JP2020035010A - Thinking support system, thinking support program, storage medium for thinking support program and storage medium for thinking support information - Google Patents

Abstract

To provide a thinking support technique which is related to a thinking object text, goes beyond the framework providing hints in traditional analogies, provides a flexible and wide hints, is heavily independent of user skills and knowledge, and needs a low effort for its use.SOLUTION: A text highly relevant with a thinking object text is presented by converting a set of wo-case neighborhood words and phrases to a set of relevant wo-case neighborhood words and phrases using the information relevant with connections between words and phrases, and by converting the wo-case to the highly relevant wo-case also using the information of the set of the converted relevant wo-case neighborhood words and phrases, where the set of the wo-case neighborhood words and phrases is the set of the words and phrases located in the neighborhood of the wo-case included in the thinking object text and the set of the relevant wo-case neighborhood words and phrases is the set of the wo-case neighborhood words and phrases of the set of the highly relevant words and phrases.SELECTED DRAWING: Figure 1

Description

The present invention relates to a natural language processing technology, and more particularly, to a thinking support technology for presenting text candidates related to an acquired text to be thought.

An analogy is one of useful methods for obtaining hints for solving various problems (see Patent Literature 1, Patent Literature 2, Non-Patent Literature 1, Non-Patent Literature 2, and the like). Non-Patent Literature 2 describes a case where a hint for conceiving a “folding blade type cutter knife” is obtained using analogy. Specifically, the similarity of "glass" and "cutter knife" is obtained from the viewpoint of "sharp when cut", and "plate chocolate" and "cutter knife" from the viewpoint of "can be cut by providing grooves". By obtaining the similarity of the above and presenting these two similarities as hints, it is intended to realize a highly creative design support that enables the idea of a "folding blade type cutter knife".

Japanese Patent No. 3373868 JP 2017-59077 A

Isao Hosoya, Thinking Analogy, Toyo Keizai Shimpo, 2011, ISBN978-4-492-55697-9 Hideaki Takeda et al., Development of Universal Abduction Studio (4th report) -Multiple Interpretation Type Knowledge Representation for UAS-, Proc. Of JSCE Academic Lecture Meeting 2004, JSCE Spring Conference, 197-197, 2004 J. Pennington, R. Socher, C. D. Manning, GloVe: Global Vectors for Word Representation, Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 1532-1543, 2014 T. Mikolov, I. Sutskever, K. Chen, G. Corrado, J. Dean, Distributed Representations of Words and Phrases and their Compositionality, Proceedings of Advances in Neural Information Processing Systems 26, 3111-3119, 2013 Masayuki Asahara, NWJC2Vec: Word embedding dataset from ‘NINJAL Web Japanese Corpus’, Terminology: International Journal of Theoretical and Applied Issues in Specialized Communication, 24: 1, 7-22, 2018 nwjc2vec (National Institute for Japanese Language): http://nwjc-data.ninjal.ac.jp/nwjc2vec/ Yoko Kanzaki et al., Acquisition of Semantic Relationship between Words from Corpus and Its Application, National Institute of Information and Communications Technology Quarterly Report, Vol. 53, No. 3, 29-37, 2007 Pang-Ning Tan, Vipin Kumar, Jaideep Srivastava, Selecting the Right Interestingness Measure for Association Pattern, Proc.ACM-SIGKDD Conf. Knowledge Discovery and Data Mining, 491-502, 2002 Japanese Writing Grammar Study Group, Contemporary Japanese Grammar 2, Kuroshio Publishing, 2009, ISBN978-4-87424-460-9 Masayasu Muraoka et al., Generation of Phrase Vectors Using Dependency Relationships, Proc. Of the 20th Annual Conference of the Linguistic Processing Society, 1055-1058, 2014

An analogy is a very effective technique if you can find hints for a task. However, there is a problem that if an attempt is made to find hints only by hand without using a support system or the like as in Non-Patent Document 2, there is a problem that it greatly depends on the knowledge and skills of the user. In addition, when such a support system is used, there is a problem that a great deal of labor is required for its construction and use.

  Also, as we will see later, the traditional analogy approach focuses on structural similarities related to "objects," so it is generally difficult to obtain flexible and broad hints based on other relevance. there were.

  For the same reason, even if the object of thinking is not a concrete "thing" but an abstract "thing", in a conventional analogy, finding structural similarity about the "object" is generally There was also a problem that it was difficult to obtain useful hints because it was difficult.

  Here, the above-described problem will be described by taking a “folded blade type cutter knife” described in Non-Patent Document 2 as a specific example. To be specific, the subject related to the "folding blade type cutter knife" is regarded as a subject text "Keeping a sharp blade", and for this subject text, for example, "Adjusting in a different direction" or "Adjusting a different length" This will be described more specifically by taking an example in which a hint has been obtained.

FIG. 2 is an example of an idea supporting example of an idea “rotary blade type cutter knife” which is conceived based on the former hint of “matching different directions”. FIG. 3 is an idea supporting example regarding the idea “multiple cutter knives” conceived based on the latter hint “match different lengths”.

As apparent from FIGS. 2 and 3, from the former hint, it is possible to conceive a solution of a different form from the “folded-cutter knife”, and the “multi-cutter” by the latter hint. The idea of "knives" can be linked to the idea of "folding blade type knife".

In this way, it can be said that the seemingly unrelated hints “Adjust different directions” and “Adjust different lengths” can effectively support the idea. However, the conventional analogy has a problem that it is generally difficult to obtain such a flexible and wide hint.

  In addition, as an example for abstract “things”, let's consider an issue text “creating new ideas”. Consider a case in which a hint "Give clear motivation" is obtained for this task text.

When asked to "create a new idea", it is better to "create a new idea" than for "whatever is good" or "to create a new idea". In general, it can be said that the idea of "creating ideas" is promoted when it is required to do so.

In this way, it can be said that the idea of "creating a new idea" and the hint of "giving a clear motivation" with little structural similarity can also be supported effectively. . However, the conventional analogy has a problem that it is generally difficult to obtain a hint for such an abstract "thing".

  The present invention focuses on a "target" of a task using textual information (hints) that are effective in solving the task in relation to the task text representing the task to be considered. Instead of focusing on the "effect" on the "object", it is possible to provide more flexible and broader hints than was possible with conventional analogies, and thinking that does not greatly depend on the knowledge and skills of the user It aims to provide assistive technology.

  In order to achieve such an object, a thinking support system according to the present invention includes information on a connection between words and phrases, and uses the information to represent an “action” of an input text that is a text to be thought. A case pair near the case is converted into a related case group that is a set of highly relevant words, and the case is related using information on the related case group. By converting to a highly relevant category, a text related to the input text can be obtained.

In addition, the thinking support system according to the present invention includes information about connections between words and phrases including information in a distributed expression format regarding the connections between words and words, and uses the information to perform parallel processing in a space having information based on the distributed expression format in particular. Using the index related to mobility, the word group that is located near the case that represents the “action” of the input text, which is the text to be thought, is converted to the related word group that is a highly relevant word set. It is configured such that a text related to the input text can be obtained by converting the case into a related word set and converting the case into a highly relevant case using the information on the related case group. I have.

In addition, the thinking support system according to the present invention includes information on a connection between phrases including information on a connection between a noun phrase and a noun phrase that is a phrase located in the vicinity of a noun, and uses the information to perform a thinking process. The word set near the case that is located near the case that represents the “action” of the input text that is the target text of the text is, in particular, the noun phrase contained in the case and the case that is located near the case. By using an index relating to co-occurrence with a phrase, a text related to the input text can be obtained by converting the phrase into a phrase set having a high degree of relevance, which is a pair of highly relevant phrases.

In addition, the thinking support system according to the present invention is a system which includes information on a connection between a noun phrase and a noun phrase, which is a phrase located in the vicinity of a noun, and information in a distributed expression form on a connection between the words. Using the information, and using the information, a set of words near the case that is located near the case that represents the “action” of the input text that is the text to be considered, particularly the nouns included in the case By using an index relating to co-occurrence of a phrase and a phrase near a ヲ case located in the vicinity of the 、 case, and an index related to parallel movement in a space based on the information of the distributed representation form, a phrase having a high relevance is used. By converting the case into a related case with high relevance using information about the related case-neighboring word / phrase set as well as the text related to the input text. It is configured to obtain a list.

In addition, the thinking support system according to the present invention is a system which includes information on a connection between a noun phrase and a noun phrase, which is a phrase located in the vicinity of a noun, and information in a distributed expression form on a connection between the words. Using the information, and using the information, a set of words near the case that is located near the case that represents the “action” of the input text that is the text to be considered, particularly the nouns included in the case An index relating to co-occurrence of a phrase and a phrase located in the vicinity of the case, and an index relating to parallel movement in a space based on the information of the distributed expression form are also used. By converting to a related ヲ case neighborhood phrase set that is a set of highly relevant terms, and by using the information about the related 近 傍 case neighborhood phrase set to convert the に case into a highly relevant related case, Input text And it is configured so as to obtain the relevant text.

  A thinking support program according to the present invention is a program for causing a computer to function as each unit configuring the above-described thinking support system.

  Further, a thinking support program storage medium according to the present invention is a computer-readable program storage medium storing the thinking support program.

  Further, a thinking support information storage medium according to the present invention is a computer-readable information storage medium storing information relating to a connection between at least a noun phrase included in a case and a phrase located in the vicinity of the noun phrase included in the connection information between the phrases. It is.

According to the present invention, a task text representing a task to be considered is not limited to providing a hint based on structural similarity to a “target” in a conventional analogy. Providing more flexible and wide-ranging hints that expands the scope of thinking not only to concrete "things" but also to abstract "things" without relying heavily on the knowledge and skills of users There is an effect that it can be easily realized.

It is a block diagram showing composition of a thought support system. This is an idea support example (“rotary blade type cutter knife”) for the problem of “keeping a sharp blade”. This is an idea support example (“multi-cutter knife”) for the problem of “keeping a sharp blade”. It is an input screen display example of a thinking support system. It is a flowchart which shows a process in a thinking support system. It is an input screen display example of a thinking support system. FIG. 4 is an explanatory diagram of parallel movement in a distributed expression space. It is an example of an output screen display of a thinking support system. It is an example of an output screen display of a thinking support system.

[Principle of the invention]
The analogy is a method of supporting thinking by paying attention to an "object" and converting the "object" to another "object" having a high structural similarity. However, there is one side that it is difficult to provide more flexible and broad hints because of focusing on the "object".

Therefore, the present invention focuses on the "action" of the object rather than the "object" itself, and converts the "action" to another "action" that is more relevant, thereby providing a more flexible and broader hint. It is intended to provide.

Also, in the conversion of the “action”, the conversion is performed not by focusing on the structural similarity of the “action” itself, but by focusing on the connection of the words constituting the “action”, thereby providing more flexibility. It seeks to provide a wide range of tips.

  The details will be described below. First, the “action” will be described.

“Action” refers to a word / phrase set located in the vicinity of a case that represents “target” in a task text that represents a task (here, “pair” means “ordered pair” and also includes “pair (ordered pair)”) ) Can be said to be represented by a set of ヲ case neighborhood phrases. In the above-mentioned task text “Keep the sharp blade”, the ヲ case becomes <the blade>, and the ヲ case neighborhood phrase set becomes the phrase set (<sharp>, <keep>). (Here, terms are enclosed in angle brackets <> to facilitate distinction.)

By converting this set of case-sensitive words and phrases into a related case-related word set, which is another highly relevant word set, conversion of “action” can be realized. As for the hint presentation of "keep sharp edge", "adjust different directions", and "adjust different lengths", please refer to the (<sharp>, <keep>) It can be regarded as a conversion to a related {case-related phrase set}. (In this example, there is a one-to-one correspondence, but a one-to-many correspondence, that is, conversion to a plurality of word sets is also included in the conversion described here.)

  In addition, in addition to representing "object of action" (here, "object of movement", "object of mental activity", etc.), "square transition area", "temporal transition area" In some cases, it may be expressed. For example, <the river> of "swim across the river" and "the new year" of "I spent the New Year at home" can be cited (see Non-Patent Document 9 and the like). Therefore, the specification in the present invention may exclude such a specification.

  On the other hand, some are not qualified, but can be identified with qualified. For example, the passive voice "the problem was solved" can be rephrased as the active voice "the problem was solved", and the <problem> can be equated with the qualification <problem>.

As another example, the compound noun "learning the concept" can be paraphrased as "learning the concept", and the noun <concept> forming the compound noun can be equated with the ヲ case of <concept>. Furthermore, "introduction of technology" of the "N1 N2" type can be rephrased as "introduction of technology", and <technical> can be equated with <technical>.

As described above, there are various words and cases that can be identified with paraphrases in other words. Therefore, a case in which such a phrase or case is paraphrased may be included in the case in the present invention.

  As described above, there are various words and cases that can be identified with paraphrases by paraphrasing. It can be said that the case in the text including such paraphrases represents "object", and the case set near the case, which is a set of words located near the case, represents "action".

  Next, a description will be given of the conversion of the “action” using the information on the connection between the phrases, that is, the conversion of the ヲ case neighborhood phrase set included in the assignment text.

First, information regarding the connection between phrases will be described. There are various types of information about the connection between phrases. The information formats can be broadly divided into two formats, a local expression and a distributed expression.

  Vectors are often used for information in the local expression format and the distributed expression format. (Here, a vector represents a column vector or a row vector, and does not include an object representing linearity such as a matrix.) If two types are represented using a dimension which is a feature amount of the vector, The local expression form is an expression form in which one word corresponds to one dimension, and the distributed expression form is an expression form in which a plurality of words correspond to a plurality of dimensions.

  Information regarding the connection between words will be described in the order of the local expression form and the distributed expression form. First, information on the connection between words in the local expression form will be described.

Here, information on the connection between the two words Wi and Wj is expressed as inter-phrase connection information R (Wi, Wj). As the inter-phrase connection information R (Wi, Wj), the following various information can be used.

For example, for two phrases Wi and Wj, the presence / frequency of co-occurrence in the same sentence, and the presence / frequency of co-occurrence in a fixed-length window (word n-gram with window size of n words before and after) In addition, connection information between various phrases such as presence / absence / frequency of a dependency relationship and presence / absence / frequency of a specific dependency relationship can be used as R (Wi, Wj).

  Further, a matrix having the information R (Wi, Wj) as an (i, j) component can be considered. For example, assuming that the frequency of co-occurrence of two words Wi and Wj in the same sentence is R (Wi, Wj), it is called a co-occurrence matrix of words. If there are n types of words, this matrix is an n-th order square matrix. Information can be stored and used as such a matrix format.

  However, as the types of phrases increase, the dimensions of the matrix increase, resulting in a large and sparse matrix (a matrix in which many of the components are 0). When dealing with a large and sparse matrix, there is a problem that the storage capacity and the amount of calculation increase. One way to solve this problem is to compress dimensions. Specifically, dimensional compression can be performed using singular value decomposition (SVD), principal component analysis (PCA), or the like.

  Information that has been dimensionally compressed in this way does not correspond to one word but one dimension, and a plurality of words corresponds to a plurality of dimensions, and can be said to be information in a distributed expression format. Methods for creating such information in the distributed representation format include methods such as Word2Vec and GloVe (see Non-Patent Documents 3 to 6) in addition to singular value decomposition (SVD) and principal component analysis (PCA).

  In the information in the distributed expression format described above, the information corresponding to each word is a vector. However, various types of information in a distributed representation format other than the vector can be used (see Non-Patent Document 10). Specifically, there are a matrix, a tensor, and a set of a vector and a matrix.

  Next, a method of converting “action” using connection information between these various phrases will be described. Specifically, the conversion of “action” is to convert a ヲ case neighborhood phrase set representing “action” in a task text into a related ヲ case neighborhood phrase set that is a highly relevant phrase set. That is, it can be said that a certain phrase set is converted into another related phrase set. First, a method of converting a phrase into another related phrase will be described, and then a method of converting a phrase set into another related phrase set will be described.

  A method for converting a phrase into another related phrase will be described below. First, information in the local expression format will be described, and then information in the distributed expression format will be described.

と表すことができる。 For example, if R (Wi, Wj) is the presence / absence of co-occurrence of two words Wi and Wj in the same sentence (Yes: 1, No: 0) as the local expression form inter-phrase connection information, Assuming that the set is Σ, a set A of terms co-occurring with the term Wi and a set B of terms co-occurring with the term Wj are obtained by using inter-phrase connection information R (Wi, Wj).

It can be expressed as.

  By using these word sets A and B, it is possible to calculate various indices of similarity and relevance (Cosine similarity, Jaccard coefficient, φ coefficient, etc.) shown in Non-Patent Document 8 and the like. Then, it can be said that the values of these indices represent the similarity and relevance of the words Wi and Wj.

を用いて算出することができる（非特許文献８など参照）。ここで、P(A)、P(B)は、それぞれ語句集合A、Bを事象と見做したとき、事象A、Bそれぞれの起こる確率である。また、P(A, B)は、事象AとBの同時確率（事象Aと事象Bがどちらも起こる確率）である。 For example, for Cosine similarity,

(See Non-Patent Document 8, etc.). Here, P (A) and P (B) are the probabilities of occurrence of events A and B, respectively, when word sets A and B are regarded as events. P (A, B) is the simultaneous probability of events A and B (the probability that both events A and B occur).

となる。ここで、|・|は集合の濃度（元の個数）である。 Also, if Equation (2) is expressed in set theory,

Becomes Here, | · | is the density (original number) of the set.

であり、標準内積<・,・>とユークリッドノルム||・||を用いてCosine類似度を表わせば、

となる。 Also, consider R (Wi, Wj) as the (i, j) component of the matrix M (co-occurrence matrix described above), and denote the i-th row vector and the j-th row vector of the matrix M by vectors x and y, respectively. Can be regarded as word vectors for the words Wi and Wj. If the matrix M is an n × n real square matrix, the vectors x and y are

And using the standard inner product <•, •> and the Euclidean norm || • ||

Becomes

  Further, the method of calculating the Cosine similarity using the inner product between the row vectors of the co-occurrence matrix M enables not only the presence / absence of co-occurrence but also co-occurrence frequency information. This can be realized by extending the value 0 or 1 (presence information) of the component of the co-occurrence matrix M to a non-negative integer (frequency information).

  Up to this point, co-occurrence information of words in the same sentence has been described as the inter-phrase connection information R (Wi, Wj). The inter-phrase connection information R (Wi, Wj) can take various information as described above in addition to such co-occurrence information. For example, it can be used as the inter-phrase connection information R (A, N) as the presence / absence of a modification / modification relation of <Noun modification A>-<Noun N> (Yes: 1, No: 0). (Here, the <unit modifier> includes <unit modifier>, and the <noun phrase> includes <noun>.)

In the case of the above-mentioned task text "Keep a sharp blade", <Sharp>-<Blade> is word-to-phrase connection information, and the hints "Adjust different directions" and "Adjust different lengths"<Different>-<Orientation> and <Different>-<Length> are inter-phrase connection information. When this is expressed using the inter-phrase connection information R (A, N), the expression (6) is obtained.

  For example, the similarity and the relevance of the <unit modification unit> can be calculated using such inter-phrase connection information. Assume that when this type of inter-phrase connection information R (Ai, Nj) is collected, there are m types of <continuous modifier A> and n types of <noun phrase N>. If this inter-phrase connection information R (Ai, Nj) is the (i, j) component of the specific dependency relationship matrix M, M is an m × n execution sequence. As described above, the i-th row vector of the specific dependency relation matrix M is a word vector of the <continuous modifier Ai>, and the j-th row vector is a word vector of the <continuous modifier Aj>. > And <union modification unit Aj> can be calculated using various indices.

  There is a relation that can be identified in another expression form, such as the relation between <modification part> -modification / modification of <noun phrase> and <ヲ case> and <verb phrase>. (Here, the <verb phrase> includes the <verb>.) For example, the compound noun “new face writer” can be paraphrased as <new face writer> − <writer>, and this relationship can be restored. In other words, there is a relationship of <modification part>-<noun phrase> modification / modification. Therefore, such a relationship may be included in the relationship between <modification unit>-<noun phrase> in the present invention.

  As described above, the case where the relation between the modification and the modification of <noun phrase>-<noun phrase> is used as inter-phrase connection information has been described. In this manner, the relation between specific types of words is used as inter-phrase connection information. Besides, it is also possible to use a plurality of relations between various phrases.

  For example, in a relationship of text of the type <continuous modifier Ad> <continuous modifier A> <noun phrase N> such as <very> <clean> <flower>, <continuous modifier Ad>-< It is also possible to consider a relation such as "continuous modifier A" or "continuous modifier Ad"-<noun phrase N>.

Also, in relations to texts of the type <ヲ case N> <continuous modifier Ad> <verb phrase V>, such as the texts <task> <fundamental> <solve>, <ヲ case N> In addition to the relationship <-verb phrase V>, a relationship << case N>-<continuous modifier V> can also be used.

Furthermore, it is like a complex structured text of <cause> <clarify> <thing> <by> <task> <fundamental> <solve> <ヲ case N1> <verb phrase V1> In relations in texts of the type <koto> <by> <ヲ case N2> <adverb phrase Ad> <verb phrase V2>, the relationship <ヲ case N1> − <verb phrase V2> or <verb phrase V1> More specifically, a relationship such as-<ヲ N2> or the like, specifically, <cause>-<solve> or <clarify>-<assignment> can be used.

Also, besides the relation of the type <ヲ case N1> <verb phrase V1> <thing> <by> <ヲ case N2> <adverb phrase Ad> <verb phrase V2>, <thing> <by> < There are also texts with similar meanings that are replaced with <to> or <to>. Specifically, it corresponds to "solving the problem fundamentally by clarifying the cause" and "solving the problem fundamentally by clarifying the cause". Such <verb phrase V1>-<@ case N2> relations in different types of texts having the same meaning may be included as the same kind of relation.

  Cosine similarity and Jaccard coefficient are defined by symmetric expressions for two variables, and the index is symmetric. However, in general, when considering relevance, it may be necessary to consider its asymmetry. For example, when considering the "action" of "applying a patent" and the "action" of "submitting documents", the "action" of "applying a patent" is related to the "action" of "submitting documents". However, it is generally difficult to say that the "action" of "submitting documents" is related to the "action" of "applying a patent". When considering relevance in such a case, a relevance index that can be evaluated including its asymmetry can also be used.

となる。この場合、全体集合の濃度が大きくなると、全体集合に対する各集合の濃度比が一定であっても、補完類似度の値も大きくなるという性質を持っているので、例えば、全体集合の濃度で除して式（８）のような指標を用いてもよい。

Indices indicating asymmetric similarity / relationship include complementary similarity (see Non-Patent Document 7 and the like). If the complementary similarity is expressed by using the density (the original number) of the sets A and B used in Expression (3),

Becomes In this case, if the density of the whole set increases, the value of the complementary similarity also increases, even if the density ratio of each set to the whole set is constant. Then, an index such as Expression (8) may be used.

  This complementary similarity was originally proposed for recognizing deteriorated printed characters having "blurred" or "dirty", and a case where the similarity was applied to the similarity of a one-to-many relationship is described in Non-Patent Document. 7 and the like. Due to the nature of the complementary similarity, in the present invention, when the complementary similarity is used as an index of the conversion from the phrase represented by the set A to the phrase represented by the set B, the sets A and B can be used interchangeably. This replacement corresponds to replacing b and c in equations (7) and (8).

  In the above, the index regarding the relevance between words and phrases using the word-phrase connection information in the local expression form has been described. Based on these indices, it is possible to select another phrase candidate highly relevant to a phrase.

  Next, a method for converting a phrase using connection information between phrases in a distributed expression format into a related phrase will be described. In many pieces of word-to-phrase connection information in a distributed expression format, the type of the information is a vector. Hereinafter, a case where the type of information is a vector will be described as an example.

  Since the information in the distributed expression format is a vector, the relevance can be evaluated using the Cosine similarity of Expression (5) described in the case of the information in the local expression format. Also, the norm between words and phrases in the vector space can be used as an index of relevancy. Although the Euclidean norm is used for the Cosine similarity in Expression (5), this corresponds to the second-order average norm, and in general, a p-order average norm or the like can be used. Various other norms are also available.

  Next, a method of converting a phrase set using connection information between phrases in a local expression form into a related phrase set will be described.

  The index of similarity / relevance between phrases described above can also be used to evaluate the relevance between phrase pairs. For example, when evaluating the relevance of a phrase pair Pi = (Wi1, Wi2) and Pj = (Wj1, Wj2), the phrase phrase Pi = (Wi1, Wi2) co-occurs with both phrases Wi1 and Wi2 Assuming that a set of phrases is A and a set of phrases co-occurring with both of the two phrases Wj1 and Wj2 of the phrase pair Pj = (Wj1, Wj2) is B, the above-described equations (2), (7), and (8) ) Can be used to represent similarity or relevance.

  In addition, for the three or more (n) phrase sets, the above-mentioned index regarding the similarity / relevance between the phrases can be used in order to evaluate the similarity and the relevance between the phrase sets. For example, when evaluating the association with the word sets Ti = (Wi1, Wi2,…, Win) and Tj = (Wj1, Wj2,…, Wjn), by selecting two words from these word sets, By the method, the similarity / relevance of two phrase pairs can be evaluated. Therefore, by selecting appropriate plural pairs of phrases from the phrase sets Ti and Tj and comprehensively evaluating the similarity / relevance index, the similarity can be obtained for three or more (n) phrase sets. And relevance can be evaluated.

  It is known that in the case of information in a distributed expression format such as Word2Vec, the analogy relationship between words can be associated as "translation" in a space based on the distributed expression information of words (hereinafter, referred to as a distributed expression space). (See Non-Patent Document 4, etc.). Therefore, when such information in the distributed representation format is also used, an index of parallel movement can be used in addition to the various indexes indicating the relevance described above.

  The index of parallel movement will be described using a simple example. The task text "Keep a sharp blade" can be generalized as <adjunct modifier A> <noun phrase N> <to> <verb phrase V>. In simple terms, the assignment text is <A> <N> <V>. A text having a relation converted from the task text <A> <N> <V> is expressed as <A '> <N'> <V '>, and the image of the conversion is expressed in a distributed expression space as a plane. The mobility can be expressed in a form as shown in FIG.

  In FIG. 7, <A>, <N>,..., <V ′> represent information in a distributed expression form of a corresponding word, that is, a word vector. Strictly speaking, it means that a symbol such as <A> is arranged at the position of the end point when the viewpoint of the phrase vector is set to the origin of the distributed expression space. Assuming that vectors corresponding to the words <A>, <N>,..., <V ′> are represented by a, n,..., V ′, the parallel movement is represented by <A> → <A ′>, < It can be said that three arrows of N> → <N ′> and <V> → <V ′> indicate the coincidence of three vectors a′−a, n′−n, and v′−v. You. That is, it can be said that it is an index for selecting the conversion candidate <A '> <N'> <V '> so that the three vectors a'-a, n'-n, and v'-v are aligned as much as possible.

  As described above, the present invention focuses on "action" rather than "object" described in an assignment text, and uses the information on the connection between words to associate a phrase set included in the assignment text. This is a method of converting into a related phrase set that is a highly likely phrase set, and particularly includes a method that takes into account, for example, co-occurrence between words and parallel translation in a space based on distributed expression information as an index. By using the present invention, it is possible to provide a flexible, wide, and effective hint without depending on the knowledge and skills of the user.

Next, a plurality of embodiments of the present invention will be described with reference to the drawings.
[Thinking support system (1)]
First, a thought support system according to the present embodiment will be described with reference to FIGS.
FIG. 1 is a block diagram showing the configuration of the thinking support system. The thinking support system 10 is an information processing system that performs information processing using a natural language by a computer as a whole, has a relationship with a part or all of an input text to be thought, and performs a structural analysis of the text. Using the information on the connection between phrases, the related phrase pair, which is a phrase pair located in the vicinity of the phrase included in the obtained text, is converted into a related phrase pair, which is a highly relevant phrase pair, using information on the connection between the phrases. A function that supports thinking by converting and presenting a related text obtained by converting the case into a related case, which is a highly relevant phrase, by using the information of the related case neighborhood phrase set. Have.

The thinking support system 10 according to the present embodiment includes an operation input unit 11, a screen display unit 12, a communication I / F unit 13, a storage unit 14, an arithmetic processing unit 15, and a database unit 16 as main functional units. Have been.

The operation input unit 11 includes an operation input device such as a keyboard, a touch panel, and a voice recognition device, and has a function of detecting a user operation and outputting the operation to the arithmetic processing unit 15. The screen display unit 12 is composed of a screen display device such as an LCD, and has a function of displaying various information on the screen according to an instruction from the arithmetic processing unit 15. In addition, an output device such as a speech synthesizer can be provided in addition to the screen display device. The communication I / F unit 13 includes a dedicated circuit for data communication, and has a function of performing data communication via a LAN line, a wireless line, or the like.

The storage unit 14 is composed of a storage device such as a hard disk or a semiconductor memory, and has a function of storing processing information 14A and a program 14P used for various processes in the arithmetic processing unit 15. The program 14P is a program that, when executed by the arithmetic processing unit 15, implements various processing units in cooperation with the arithmetic processing unit 15, and is read in advance from an external device or a storage medium (both not shown). And stored in the storage unit 14.

The database unit 16 is composed of a storage device such as a hard disk or a semiconductor memory, and has a function of storing a database used for various processes in the arithmetic processing unit 16. The main database stored in the database unit 16 is an inter-phrase connection information database that stores information on the connection between words and phrases. The various types of inter-phrase connection information described in [Principle of the Invention] are stored in the database unit 16.

  FIG. 4 shows an input screen display example of the thinking support system, and FIGS. 8 and 9 show an output screen display example of the thinking support system. FIG. 5 is a flowchart showing the processing of the thinking support system. Hereinafter, description will be made along each step of the flowchart of FIG.

  First, step S10 of the flowchart of FIG. 5 “obtain a text to be considered as input text” will be described. FIG. 4 shows an input window for a text (task text) to be considered displayed on the screen display unit 12 of the thinking support system 10. In this input window, a sentence “Keep the sharp blade”, which is a text to be thought input by the keyboard or the like of the operation input unit 11, is displayed. When the related text display button on this screen is selected by clicking or the like using the mouse or the like of the operation input unit 11, the input processing unit 15A of the arithmetic processing unit 15 receives the input text “Keep sharp blade” and stores the input text. 14 is stored.

  Next, step S11 “perform morphological analysis and structural analysis of input text” and step S12 “acquire each case and each case close phrase set” in the flowchart of FIG. 5 will be described. After step S10, the structural analysis unit 15B uses the natural language processing technology such as morphological analysis and structural analysis to analyze the input text to be considered, so that <sharp>-<blade>, <blade>- The relationship between the phrases <keep> is extracted, and <case> is used as the ヲ case, <blade> is used as the noun phrase included in the ヲ case, and ヲ case is used as a pair of words (<sharp>, <keep> ) Are specified and stored in the storage unit 14. When there are a plurality of ヲ cases, a ヲ case neighborhood word set or the like corresponding to each ヲ case is specified and stored in the storage unit 14.

  Next, step S13 of the flowchart in FIG. 5, “Converting each set of ヲ case neighboring words and phrases to a related ヲ case neighborhood word and phrase set” will be described. Since the related phrase processing unit 15C performs the conversion by focusing on the “action” rather than the “object”, first, the related case phrase pair (<sharp>, <keep>) is highly related to the related {case nearby phrase. A process of converting to a pair (<different>, <match>) or the like is performed. The conversion process to the related ヲ case neighborhood phrase set is performed by any one of the <sharp>, <keep>, or (<sharp>, <keep>) pair or the phrase set in the ヲ case neighborhood phrase set. Is executed first.

Here, an example will be described in which the related phrase processing unit 15C first acquires a phrase having a high relevance for a <verb phrase> of <keep> included in the #case neighborhood phrase set. At this time, the inter-phrase connection information database 16A includes local expression information LRD, which is information in a local expression format regarding the presence or absence of a <adjunct modifier>-<noun phrase> relationship and a <ヲ case>-<verb phrase> relationship, It is assumed that at least at least distributed expression information DRD, which is information of a distributed expression format of a phrase corresponding to the <nominal modifier>, <noun phrase>, <verb phrase>, is stored.

Here, a set including all the phrases of the <union modifier> stored in the local expression information LRD as elements is included in the set A, <noun phrase>, and <ヲ case> (exactly, in <ヲ case>). Assuming that a set of all noun phrases as elements is a set N and a set of all <verb phrases> as elements is a set V, the information on the <adnominal modifier>-<noun phrase> relationship is set A. It can be represented as a subset of the direct product set A × N of the set N, and the information on the <ヲ case> − <verb phrase> relation can be represented as a subset of the direct product set N × V of the set N and the set V.

Also, when a subset of the direct product set A × N is represented using a matrix, the element of the matrix corresponding to the element of this subset is set to 1 and the other components are set to 0, thereby defining the <union modification unit. | A | × | N | matrix representing the relation>-<noun phrase> can be generated. Similarly, a | N | × | V | matrix representing a <ヲ case>-<verb phrase> relationship can be generated. (Where | · | represents the density (original number) of the set.)

  First, the related phrase processing unit 15C searches for <ヲ case>-<verb phrase> related information of the local expression information LRD stored in the inter-phrase connection information database 16A, and states that <verb phrase> = <keep>. Acquire a set N (<keep>), which is a set based on (included in) << case> that satisfies the condition. This set N (<keep>) is, for example, a set composed of elements such as N (<keep>) = {<identity>, <brightness>, <accent>, ...} . Note that there is also a meaning to simplify the processing, and the set of <ヲ case> is based on <noun phrase> included in <ヲ case> instead of <ヲ case>. Assuming that <keep> is the i-th component when the set V is regarded as an ordered set, the set N (<keep>) corresponds to the column vector of the i-th column of the | N | × | V | matrix.

A set N (<V>) having a high relevance to this set N (<keep>) is extracted using <ヲ case> − <verb phrase> relation information. Using a | N | × | V | matrix, it can be said that this is a process of searching for another column vector having a high similarity to the column vector of the i-th column. A specific method of searching for a column vector having a high similarity / relevance is as described in [Principle of the Invention], and various methods can be used.

  When searching for a set N (<V>) with a high degree of relevance to the set N (<keep>), if a set N (<V>) with too high similarity is selected, it becomes too obvious as a hint. It is highly possible that Therefore, by setting the value of “max” of the “relevance threshold” shown in FIG. 4 to an appropriate value, it is possible to remove the set N (<V>) having too high a similarity. Also, if the relevance is too low, it can be said that the usefulness as a hint is likely to decrease. Therefore, by setting the value of “min” of the “relevance threshold” shown in FIG. 4 to an appropriate value, it is possible to remove the set N (<V>) having too low relevance. This value is a value calculated using, for example, the complementary similarity shown in Expression (8).

  Further, by changing the setting of the “max” and “min” values of the “parallel mobility threshold” shown in FIG. 4, it is possible to narrow down the range of the “parallel mobility” that is appropriate as a hint. As the index relating to the “parallel movement”, a value calculated by a Cosine coefficient or the like as shown in Expression (5) for a vector corresponding to the conversion shown in FIG. 7 can be used.

  When a set N (<V>) having a high similarity and relevance to N (<keep>) using the complementary similarity of Expression (8), for example, <enhance>, <given>, <change>, Verbs such as ... can be obtained as related ヲ case neighborhood phrases.

As described in [Principle of the Invention], as a method of converting a ヲ case neighborhood phrase set to a related ヲ case neighborhood phrase set, in addition to a method of calculating relevance in terms of phrases, the relevancy is obtained as a phrase set. A conversion method can also be used.

  A description will be given of an embodiment of a method for converting a case pair (<sharp>, <keep>) in a task text "Keep a sharp blade" into a related case pair.

  Similarly to N (<Keep>), a set of <Noun phrases> satisfying the condition <Continuous modifier> = <Sharp> is defined as N (<Sharp>) from the relation information of <Adnominal modifier>-<Noun phrase>. ). Also, a common part N (<sharp>) of a set N (<sharp>) of <noun phrases> to be modified by <sharp> and a set N (<keep>) of <noun phrases> that <keep> has as a ヲ case ) ∩N (<keep>) is represented as N (<sharp>, <keep>).

  A ヲ case neighborhood phrase set (<A>, <V>) having another N (<A>, <V>) with high similarity and relevance to this N (<sharp>, <keep>) It is a candidate for the related ヲ case neighborhood phrase set. Similarly to the case of N (<keep>) described above, when a related ヲ case neighborhood phrase set (<A>, <V>) having a high relevance is obtained by complementary similarity as shown in Expression (8), ( Candidates for related {case-nearby word pairs, such as <quiet>, <given>), (<different>, <fit>), (<different>, <live>), ..., etc. can be obtained.

Among such candidates, a set (<A>, <V>) in which <keep> → <V> and <sharp> → <A> are as close as possible to translation in the space based on the distributed expression information. >) Can also be selected (see FIG. 7). As an index for the selection, a method of evaluating the parallel shift to <A> based on the parallel shift to <keep> → <V> can be expressed as follows. You can also. Equation (9) expresses the similarity between the vector a0 + v1−v0 and the vector a1 using the Cosine coefficient. Therefore, it can be represented by using another index for calculating the degree of similarity between vectors.

  By using this Cosine coefficient, for example, the words “(quiet), <given>), (<different>, <fit>), (<different>, <live>), etc. As shown in the output screen display example of the thinking support system in FIG. 8, the Cosine coefficient is high (<different>, <match>), (<transmitted>, <given>), (<transmitted>, <learned> ),... Can be selected as a related ヲ case neighborhood phrase set. (This result is calculated using nwjc2vec (National Institute for Japanese Language: Distributed Expression Information) shown in Non-Patent Document 6.)

  Next, steps S14 “convert each case into related cases” and S15 “output converted text” in the flowchart of FIG. 5 will be described. Fig. 9 is an output screen display example in which, in the output screen display example of Fig. 8, a conversion candidate of a case when (<no>, (match)) is selected as a related no-case neighborhood phrase set is displayed. Related qualifications, which are candidates for qualification conversion, can be obtained as follows.

When (<No>, <Match>) is selected as the related {case} word pair, the candidates for the related {case} to convert {case} are set N (<no>) and set N (<match>) ), Which is a common part of N (<different>) ∩N (<match>), can be used as a candidate. For example, N (<different>) ∩N (<fit>) = ｛<item>, <value>, <opinion>, <position>, <image>, <color>,... Among the candidates, the parallel movement can be used as a condition in the same manner as the selection of the related-case-neighboring phrase set.

For example, in the above example of <sharp><holdblade><keep> → <different><N(a)><match>, two sets of <sharp> → <different> and <keep> → <given> The following Cosine coefficient can be used as an index for evaluating <the blade> → <N (（)> based on the parallel movement corresponding to the conversion (see FIG. 7). Equation (10) expresses the similarity between the vector n0 + ((v1-v0) + (a1-a0)) / 2 and the vector n1 using the Cosine coefficient. Therefore, it can be represented by using another index for calculating the degree of similarity between vectors.

  Using the Cosine coefficient of the equation (10) as an evaluation index and arranging the elements of N (<different>) ∩N (<fit>) in ascending order of the Cosine coefficient, for example, as shown in FIG. Related angle candidates such as angle>, <length>, <size>, <direction>, ... can be obtained. By selecting a phrase that is likely to be a hint for the task text “Keep the sharp blade” from among these related cases, for example, as shown in FIG. 2 and FIG. Adjust the length ".

  In the above example, in order to obtain a highly relevant word / phrase set candidate, the index based on the complementary similarity shown in Expression (8) as the index of the word / phrase relevance is used as the parallel movement index. The case where the Cosine similarity shown in (5), (9), and (10) is used has been described. However, as described in [Principle of the Invention], various types of indices can be used as indices indicating the relevance and parallel movement between word sets, and further, these indices are used as arguments. Can be used as a new index, and the present invention is not limited to this embodiment.

Further, in this embodiment, after first obtaining other <verb phrase> candidates having high relevance to <keep> which is <verb phrase>, the relevance to <sharp> which is <adjunct modifier> is obtained. We are seeking other candidates for <modification unit> that are high and have high parallelism, but we first find candidates for <modification unit>, and then have a relevance with <keep> which is a <verb phrase> Other <verb phrase> candidates that are high and have high parallel movement may be obtained. In addition, a candidate for another set highly relevant to the set (<sharp>, <keep>) of <adjunct modifier> and <verb phrase> is determined, and a set with high parallel movement is selected from the candidates. It may be. As described above, the procedure for obtaining the related phrase set candidate is not limited to the present embodiment.

[Thinking support system (2)]
Next, a thinking support system according to the present embodiment will be described with reference to FIG. In the embodiment [thinking support system (1)], the thinking target text could be input in the form of a sentence as shown in FIG. However, a user who is not accustomed to using the thinking support system may be able to freely input the thinking target text, which may make it difficult to create the thinking target text. As a method of solving the problem, for example, by providing an input screen as shown in FIG. 6 to the thinking support system, preprocessing of input text is simplified, and the user is not used to using the thinking support system. It is also possible to make it easier for the user to create a thought target text. It is almost the same as the embodiment [thinking support system (1)] except that the preprocessing of the input text is simplified.

[Expansion of Embodiment]
As described above, the present invention has been described with reference to various embodiments, but the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

10 thinking support system 11 operation input unit 12 screen display unit 13 communication I / F unit 14 storage unit 15 arithmetic processing unit 16 database unit

Claims (7)

Input processing means for acquiring text to be thought as input text;
First information storage means for storing information relating to a connection between phrases;
Using the information stored in the first information storage means, a set of phrases near the case, which is located near the case, included in the input text is a set of phrases having a high relevance. A related phrase set conversion process that converts the case into a related case that is a highly relevant phrase using the input text and the converted information on the related case nearby phrase set. Means,
Output processing means for outputting a set of phrases converted by the related phrase set conversion processing means,
A thinking support system comprising: The first information storage means includes information in a distributed expression form relating to a connection between words and phrases, and the related word / phrase set conversion processing means uses an index relating to parallel movement in a space based on the information in the distributed expression form. And converting a set of phrases near the case included in the input text, which is a set of phrases located in the vicinity of the case, into a set of words related to the case near the case, which is a set of highly relevant phrases. 2. The thinking support system according to 1. Input processing means for acquiring text to be thought as input text;
A second information storage unit that stores information on a connection between phrases including information on a connection between a ヲ case neighboring phrase and a noun phrase, which is a phrase located near the ヲ case,
Using the information stored in the second information storage means, an index relating to the cooccurrence of a noun phrase included in the case in the input text and a word near the case close to the case is also used. Thereby, a related phrase set conversion for converting a set of phrases close to the case included in the input text, which is a set of phrases located in the vicinity of the case, into a set of related relevant words near the case, which is a set of highly relevant phrases. Processing means;
Output processing means for outputting the phrase set converted by the related phrase set conversion processing means,
A thinking support system comprising: The first information storage means includes information on a connection between a noun phrase and a noun phrase near the noun phrase, which is a phrase located in the vicinity of the noun, and the related phrase set conversion processing means includes the information in the noun case in the input text. By using an index related to co-occurrence between a noun phrase and a phrase located in the vicinity of the case, the pair of phrases located in the vicinity of the case included in the input text can be related. The thinking support system according to claim 2, wherein the thought support system converts the phrase into a related ヲ case neighborhood phrase group that is a group of highly likely phrases. A thinking support program for causing a computer to function as each unit constituting the thinking support system according to claim 1. A computer-readable thinking support program recording medium storing the thinking support program according to claim 5. 5. Information relating to a connection between a noun phrase included in at least a ヲ case and a phrase located in the vicinity thereof, among the information stored in the first or second information storage means according to claim 3. Computer-readable thinking support information recording medium that stores the information.

Images