JP7553314B2 - Estimation device, estimation method, and program - Google Patents

Description

The present invention relates to an estimation device, an estimation method, and a program.

Currently, there are services that allow researchers to freely post papers. Researchers can freely view the posted papers and use them in their own research. Patent Document 1 discloses technology that can identify a user's research interests based on the papers collected by the user.

JP 2005-346225 A

Because the number of submitted papers is enormous, it is not realistic for users to check all of the papers in order to catch up on the latest technologies that are attracting attention. Therefore, it is possible to narrow down the number of papers to be checked by searching for papers using technical words used in the latest technologies. However, since the technical words used in the latest technologies are not listed in dictionaries and users themselves are often not familiar with them, narrowing down the papers using technical words is difficult in itself. Note that this issue is not limited to papers, but can occur with any type of text, such as books or online documents.

The present invention aims to provide a technology that can extract specific terms used in multiple sentences by analyzing multiple sentences without using a dictionary.

The estimation device according to one aspect of the present invention includes a receiving unit that receives input of a plurality of sentences, a generating unit that generates N-grams from a word group including a plurality of consecutive words contained in the plurality of sentences, and an estimating unit that estimates a specific term used in the plurality of sentences by evaluating the similarity between two adjacent N-grams among the generated N-grams.

The present invention provides a technology that can analyze multiple sentences and extract specific terms used in multiple sentences without using a dictionary.

FIG. 1 illustrates an example of a document analysis system. FIG. 2 illustrates an example of a hardware configuration of an analysis device and a terminal. FIG. 2 is a diagram illustrating an example of a functional block configuration of an analysis device. 13 is a flowchart illustrating an example of a processing procedure performed by the analysis device. FIG. 13 is a diagram for explaining an example of a process for estimating technical words. FIG. 13 is a diagram showing an example of a graph showing the number of papers and points of change, which is displayed on a terminal.

The following describes an embodiment of the present invention with reference to the attached drawings. In each drawing, the same reference numerals denote the same or similar configurations.

<System Configuration>
1 is a diagram showing an example of a document analysis system 1. The document analysis system 1 includes an analysis device 10 and a terminal 20. There is no limit to the number of terminals 20 included in the document analysis system 1. The analysis device 10 and the terminal 20 are connected via a wireless or wired communication network N and can communicate with each other.

The analysis device 10 estimates technical words (specific terms) used in the papers without using a dictionary by analyzing a large number of papers published on the Internet, etc. Furthermore, the analysis device 10 detects signs of the latest technology trends based on changes in the frequency with which the estimated technical words are used in the papers.

The analysis device 10 may be configured with one or more physical servers, or may be configured with a virtual server that runs on a hypervisor, or may be configured with a cloud server.

The terminal 20 is a device that displays the estimation results by the analysis device 10. The terminal 20 displays, for example, a graph showing the time series of changes in the frequency with which the latest technical words estimated by the analysis device 10 are used in papers. The terminal 20 is a personal computer (PC), a notebook PC, a smartphone, a tablet terminal, a mobile phone, a personal digital assistant (PDA), etc.

<Hardware Configuration>
2 is a diagram showing an example of the hardware configuration of the analysis device 10 and the terminal 20. The analysis device 10 and the terminal 20 each include a processor 11 such as a central processing unit (CPU) or a graphical processing unit (GPU), a storage device 12 such as a memory, a hard disk drive (HDD) and/or a solid state drive (SSD), a communication interface (IF) 13 for wired or wireless communication, an input device 14 for accepting input operations, and an output device 15 for outputting information. The input device 14 is, for example, a keyboard, a touch panel, a mouse, and/or a microphone. The output device 15 is, for example, a display, a touch panel, and/or a speaker.

<Function block configuration>
FIG. 3 is a diagram showing an example of a functional block configuration of the analysis device 10. The analysis device 10 includes a storage unit 100, a reception unit 101, a generation unit 102, an estimation unit 103, a counting unit 104, a detection unit 105, and a learning unit 106. The storage unit 100 can be realized by using a storage device 12 provided in the analysis device 10. The reception unit 101, the generation unit 102, the estimation unit 103, the counting unit 104, the detection unit 105, and the learning unit 106 can be realized by the processor 11 of the analysis device 10 executing a program stored in the storage device 12. The program can be stored in a storage medium. The storage medium storing the program may be a non-transitory computer readable storage medium. The non-transitory storage medium is not particularly limited, and may be, for example, a storage medium such as a USB memory or a CD-ROM.

The storage unit 100 stores a paper DB 110, a technical word estimation model M110 used to estimate technical words, and a group estimation model M120 used when grouping similar technical words. The paper DB 110 is a database that stores paper data acquired from the Internet, for example. The paper DB 110 may be stored in a storage device 12 provided in the analysis device 10, or in an external device that can communicate with the analysis device 10.

The reception unit 101 receives input of multiple papers (multiple sentences) to be analyzed. The reception unit 101 also stores the multiple papers it receives in the paper DB 110. Each of the multiple papers received by the reception unit 101 includes information indicating a date. The information indicating the date may be, for example, the date the paper was submitted or the date the paper was created.

The generation unit 102 generates an N-gram from a group of words containing multiple consecutive words obtained from multiple papers to be analyzed.

In this embodiment, an N-gram is a character string generated by dividing a sentence containing multiple words into units of N consecutive words. In addition, when N is 1, it is called a Uni-gram, when N is 2, it is called a Bi-gram, when N is 3, it is called a Tri-gram, when N is 4, it is called a Four-gram, when N is 5, it is called a Five-gram, when N is 6, etc. The generation unit 102 generates multiple N-grams from 1 to n, where n is the number of words contained in the word group.

The estimation unit 103 estimates technical words (specific terms) used in the multiple papers to be analyzed by evaluating the similarity between two adjacent N-grams among the generated N-grams. Furthermore, when there are multiple estimated technical words, the estimation unit 103 estimates a group of similar technical words (a group of technical words) by evaluating the similarity based on the distributed representation (vector) of each of the multiple technical words.

The counting unit 104 counts the number of papers that contain a technical word from among the multiple papers to be analyzed for each specified period. The counting unit 104 may also be configured to count the number of papers that contain at least one technical word from a group of technical words from among the multiple papers to be analyzed for each specified period.

The detection unit 105 detects a change point in the time series where the number of papers including a technical word increases, based on the number of papers for each specified period tallied by the aggregation unit 104. The detection unit 105 may also detect a change point in the time series where the number of papers including at least one technical word from a group of technical words increases, based on the number of papers for each specified period tallied by the aggregation unit 104.

The learning unit 106 uses the paper data stored in the paper DB 110 to train the technical word estimation model M110 and the group estimation model M120.

<Processing Procedure>
Fig. 4 is a flowchart showing an example of a processing procedure performed by the analysis device 10. Below, a series of processing procedures from when the analysis device 10 reads a large amount of paper data, performs analysis, and displays technical terms that are attracting attention on the screen of the terminal 20 will be described with reference to Fig. 4.

In step S10, the reception unit 101 receives input of multiple papers to be analyzed and stores them in the paper DB 110. For example, the reception unit 101 may access a service provided on the Internet that allows researchers to freely post papers, download papers posted in the past (which may be the entire past, or a partial period such as the past five years), and store them in the paper DB 110. Also, rather than downloading all of the papers and storing them in the paper DB 110, only the abstract portion or the main text of the papers may be downloaded and stored in the paper DB 110.

In step S11, the generation unit 102 breaks down each sentence included in the paper to be analyzed from among the paper data stored in the paper DB 110 into words. For example, the generation unit 102 breaks down Japanese into words by performing morphological analysis, and for English, recognizes spaces as word separators.

In step S12, the generation unit 102 performs a cleansing process to delete unnecessary characters and words (e.g., articles, subjects, conjunctions, be verbs, etc.) and leave words that can become technical words. At this time, the generation unit 102 divides the sentence at the part where the deleted characters or words were present, so that it can recognize word groups that include one or more words contained in the divided part. In addition, by performing stem processing on each word, for verbs and adjectives with invariable endings, only the stem remains and parts other than the stem are deleted.

Here, we will explain a specific example of the processing procedures of steps S11 and S12 using the sentence "In this paper, we propose the simple Generative Adversarial Network model which allows long-range dependency modeling for image generation tasks." as an example.

First, the generation unit 102 breaks down the sentence into words by recognizing spaces and punctuation marks (commas, periods, etc.) as word separators. Next, the generation unit 102 performs a cleansing process and inserts separators (for convenience of explanation, assume ":") so that the parts where deleted characters or words exist can be recognized. As a result, the generation unit 102 outputs the string "paper : propose : simple Generative Adversarial Network model : allow long-range dependency modeling : image generation tasks".

Next, the generation unit 102 performs stem processing to leave only the stem. As a result, the generation unit 102 outputs the character string "paper : propose : simple Gener Adversar Network model : allow long-range depend model : image generator task".

In step S13, the estimation unit 103 estimates technical words. Here, for a portion of the sentence generated in the processing procedure of step S12 that contains only one word separated by delimiters, the estimation unit 103 estimates that one word as a technical word. For example, for the character string "paper : propose : simple generative adversarial network model : allow long-range dependency modeling : image generation tasks", the estimation unit 103 estimates that "paper" and "propose" are technical words.

Next, for a word group in which multiple words are contained in a portion separated by a delimiter, the estimation unit 103 estimates whether each word is a technical word, or whether a phrase consisting of multiple consecutive words is a technical word. First, the generation unit 102 generates an N-gram for each word group.

Below, an example of generating an N-gram for the word group "simple Generative Adversarial Network model" will be described. Since the word group includes five words, the generation unit 102 generates multiple N-grams, 1 to 5. Specifically, for unigrams, five character strings are generated: "simple", "Generative", "Adversarial", "Network", and "model". For bigrams, four character strings are generated: "simple Generative", "Generative Adversarial", "Adversarial Network", and "Network model". For trigrams, three character strings are generated: "simple Generative Adversarial", "Generative Adversarial Network", and "Adversarial Network model". For fourgrams, two character strings are generated: "simple Generative Adversarial Network" and "Generative Adversarial Network model". For fivegrams, one character string is generated: "simple Generative Adversarial Network model".

Then, the estimation unit 103 estimates technical words using the generated N-grams. The estimation unit 103 determines whether or not the similarity between two adjacent N-grams (which may be referred to as a pair of adjacent N-grams) for an N-gram with the value of N subtracted by 1 from the maximum value of N is equal to or greater than a predetermined threshold (first threshold). If there are no adjacent N-grams with a similarity equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 repeats the process of determining whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold (first threshold) for an N-gram with the value of N further subtracted by 1, until the value of N becomes 1. If there is an N-gram with a similarity between two adjacent N-grams equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 estimates, as technical words, phrases that combine, in order, words included in an N-gram with an N value that is one larger than the two adjacent N-grams.

In the case of unigrams, "two adjacent N-grams" refers to a pair of two consecutive words in a word group, and "a phrase formed by sequentially combining words included in an N-gram with an N value one larger corresponding to two adjacent N-grams" refers to a phrase formed from the two consecutive words. In the case of bigrams or higher, "two adjacent N-grams" refers to a pair of an N-gram formed from consecutive words p to (p+N-1) among the words included in the word group, and an N-gram formed from consecutive words (p+1) to (p+N). In addition, "a phrase formed by sequentially combining words included in an N-gram with an N value one larger corresponding to two adjacent N-grams" refers to an N-gram phrase formed from consecutive words p to (p+N).

For example, suppose that the "two adjacent N-grams" in the case of bigrams are to be found for the word group "A B C D" (A to D represent words). In this case, since n = 4 and N = 2, the value of p that satisfies p + 2 ≦ 4 is 1 or 2. Note that if there are multiple values of p that satisfy p + N ≦ n, the estimation unit 103 determines for each of the multiple values whether the similarity between the two adjacent N-grams is equal to or greater than a predetermined threshold (first threshold).

When p=1, the "two adjacent N-grams" are the consecutive words from the 1st to the (1+2-1)th and the consecutive words from the (1+1)th to the (1+2)th, i.e., "A B" and "B C." Also, the "phrase formed by combining, in order, the words contained in the N-gram with an N value one larger that corresponds to the two adjacent N-grams" is the consecutive words from the 1st to the (1+2)th, i.e., "A B C."

When p=2, the "two adjacent N-grams" are the consecutive words from the second to the (2+2-1)th word and the consecutive words from the (2+1)th to the (2+2)th word, i.e., "B C" and "C D." Also, the "phrase formed by combining, in order, the words contained in the N-gram with an N value one larger that corresponds to the two adjacent N-grams" is the consecutive words from the second to the (2+2)th word, i.e., "B C D."

Here, the estimation unit 103 determines the similarity between two adjacent N-grams using a technical word estimation model M110 that outputs the similarity between two input words. The technical word estimation model M110 is a trained model that outputs the similarity between words, and is generated in advance using paper data stored in the paper DB 110. The technical word estimation model M110 can be generated, for example, using a technology called word2vec.

As shown in FIG. 3, in this embodiment, the technical word estimation model M110 includes a unigram model M111 that outputs similarities between unigram words, a bigram model M112 that outputs similarities between bigram words, a trigram model M113 that outputs similarities between trigram words, and a four-gram model M114 that outputs similarities between four-gram words. Note that the five models included in the technical word estimation model M110 are merely examples, and models with even larger values of N, such as a five-gram model or a six-gram model, may also be included.

Here, a method for generating the bigram model M112 will be described. The model can be generated by performing the process described in the processing procedure of steps S11 and S12 for each sentence of the paper data stored in the paper DB 110, creating sentences by connecting each sentence in two-word units, and having word2vec learn the created sentences. For example, when the sentence "paper propose simple Gener Adversar Network model allow long-range depend model image generator task" output by the processing procedure of steps S11 and S12 exists, the learning unit 106 generates the sentence "paper_propose propose_simple simple_Gener Gener_Adversari Adversari_Network Network_model model_allow allow_long-range long-range_dependency dependency_modeling modeling_image image_generation generation_task" and has word2vec learn it. The learning unit 106 repeatedly performs such a process for all sentences included in the paper data stored in the paper DB 110. This allows a distributed representation (vector) to be determined for a bigram linking two words included in the paper data stored in the paper DB 110, and the estimation unit 103 can use the learned word2vec to evaluate the similarity between two bigrams based on the distributed representation. Note that word2vec is merely an example, and any technology that evaluates the similarity between words based on a distributed representation can be used.

Figure 5 is a diagram for explaining an example of a process for estimating technical words. Using Figure 5, an example will be described in which the estimation unit 103 estimates technical words present in the word group "simple gener adversarial network model". First, when the number of words included in the word group is n, the estimation unit 103 determines whether or not the similarity between two adjacent N-grams, for an N-gram in which the value of N is the maximum value of n minus 1, is equal to or greater than a predetermined threshold. In the example of Figure 5, the word group includes five words, so the value of n is 5.

First, the estimation unit 103 estimates the similarity between the adjacent four-gram pair, i.e., "simple gener adversari network" and "gener adversari network model", for the four-gram obtained by subtracting 1 from 5, using the four-gram model M114. If the similarity is equal to or greater than a predetermined threshold (first threshold), the estimation unit 103 estimates the five-gram phrase one level above these four-gram pairs (i.e., simple gener adversari network model) as a technical word.

On the other hand, if the similarity is less than the predetermined threshold (first threshold), the estimation unit 103 uses the trigram model M113 to estimate the similarity between adjacent trigram pairs, i.e., between "simple gener adversari" and "gener adversari network", and between "gener adversari network" and "adversari network model", for the trigram obtained by subtracting 1 from 4.

If the similarity between "simple gener adversari" and "gener adversari network" is equal to or greater than a predetermined threshold (first threshold), the estimation unit 103 estimates the four-gram phrase one level above these trigram pairs (i.e., simple gener adversari network) as a technical word. If the similarity between "gener adversari network" and "adversari network model" is equal to or greater than a predetermined threshold (first threshold), the estimation unit 103 estimates the four-gram phrase one level above these trigram pairs (i.e., gener adversari network model) as a technical word. If both the similarity between "simple gener adversari" and "gener adversari network" and the similarity between "gener adversari network" and "adversari network model" are equal to or greater than a predetermined threshold (first threshold), the estimation unit 103 estimates the four-gram phrase one level above these trigram pairs with the higher similarity as a technical word.

On the other hand, if all similarities are less than the predetermined threshold (first threshold), the estimation unit 103 estimates the similarity between adjacent bigram pairs, i.e., the similarity between "simple gener" and "gener adversari", the similarity between "gener adversari" and "adversari network", and the similarity between "adversari network" and "network model", for the bigrams obtained by subtracting 1 from 3, using the bigram model M112. If there is a pair of adjacent bigrams whose similarity is equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 estimates the trigram phrase one level above the two bigrams as the technical word. If there are multiple pairs of adjacent bigrams whose similarity is equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 estimates the trigram phrase one level above the pair with the greatest similarity as the technical word.

On the other hand, if all similarities are less than the predetermined threshold (first threshold), the estimation unit 103 uses the unigram model M111 to estimate the similarity between adjacent unigram pairs, i.e., the similarity between "simple" and "gener", the similarity between "gener" and "adversari", the similarity between "adversari" and "network", and the similarity between "network" and "model", for the unigram obtained by subtracting 1 from 2. If there is an adjacent unigram pair among these whose similarity is equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 estimates the bigram phrase one level above the pair as the technical word. If there are multiple unigram pairs whose similarity is equal to or greater than the predetermined threshold (first threshold), the estimation unit 103 estimates the bigram phrase one level above the pair with the greatest similarity as the technical word.

On the other hand, if all similarities are less than the predetermined threshold (first threshold), the estimation unit 103 estimates all unigrams (i.e., simple, gener, adversarial, network, model) as technical words.

In step S14, if there are multiple technical words estimated in the processing procedure of step S13, the estimation unit 103 estimates a group (a group of technical words) that includes similar technical words by evaluating the similarity based on the distributed representation (vector) of each of the multiple technical words. The estimation unit 103 estimates the group by using the group estimation model M120.

The group estimation model M120 can be generated, for example, by carrying out the processing procedures of steps S12 and S13 for each sentence of the paper data stored in the paper DB110, preparing sentences in which the technical words of phrases estimated in the processing procedure of step S14 are converted into a single character string by combining them with underscores or the like, and having word2vec learn the prepared sentences.

Here, an example of generating the group estimation model M120 will be described. It is assumed that "simple gener adversari" is estimated as a technical word of the phrase in the processing procedure of step S14. First, the generation unit 102 performs the processing procedures of steps S11 and S12 to convert the sentence "In this paper, we propose the simple Generative Adversarial Network model which allows long-range dependency modeling for image generation tasks." into the sentence "paper propose simple Gener Adversar Network model allow long-range depend model image generator task". Next, the generation unit 102 combines multiple words included in the technical words of the phrase with underscores so that the phrase is recognized as one word in word2vec. Specifically, the generation unit 102 converts the sentence into "paper propose simple_Gener_Adversari Network model allow long-range depend model image generator task". Next, the learning unit 106 trains the converted sentence in word2vec. This makes it possible to find the distributed representation (vector) of each technical word, so that by inputting two technical words, a learning model can be generated that can output the similarity between the two technical words.

The estimation unit 103 estimates the similarity between two technical words for all technical words estimated in the processing procedure of step S13 in a brute force manner, and sets combinations of technical words with similarities close to each other as a group of technical words. For example, the estimation unit 103 may regard combinations of technical words whose similarity is equal to or greater than a predetermined threshold (second threshold) in all combinations as a group of technical words. For example, if the similarity is equal to or greater than the predetermined threshold (second threshold) in all combinations of the four words "blockchain", "smart contract", "bitcoin", and "ethereum", the estimation unit 103 may regard the four words "blockchain", "smart contract", "bitcoin", and "ethereum" as a group of technical words. Note that the method of estimating groups of technical words is not limited to this, and other clustering methods may be used.

In step S15, the counting unit 104 counts the number of papers that contain at least one of the technical words in the technical word group for all papers stored in the paper DB 110 for each group of technical words and for each predetermined period based on the information indicating the date included in each paper (here, the posting date). For example, if there is a group of technical words consisting of four words, "blockchain", "smart contract", "bitcoin", and "ethereum", papers that contain at least one of these words are searched for, and the number of papers is counted for a predetermined period (e.g., one-month intervals) using the posting dates included in the searched papers. This makes it possible to obtain data such as the number of posted papers containing the above technical word group being 5 in January 2015, 7 in February 2015, 10 in March 2015, and 20 in April 2015.

Then, the detection unit 105 detects a change point in the time series where the number of papers including at least one technical word from the group of technical words increases, based on the number of papers for each specified period tallied by the aggregation unit 104. The detection unit 105 may detect a change point in the time series by using a known change point detection algorithm such as Change finder, or may detect a change point in the time series based on the ratio of the number of papers by month (e.g., a 20% increase from the previous month). The detection unit 105 is not limited to these, and may detect a change point in any manner.

Figure 6 shows an example of a graph showing the number of papers and change points displayed on the terminal 20. Figure 6(a) is a graph showing the change in the number of papers that contain at least one technical word from a group of specific technical words. The vertical axis represents the number of papers submitted, and the horizontal axis represents the year and month. Figure 6(b) shows the score of the change point calculated based on the number of papers submitted shown in Figure 6(a). In Figure 6(b), the points where the score increases suddenly are the change points. Figure 6(b) shows that the number of papers submitted changed significantly around June 2018 and around September 2018.

In the processing procedure described above, the processing procedure of step S14 may be omitted. This is because it is expected that grouping may not be necessary depending on the technical word. If the processing procedure of step S14 is omitted, in the processing procedure of step S15, the aggregation unit 104 may aggregate the number of papers containing the technical word for each technical word and for each predetermined period for all papers stored in the paper DB 110, based on information indicating the date included in each paper. Furthermore, the detection unit 105 may detect a change point in the time series where the number of sentences containing the technical word increases, based on the number of sentences for each predetermined period aggregated by the aggregation unit 104.

<Summary>
According to the embodiment described above, the analysis device 10 generates N-grams from the paper to be analyzed and evaluates the similarity between adjacent N-grams. This makes it possible to extract the latest technical words used in multiple papers without using a dictionary. The analysis device 10 also estimates a group of similar technical words and detects signs of the latest technology trend based on the change in the frequency of at least one technical word in the estimated group of technical words used in the paper. There are many cases where the names of the latest technical words are not uniquely determined, but by counting the number of papers taking similar technical words into consideration, it becomes possible to more appropriately detect signs of the latest technology trend.

In the embodiment described above, the analysis device 10 is assumed to analyze a submitted paper, but the present embodiment is not limited to this. The present embodiment is not limited to papers, and can be applied to the analysis of various texts.

The above-described embodiments are intended to facilitate understanding of the present invention, and are not intended to limit the present invention. The flow charts, sequences, elements included in the embodiments, and their arrangements, materials, conditions, shapes, sizes, etc., described in the embodiments are not limited to those exemplified, and may be modified as appropriate. In addition, configurations shown in different embodiments may be partially substituted or combined.

1... document analysis system, 10... analysis device, 11... processor, 12... storage device, 13... communication IF, 14... input device, 15... output device, 20... terminal, 100... storage unit, 101... reception unit, 102... generation unit, 103... estimation unit, 104... aggregation unit, 105... detection unit, 106... learning unit

Claims (5)

A reception unit that receives input of a plurality of sentences;
a generation unit that generates an N-gram from a word group including a plurality of consecutive words included in the plurality of sentences;
an estimation unit that estimates a specific term used in the plurality of sentences by evaluating a similarity between two adjacent N-grams among the generated N-grams;
an output unit that outputs information related to the specific term;
having
the maximum value of N in the N-gram is the number of words included in the word group,
The estimation unit is
For an N-gram obtained by subtracting 1 from the maximum value of N, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold. If there are no adjacent N-grams whose similarity is equal to or greater than the first threshold, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold for an N-gram obtained by further subtracting 1 from the value of N, and repeat this process until the value of N becomes 1;
When there is an N-gram in which the similarity between two adjacent N-grams is equal to or greater than the first threshold, a phrase formed by combining, in order, words included in the N-gram corresponding to the two adjacent N-grams and having an N value that is one larger than the first threshold, is estimated as the specific term .
Estimation device. Each of the plurality of sentences includes information indicating a date; and
a counting unit that counts the number of sentences including the specific term from among the plurality of sentences for each predetermined period based on the information indicating the date;
a detection unit that detects a change point in the time series where the number of sentences including the specific term increases based on the number of sentences for the predetermined period tallied by the counting unit,
The output unit outputs information regarding change points on the time series.
The estimation device according to claim 1 . the estimation unit, when there are a plurality of estimated specific terms, estimates a similarity based on a distributed representation of each of the plurality of specific terms to estimate a group of the specific terms;
the counting unit counts, for each predetermined period, the number of sentences including at least one specific term in the group of specific terms from among the plurality of sentences based on the information indicating the date;
the detection unit detects a change point in the time series at which the number of sentences including at least one specific term among the group of specific terms increases, based on the number of sentences for the predetermined period tallied by the counting unit;
The estimation device according to claim 2 . An estimation method performed by an estimation device,
A step of receiving an input of a plurality of sentences by the estimation device ;
A step in which an estimation device generates an N-gram from a word group including a plurality of consecutive words included in the plurality of sentences;
A step in which an estimation device estimates a similarity between two adjacent N-grams among the generated N-grams, thereby estimating specific terms used in the plurality of sentences;
The estimation device outputs information about the specific term ,
the maximum value of N in the N-gram is the number of words included in the word group,
The estimating step includes:
For an N-gram obtained by subtracting 1 from the maximum value of N, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold. If there are no adjacent N-grams whose similarity is equal to or greater than the first threshold, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold for an N-gram obtained by further subtracting 1 from the value of N, repeating this process until the value of N becomes 1;
When there is an N-gram in which the similarity between two adjacent N-grams is equal to or greater than the first threshold, a phrase formed by combining, in order, words included in the N-gram corresponding to the two adjacent N-grams and having an N value that is one larger than the first threshold, is estimated as the specific term.
Estimation method. On the computer,
accepting input of a plurality of sentences;
generating N-grams from a word group including a plurality of consecutive words included in the plurality of sentences;
A step of estimating specific terms used in the plurality of sentences by evaluating a similarity between two adjacent N-grams among the generated N-grams;
an estimation device outputting information about the specific term;
Run the command ,
the maximum value of N in the N-gram is the number of words included in the word group,
The estimating step includes:
For an N-gram obtained by subtracting 1 from the maximum value of N, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold. If there are no adjacent N-grams whose similarity is equal to or greater than the first threshold, determine whether or not the similarity between two adjacent N-grams is equal to or greater than a predetermined threshold for an N-gram obtained by further subtracting 1 from the value of N, and repeat this process until the value of N becomes 1;
When there is an N-gram in which the similarity between two adjacent N-grams is equal to or greater than the first threshold, a phrase formed by combining, in order, words included in the N-gram corresponding to the two adjacent N-grams and having an N value that is one larger than the first threshold, is estimated as the specific term.
program.

Images