observed event is, as described above, an observer, an observation location, a timing of the observation, or a document in which the observed event is described, but is not limited to this. In the specific example of FIG. 6, the attributes of the observed event are represented by events "e1" to "e4". For example, the attribute of the observed event related to the text element "previous chapter" extracted in step S401 is the event "e1". This is because the logical expression indicated by the node N1 from which the text element "previous chapter" is extracted includes the event "e1". Similarly, the attributes of the observed event related to the text element "bleeding" are the events "e1" and "e3". Also, the attribute of the observed event related to the text element "language disorder" is the event "e2". The generation unit 22 divides these text elements into a group related to the event "e1" and a group related to the event "e2". In this specific example, since the event "e1" is equal to the event "e3", the generation unit 22 does not generate a group related to the event "e3". The group related to the event "e1" includes the text elements "Previous Chapter" and "Bleeding". The group related to the event "e2" includes the text element "Speech disorder".

[0054] (Details of the process of generating natural language sentence candidates) Details of the process of generating candidates for natural language sentences in step S303 will be described with reference to Fig. 8. Fig. 8 is a flow diagram showing a detailed flow of the process of generating candidates for natural language sentences (step S303). The process shown in Fig. 8 is executed for each of the multiple groups generated in step S302. As shown in Fig. 8, step S303 includes steps S501 to S506.

[0055] (Step S501) In step S501, the generating unit 22 generates a string in which the text elements included in the target group are arranged in n different orders (n is an integer equal to or greater than 1).

[0056] For example, if a target group contains one text element, there is only one possible order for arranging the text elements. In the example of Figure 6, the group related to the event "e2" contains one text element, "language disorder." In this case, a string {"language disorder"} is generated in which the text element is arranged in one possible order.

[0057] Furthermore, for example, when a target group contains multiple text elements, there are multiple possible orders for arranging these text elements. For example, in the example of FIG. 6, the group related to the event "e1" contains two text elements, "Previous Chapter" and "Bleeding". In this case, a first sequence {"Previous Chapter", "Bleeding"} and a second sequence {"Bleeding", "Previous Chapter"} are generated, arranging these text elements in two different orders.

[0058] Thereafter, information processing device 2 executes steps S502 to S505 for each of the n sequences of text elements generated in step S501. Hereinafter, the sequence of text elements to be processed is referred to as the i-th sequence (i is an integer between 1 and n).

[0059] (Step S502) In step S502, the generation unit 22 inputs the i-th sequence into the language model L. For example, the generation unit 22 inputs the first sequence {"Previous chapter", "Bleeding"} of two possible sequences of orders for two text elements included in a group related to the event "e1" into the language model L.

[0060] (Step S503) In step S503, the generation unit 22 generates a word candidate that complements the i-th string with reference to the output from the language model L. The generation unit 22 may generate an empty candidate that indicates that the i-th string is to be concatenated without complementing a word. For example, the generation unit 22 generates an empty candidate as word candidate 1 for the first string {"previous chapter", "bleeding"}. The generation unit 22 also generates a word "from" that complements the space between these two text elements as word candidate 2. The generation unit 22 also generates a word "none" that complements the rear of these two text elements as word candidate 3. The generation unit 22 also generates other word candidates 4, 5, ....

[0061] (Step S504) In step S504, the generation unit 22 evaluates the probability of the generated word candidates. Evaluating the probability means, for example, selecting a candidate whose probability output from the language model L satisfies an evaluation condition. The evaluation condition may be, for example, a threshold value or more. For example, the generation unit 22 selects word candidates 1, 2, and 3 whose probabilities are equal to or greater than a threshold value from among the above-mentioned word candidates 1, 2, 3, 4, 5, ....

[0062] (Step S505) In step S505, the generation unit 22 generates natural language sentence candidates using the word candidates selected by the evaluation. For example, the generation unit 22 generates natural language sentence candidate 1 "Previous chapter bleeding" by filling in the gap between two text elements in the i-th string {"Previous chapter", "Bleeding"} with word candidate 1, which is a blank candidate. The generation unit 22 also generates natural language sentence candidate 2 "Previous chapter bleeding" by filling in the gap between these two text elements with word candidate 2 "From". The generation unit 22 also generates natural language sentence candidate 3 "No bleeding from previous chapter" by filling in the gap between these two text elements with word candidate 2 "From" and filling the gap behind it with word candidate 3 "None".

[0063] When the above-mentioned processes of steps S502 to S505 are completed for n sequences, the process of generating a natural language sentence for the group (step S303) ends. In the natural language sentence candidates generated in this way, the text elements included in the group appear in the i-th order. Here, when the group includes multiple text elements, any two of the n orders correspond to examples of the "first order" and the "second order" described in the claims. In other words, when the group includes multiple text elements, the generation unit 22 generates a first candidate for a natural language sentence in which multiple text elements appear in a first order and a second candidate for a natural language sentence in which multiple text elements appear in a second order different from the first order, and generates one or both of the first candidate and the second candidate as a natural language sentence. When one or both of the first candidate and the second candidate are not excluded by the process of step S304 or S305 described later, they are adopted as natural language sentences. Note that, when a group contains one text element, there is a possibility that only one natural language sentence candidate is generated from the text element contained in the group. Therefore, the generation unit 22 generates one or more natural language sentence candidates for each group. Furthermore, the generation unit 22 adds the one or more natural language sentence candidates generated for each group to a set of output sentences. However, some or all of the natural language sentence candidates added to the set of output sentences may be excluded from the set of output sentences by the processing of step S304 or S305 described below.

[0064] (Details of the process of excluding texts similar to the original text) Details of the process of excluding natural language sentence candidates similar to the original sentence D1 in step S304 will be described with reference to Fig. 9. Fig. 9 is a flow diagram showing a detailed flow of the process of excluding sentences similar to the original sentence D1 (step S304). In this example, an example will be described in which the condition that the original sentence D1 contains a sentence matching the natural language sentence candidate is used as the condition for determining whether or not there is similarity.

[0065] The process shown in Fig. 9 is executed for each of the candidates for the natural language sentence included in the set of output sentences. As shown in Fig. 9, step S304 includes steps S601 to S602.

[0066] (Step S601) In step S601, the first determination unit 24 determines whether or not a sentence matching a candidate of the natural language sentence to be determined is included in the original sentence D1. For example, the part of the original sentence D1 to be determined is a part indicated by the grounds information associated with the rule used to infer the inference result. In the example of FIG. 6, the grounds information "document d11: p10, L1", "document d11: P10, L2", and "document d12: P30, L5" are associated with the rules r1 to r3 used to infer the inference result. Therefore, the first determination unit 24 determines whether or not a sentence matching the candidate is included in the corresponding part of the original sentence D1 for each of the natural language sentence candidates 1 to 3.

[0067] For example, suppose that the relevant portion in original text D1 contains the content "Bleeding in previous chapter... (omitted)... presents speech disorder." In this case, the first determination unit 24 determines that candidate natural language sentence 1 "Bleeding in previous chapter" is included in the relevant portion in original text D1 (i.e., similar). The first determination unit 24 also determines that candidate natural language sentence 2 "Bleeding from previous chapter" is not included in the relevant portion in original text D1 (i.e., not similar). The first determination unit 24 also determines that candidate natural language sentence 3 "No bleeding from previous chapter" is not included in the relevant portion in original text D1 (i.e., not similar).

[0068] (Step S602) If it is determined in step S601 that the natural language sentence candidate is included in the original sentence D1 (Yes), the first determination unit 24 excludes the natural language sentence candidate from the set of output sentences. If it is determined in step S601 that the natural language sentence candidate is not included in the original sentence D1 (No), step S602 is not executed. For example, in the above example, natural language sentence candidate 1 is excluded from the set of output sentences, and natural language sentence candidates 2 and 3 remain.

[0069] (Details of the process of excluding content that is inconsistent with the original text) Details of the process of excluding natural language sentence candidates whose contents do not match the original text D1 in step S305 will be described with reference to Fig. 10. Fig. 10 is a flow diagram showing a detailed flow of the process (step S305) of excluding natural language sentences whose contents do not match the original text D1. Here, an example will be described in which a condition that refers to a comparison result of the number of negation words is used as a condition for determining whether the contents match the original text D1. The process shown in Fig. 10 is executed for each natural language sentence candidate included in the set of output sentences. As shown in Fig. 9, step S305 includes steps S701 to S702.

[0070] (Step S701) In step S701, the second judgment unit 25 refers to a comparison result between the number of negation words included in the candidate natural language sentence and the number of negation words included in the original sentence D1, and judges whether the candidate is consistent with the content of the original sentence D1. More specifically, for example, the second judgment unit 25 judges whether the numbers of these negation words match. For example, the part of the original sentence D1 to be compared for the number of negation words is a part indicated by the grounds information associated with the rule used to infer the inference result, as is the part of the original sentence D1 to be judged for similarity. The relevant part in the example of FIG. 6 is as described above.

[0071] For example, the number of negation words contained in original text D1 "Bleeding from the previous chapter... (omitted)... presents speech disorder." is zero. The number of negation words contained in natural language sentence candidate 2 "Bleeding from the previous chapter" is zero. The number of negation words contained in natural language sentence candidate 3 "No bleeding from the previous chapter" is one. Therefore, the second judgment unit 25 judges that the number of negation words in natural language sentence candidate 2 matches, i.e., it is consistent with the content of the corresponding part of original text D1. The second judgment unit 25 also judges that the number of negation words in natural language sentence candidate 3 does not match, i.e., it is not consistent with the content of the corresponding part of original text D1.

[0072] (Step S702) If it is determined in step S702 that the numbers of negated words do not match (No), the second determination unit 25 excludes the target natural language sentence candidate from the set of output sentences. If it is determined in step S701 that they match (Yes), step S702 is not executed. For example, in the above example, natural language sentence candidate 3 is further excluded from natural language sentence candidates 2 and 3 remaining in the set of output sentences, and candidate 2 remains.

[0073] As described above, steps S303 to S305 are executed for each group, and the natural language sentence candidates remaining in the set of output sentences are adopted and output in step S306.

[0074] (Specific examples) For instance, in the example in Figure 6, for the first sequence in the group related to event “e1” {“bleeding”, “previous chapter”}, as mentioned above, candidate natural language sentence 3 “no bleeding from previous chapter” remains in the set of output sentences.

[0075] In addition, in this example, for the second column in the group related to the event “e1”, {“bleeding”, “previous chapter”}, a candidate natural language sentence was generated in step S303 and added to the set of output sentences, but was excluded from the set of output sentences in steps S304 and S305.

[0076] Also, for the group related to the event "e2", step S303 is executed with only one text element {"language disorder"} included in the group as a column. As a result, natural language sentence candidate 4 "has a language disorder", candidate 5 "has a language disorder", and candidate 6 "has no language disorder" are generated and added to the set of output sentences. Also, step S304 is executed to exclude candidate 5 "has a language disorder" included in the corresponding part of the original text D1 from natural language sentence candidates 4 to 6. Also, step S305 is executed to exclude candidate 6 "has no language disorder" from the remaining natural language sentence candidates 4 and 6, since the number of negation words does not match that of the corresponding part of the original text D1. As a result, candidate 4 remains in the set of output units.

[0077] Therefore, in step S306, candidates 2 and 4 remaining in the set of output sentences are adopted, and the natural language sentences "No bleeding from the previous chapter" and "Language disorder" are output.

[0078] <Advantages of this exemplary embodiment> As described above, in this exemplary embodiment, a configuration is adopted in which, from among the candidates for natural language sentences generated using the language model L, those that are not included in the corresponding portion of the original text D1 and have the same number of negated words as the corresponding portion of the original text D1 are selected and output.

[0079] Here, even if a natural language sentence generated using the language model L includes an expression included in the corpus D2, it may coincidentally match a sentence included in the original text D1. By adopting the above configuration, this exemplary embodiment can generate a natural language sentence that is more reliably diverse than the corresponding part of the original text D1.

[0080] In addition, the natural language sentence generated using the language model L may include expressions that are not included in the original text D1 and are included in the corpus D2. Therefore, the natural language sentence may not match the content of the original text D1. By adopting the above configuration, the present exemplary embodiment can generate a natural language sentence that has more reliable diversity compared to the corresponding part of the original text D1 and more accurately indicates the content of the original text D1.

[0081] In addition, in this exemplary embodiment, in order to generate a natural language sentence from a logical formula indicating an inference result, a language model L is used that does not include the original text D1 used to generate the inference rule R and that is generated based on a corpus D2 in the same field as the original text D1.

[0082] As a result, this exemplary embodiment generates natural language sentences that are more diverse than the sentences contained in original text D1, while also generating natural language sentences that include expressions specific to the same field as original text D1, making it possible to generate natural language sentences that more accurately express the content of the inference result.

[0083] In addition, in this exemplary embodiment, a configuration is adopted in which multiple text elements included in a logical formula are divided into multiple groups by including those with the same attributes of observed events in one group, and natural language sentence candidates are generated for each group.

[0084] As a result, the natural language sentence generated by this exemplary embodiment does not include multiple text elements that are less relevant because the attributes of the observed event are different, and the content of the inference result can be expressed more accurately.

[0085] <Modification> In this exemplary embodiment, an example has been described in which the condition that the candidate is included in the relevant portion is used as the condition for determining whether the candidate of the natural language sentence is similar to the relevant portion of the original text D1. That is, in this case, the candidate of the natural language sentence is excluded if a sentence matching the candidate is included in the relevant portion of the original text D1, and is adopted otherwise. The condition for the determination is not limited to this, and can be changed to the condition that a sentence having a similarity to the candidate equal to or greater than a threshold is included in the relevant portion. In this case, the candidate of the natural language sentence is excluded if a sentence similar to the candidate is included in the relevant portion of the original text D1, and is adopted otherwise. Even when the condition for the determination is changed in this way, the exemplary embodiment can output natural language sentences with more diversity for the original text D1.

[0086] In addition, in this exemplary embodiment, an example has been described in which the condition that the number of negation words included in the candidate matches the number of negation words included in the corresponding portion is used as a condition for determining whether or not a candidate natural language sentence is consistent with the content of the original sentence D1. This is not limited to this, and for example, the condition for the determination can be changed to other conditions that refer to the comparison result of the number of negation words. For example, the other condition for the comparison result of the number of negation words may be that the difference in the number of negation words is a multiple of 2. Furthermore, the condition for the determination is not limited to the condition for the comparison result of the number of negation words, and may be other conditions. Even when the condition for the determination is changed in this way, the exemplary embodiment can output a natural language sentence that more accurately represents the content of the original sentence D1.

[0087] In addition, in the present exemplary embodiment, an example has been described in which, when a logical formula expressing an inference result includes multiple text elements, multiple groups are generated by grouping together related observed events having the same attribute. This is not a limitation, and the generation unit 22 may generate multiple groups by grouping together related observed events having attributes that satisfy a predetermined condition. Furthermore, the generation unit 22 may generate multiple groups based on other attributes related to each text element, not limited to the attributes of observed events.

[0088] [Software implementation example] Some or all of the functions of the information processing devices 1 and 2 may be realized by hardware such as an integrated circuit (IC chip), or may be realized by software.

[0089] In the latter case, the information processing devices 1 and 2 are realized, for example, by a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter, referred to as computer C) is shown in FIG. 11. The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for operating the computer C as the information processing devices 1 and 2. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing each function of the information processing devices 1 and 2.

[0090] The processor C1 may be, for example, a central processing unit (CPU), a graphic processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a microcontroller, or a combination of these. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination of these.

[0091] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may further include a communication interface for transmitting and receiving data to and from other devices. The computer C may further include an input / output interface for connecting input / output devices such as a keyboard, a mouse, a display, and a printer.

[0092] Furthermore, the program P can be recorded on a non-transitory tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can obtain the program P via such a recording medium M. Furthermore, the program P can be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also obtain the program P via such a transmission medium.

[0093] [Additional Note 1] The present invention is not limited to the above-described embodiment, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiment are also included in the technical scope of the present invention.

[0094] [Additional Note 2] Some or all of the above-described embodiments can be described as follows. However, the present invention is not limited to the aspects described below.

[0095] (Appendix 1) An acquisition means for acquiring a logical formula representing an inference result based on an inference rule for an observed event; a generation means for generating a natural language sentence from a text element, which is a text among the components of the logical formula, by using a language model generated based on a corpus that does not include the document used in generating the inference rules; and an output means for outputting the natural language sentence.

[0096] The above configuration is expected to generate natural language sentences that include expressions used in a corpus that does not include documents used to generate inference rules. Therefore, natural language sentences that are more diverse than those used in the documents used to generate inference rules are generated from logical expressions that represent inference results.

[0097] (Appendix 2) 2. The information processing device according to claim 1, wherein the generation means uses, as the language model, a language model that does not include the document and that is generated using a corpus in the same field as the document.

[0098] With the above configuration, the generated natural language sentences are expected to include expressions specific to the target field, and therefore it is possible to generate natural language sentences that accurately express the contents of the inference results while having more diversity than the documents used to generate the inference rules.

[0099] (Appendix 3) a first determination means for determining whether the natural language sentence is similar to the document by referring to a comparison result between the natural language sentence and the document; 3. The information processing device according to claim 1, wherein the output means outputs the natural language sentence when the first determination means determines that the natural language sentence is not similar.

[0100] With the above configuration, natural language sentences that are not similar to the documents used to generate the inference rules are adopted, making it possible to generate more reliable and diverse natural language sentences for the documents.

[0101] (Appendix 4) and a second determination means for determining whether the natural language sentence matches the content of the document by referring to a comparison result between the natural language sentence and the document, 4. The information processing device according to claim 1, wherein the output means outputs the natural language sentence when the second determination means determines that there is consistency.

[0102] The above configuration employs natural language sentences that are consistent with the contents of the documents used to generate the inference rules, making it possible to generate natural language sentences that are more diverse than the documents used to generate the inference rules, but that more accurately express the contents of the inference results.

[0103] (Appendix 5) The information processing device according to claim 4, wherein the second determination means refers to a result of a comparison between the number of negation words included in the natural language sentence and the number of negation words included in the document.

[0104] With the above configuration, natural language sentences that express the opposite meaning to the content of the documents used to generate the inference rules are not adopted, and therefore it is possible to generate natural language sentences that more accurately express the content of the inference results while still having more diversity than the documents used to generate the inference rules.

[0105] (Appendix 6) The information processing device according to any one of appendices 1 to 5, wherein when the logical formula includes multiple text elements, the generation means divides the multiple text elements into multiple groups and generates the natural language sentence for each group.

[0106] With the above configuration, the generated natural language sentence does not include multiple text elements that are so poorly related to each other that they are classified into different groups, and therefore it is possible to generate natural language sentences that are more diverse than the documents used to generate the inference rules, while still more accurately expressing the content of the inference results.

[0107] (Appendix 7) The information processing device according to claim 6, wherein the generating means divides the plurality of text elements into a plurality of groups based on attributes of the observed events associated with each text element.

[0108] With the above configuration, the generated natural language sentence does not include multiple text elements with low relevance to the attributes of the observed events, and therefore it is possible to generate natural language sentences that are more diverse than the documents used to generate the inference rules, while still expressing the content of the inference results more accurately.

[0109] (Appendix 8) The information processing device according to any one of appendices 1 to 7, wherein, when the logical formula includes a plurality of the text elements, the generation means generates a first candidate natural language sentence in which the plurality of text elements appear in a first order and a second candidate natural language sentence in which the plurality of text elements appear in a second order different from the first order, and generates one or both of the first candidate and the second candidate as the natural language sentence.

[0110] With the above configuration, the order in which multiple text elements appear is taken into consideration, making it possible to generate natural language sentences that more accurately express the content of the inference results while still having greater diversity than the documents used to generate the inference rules.

[0111] (Appendix 9) obtaining a logical formula representing an inference result based on the inference rules for the observed event; generating a natural language sentence from a text element that is a text among components of the logical formula, using a language model generated based on a corpus that does not include the document used to generate the inference rule; and and outputting the natural language sentence.

[0112] The above configuration provides the same effects as those of Supplementary Note 1.

[0113] (Appendix 10) A program for causing a computer to function as an information processing device, the program comprising: An acquisition means for acquiring a logical formula representing an inference result based on an inference rule for an observed event; a generation means for generating a natural language sentence from a text element, which is a text among the components of the logical formula, by using a language model generated based on a corpus that does not include the document used in generating the inference rules; and a program that functions as an output means for outputting the natural language sentence.

[0114] The above configuration provides the same effects as those of Supplementary Note 1.

[0115] [Additional Note 3] A part or all of the above-described embodiments can be further expressed as follows.

[0116] At least one processor, the processor comprising: An acquisition process for acquiring a logical formula representing an inference result based on an inference rule for an observed event; a generation process for generating a natural language sentence from a text element, which is a text among the components of the logical formula, using a language model generated based on a corpus that does not include the document used in generating the inference rules; and an information processing device that executes an output process for outputting the natural language sentence.

[0117] The information processing device may further include a memory that stores a program for causing the processor to execute the acquisition process, the generation process, and the output process. The program may be recorded in a computer-readable, non-transitory, tangible recording medium. [Explanation of symbols]

[0118] 1, 2 Information processing device 12, 22 Generation part 11, 21 Acquisition Department 13, 23 Output section 24 1st Judgment Section 25 Second Judgment Section 26 Language model generation unit C. Computer C1 Processor C2 Memory

Claims

1. An acquisition means for acquiring a logical formula representing an inference result based on an inference rule for an observed event; a generation means for generating a natural language sentence from a text element, which is a text among the components of the logical formula, by using a language model generated based on a corpus that does not include the document used in generating the inference rules; and an output means for outputting the natural language sentence.

2. The information processing apparatus according to claim 1 , wherein the generating means uses, as the language model, a language model that does not include the document and that is generated using a corpus in the same field as the document.

3. a first determination means for determining whether the natural language sentence is similar to the document by referring to a comparison result between the natural language sentence and the document; The information processing apparatus according to claim 1 , wherein the output means outputs the natural language sentence when the first determination means determines that the natural language sentence is not similar.

4. a second determination means for determining whether the natural language sentence matches the content of the document by referring to a comparison result between the natural language sentence and the document; The information processing apparatus according to claim 1 , wherein the output means outputs the natural language sentence when the second determination means determines that the natural language sentence matches.

5. The information processing apparatus according to claim 4 , wherein the second determination means refers to a result of a comparison between the number of negated words included in the natural language sentence and the number of negated words included in the document.

6. The information processing apparatus according to claim 1 , wherein, when the logical formula includes a plurality of the text elements, the generating means divides the plurality of the text elements into a plurality of groups and generates the natural language sentence for each group.

7. The information processing apparatus according to claim 6 , wherein the generating means divides the plurality of text elements into a plurality of groups based on an attribute of the observed event related to each text element.

8. 8. The information processing device according to claim 1, wherein, when the logical formula includes a plurality of the text elements, the generation means generates a first candidate natural language sentence in which the plurality of the text elements appear in a first order and a second candidate natural language sentence in which the plurality of the text elements appear in a second order different from the first order, and generates one or both of the first candidate and the second candidate as the natural language sentence.

9. obtaining a logical formula representing an inference result based on the inference rules for the observed event; generating a natural language sentence from a text element that is a text among components of the logical formula, using a language model generated based on a corpus that does not include the document used to generate the inference rule; and and outputting the natural language sentence.

10. A program for causing a computer to function as an information processing device, the program comprising: An acquisition means for acquiring a logical formula representing an inference result based on an inference rule for an observed event; a generation means for generating a natural language sentence from a text element, which is a text among the components of the logical formula, by using a language model generated based on a corpus that does not include the document used in generating the inference rules; and a program that functions as an output means for outputting the natural language sentence.

Images