JP6543169B2 - Air category detection device, air category detection model learning device, method, and program - Google Patents

Description

  The present invention relates to an air category detection device, an air category detection model learning device, a method, and a program for detecting an empty category of input text.

  The empty category is an abbreviated pronoun called pro (or small pro) in linguistics, especially the generative grammar, a controlled unexplained subject called PRO (or big pro), T ( A null element (an element having no sound form) that expresses the trace of movement in WH questions, relative clauses, etc., which are called trace. The sky category is sometimes called a gap (gap).

  Empty category detection is a process of automatically identifying the position, type, and grammatical role of an empty category in an input sentence. In order to detect an abbreviated pronoun (pro), a subject such as English is required from a pro-drop language that omits an understandable subject from a context such as Japanese (does not allow omission of the subject) Important in translating into non-pro-drop language. For example, a subject that is empty in the sentence "I want to see a movie" is a speaker in normal interpretation. If you can analyze this relationship correctly, you can translate this sentence correctly as "I want to see a movie". Specifically, it is an analysis of "(pro) wants to watch a movie, pro = speaker".

  A control is a phenomenon in which a clause without an explicit subject (embedded clause such as infinitive) has the same co-reference relationship with the other clauses of that clause. It is. For example, in the sentence "I want to see a movie", the subject of the infinitive "to see a movie" is not specified, but matches the subject I of the verb want having the infinitive as a term. In English grammar that is learned in schools, this is expressed as "the subject of the meaning of the infinitive", but in generative grammar, PRO assumes the existence of empty category PRO like "I want (PRO to see a movie)". It is said that it is controlled by the subject I of the verb want.

  Detecting a controlled element (PRO) explains the substitution phenomenon of case particles in passive, causative, and receiving syntax, and helps to understand the situation correctly. For example, the sentence "Taro makes Hanako make curry" is an empty sentence of "making curry" like "Taro makes Hanako make (PRO makes curry), PRO = Hanako". The subject in question is controlled by the object of the sentence, "Taro made to Hanako." If you can analyze this relationship correctly, you can answer "Hanako" to the question "Who made curry?"

  Detecting the trace of movement (T) is important for the correct translation of interrogative sentences and relative clauses. For example, in the sentence "I sent him a picture I took yesterday", the object that is empty of the sentence "I took yesterday" is the destination of the relative clause. It is a noun "photograph". If this relationship can be analyzed correctly, it can be translated as "I sent him the picture I took yesterday." Specifically, the analysis is "I sent him a photo (I took yesterday (T)), T = photo".

  Because relative pronouns in English change personally like who and whom, it is also important to distinguish whether the void is the subject or the object in the relative clause of Japanese in the translation from Japanese to English It is. For example, in the sentence "I met a girl who speaks English well", the subject in the open space of the sentence "Speaks English well" is the noun "girl" who is the target of modification of the relative clause If you can analyze the relationship correctly and you have the knowledge that “girl” is “person”, then you will correctly generate a relative pronoun who who indicates that the subject is a person like “I met a girl who speaks English fluently”. it can. Specifically, it is an analysis of "I met a girl ((T) speaks English well), T = girl".

  In addition, in the adnomary modification clause of Japanese, there is a noun to which the modification destination of the adjoint modification clause is not a term of the predicate of the adjoint modification clause (there is no empty space). Called "embedded phrase". For example, "I sent him a picture (with a child swimming)".

  The relative clauses of the English language can only be translated into English because the relative clauses of the English language are only permitted if the noun to which they are modified is a term of the predicate of the relative clause. Some ingenuity such as changing Japanese or English expressions or supplementing words is necessary. For example, "I sent him a picture which shows that a child swam" (= I sent him a picture showing that a child is swimming).

  Therefore, in machine translation from Japanese to English, it is also important to identify that there is no space in the adnominal modification section (it is an external relation) by means of air category detection processing.

  Most conventional statistical learning based parsers ignore empty categories. Although Penn Treebank, which is the correct data for typical parsing in English, is annotated about controlled element PRO and trace T of movement, most parsers developed so far are empty. Output a labeled syntax tree ignoring categories.

  Also, the conventional air category detection method implements air category detection as preprocessing for syntactic analysis, a method for implementing air category detection as post processing for syntactic analysis, and simultaneously performs air category detection and syntactic analysis or integrated There are three major ways to do this.

  As a conventional sky category detection method, there is known a method of statistically detecting a void based on a pattern based on a path from a void node to its antecedent (Non-Patent Document 4). The method is a method in which air category detection is implemented as post-processing on the output of the parser.

  Also, a trace tagger that predicts an empty category from words and parts of speech before and after is known (Non-Patent Document 3). The method is a method in which air category detection is realized as preprocessing of syntactic analysis.

  There is also known a method of integrating empty category detection and syntactic analysis by performing syntactic analysis using a lattice that expresses the possibility that empty categories appear between words as an input (Non-Patent Document 6). ).

  In addition to PRO and T, the Chinese penn Treebank, which is the correct data of typical syntactic analysis in Chinese, is annotated on the omitted pronoun pro. The research on sky category detection for the abbreviated pronoun pro is prosperous with the release of the Chinese Penn Treebank.

  In addition, while the empty category detection method in English used only phrase structure as the syntactic structure (output format of the parser), the Chinese empty category detection method used the phrase structure as the syntactic structure There are known a method of using and a method of using a dependency structure.

  In addition, there is known a machine learning algorithm that formalizes sky category detection as a classification problem in which labels are assigned to pairs of positions and modifications of sky categories in a sentence dependent structure (non-patent document) 8).

  In addition, there is known a machine learning algorithm that formalizes empty category detection as a classification problem that labels an inflection phrase (IP, in the phrase structure of a sentence) and realizes this classification (Non-Patent Document 7) ). Note that the non-terminal symbol IP representing the inflection phrase in the Chinese Penn Treebank substantially corresponds to the non-terminal symbol S representing a sentence in the Penn Treebank.

  In addition, no information on sky category is given to "Kyoto University Text Corpus" which is the correct data of typical syntactic analysis in Japanese. In addition, there is a "Keyaki Treebank" to which an annotation on an abbreviated pronoun pro and a trace T of movement is added. However, there is no report that the sky category detection method devised for English and Chinese has been applied to Japanese, as far as the inventors know.

  Also, a method similar to the above-mentioned Non-Patent Document 4 in the Japanese lexical analyzer HARUNIWA (http://www.compling.jp/haruniwa/index.html) which uses a statistical model learned from the Sakai Tree Bank. There is known an air category detection method for detecting an air category according to manually generated pattern rules. However, there is no document that reports the content and accuracy of this implementation, to the best of the inventors' knowledge.

  Also, as a technique similar to Japanese sky category detection, a technique called Japanese predicate term structure and zero pronoun complementation (or zero pronoun analogy) is known. Predicate term structure analysis is a task of extracting a predicate and its terms from a sentence. If there is no term in the sentence, such as relative clauses or abbreviated pronouns (zero pronouns), predicated term structure analysis and sky category detection should solve similar problems. The difference between predicate term structure analysis and empty category detection is that while empty category detection detects the position of empty space in the sentence, there is no concept of empty space in predicate term structural analysis, empty category detection is in the sentence In addition to detecting the existence of vacancies in the domain, predicated term structure analysis includes processing to identify the reference destination of the vacancies, and the detection of vain categories is based on the theory of generative grammar and controls. The pronounial element PRO with the obligatory anaphorism called is distinguished from the abbreviated pronoun pro, but predicate term structure analysis differs in that both are collectively called zero pronouns.

Alastair Butler, Tomoko Hotta, Ruriko Otomo, Kei Yoshimoto, Zhen Zhou, and Hong Zhu. Keyaki treebank: phrase structure with functional information for japanese. Text Annotation Workshop, 2012. Alastair Butler, Shota Sasayama, Kei Yoshimoto. Coindexed null elements for a japanese parsed corpus. Proceedings of the 21st Annual Conference of the Association for Language Processing, pp. 708-711, 2015. Psyneter Diens and Amit Dubey. Deep syntactic processing by combining shallow methods. In ACL-2003, pp. 431-438, 2003. Mark Johnson. A simple pattern-matching algorithm for recovering empty nodes and their an-tecedents. In ACL-2002, pp. 136-143, 2002. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In EMNLP-2014, pp. 1532-1543, 2014. Cai Shu, David Chiang, and Yoav Goldberg. Language-independent parsing with empty ele-ments. In ACL-2011, pp. 212-216, 2011. Eninging the ghost: Modeling empty categories for machine translation. In ACL-2013, pp. 822-831, 2013. Bing Xiang, Xiaoqiang Lue, and Bowen Zhou. Dependency-based empty category detection via phrase structure trees. In NAACL-HLT-2013, pp. 1051-1060, 2013. Nianwen Xue and Yaqin Yang. Daisuke Kawahara, Sadao Kurohashi. Incremental Automatic Construction of Case Frame Dictionary. Natural Language Processing, Vol. 12, No. 2, pp. 109-131, 2005.

  However, the process corresponding to the empty category detection in the conventional predicate term structure analysis described above is not in the grammatical knowledge inside one sentence, but in the context or area, in order to try to determine the zero pronoun's reference destination. There is a problem that it can not be realized as a process closed in one sentence because it requires dependent knowledge.

  In addition, the Japanese sky category is an established concept in linguistics, and there are applications as described above. However, as a technique for automatically performing airspace detection in natural language processing for Japanese, as far as the inventors know, there is only the rule-based airspace detection attached to the Japanese parser HARUNIWA. Also, as described later, the accuracy is 62.6% even when given the correct syntax analysis, and there is a problem that it can not be said to be high.

  Also, the conventional analysis technology for zero pronouns in Japanese is a technology that is not similar to sky category detection. These are considered to be part of or similar to predicate term structure analysis, treating cases where terms related to predicates exist and do not exist in sentences (zero pronouns) in the same line, and identifying reference destinations of zero pronouns Including up to. Therefore, there is a problem that the technology is not necessarily optimized for the detection of Japanese sky categories including zero pronouns.

  In addition, although air category detection for English and Chinese has been relatively well studied, it is not well understood whether the proposed technology is applicable to Japanese or to what extent it is effective. There is a problem of

  The present invention has been made to solve the above-mentioned problems, and an air category detection device, an air category detection model learning device, a method, and an air category detection device capable of accurately detecting an air category for Japanese text. The purpose is to provide a program.

  In order to achieve the above object, an air category detection device according to the first invention is an air category detection device for detecting an empty category which is a noun phrase generated by omission or movement from input text. Then, based on the phrase structure tree obtained by parsing the input text, for each inflection phrase (IP) node included in the phrase structure tree, a label of a child node of the inflection phrase (IP) node Using a child node feature based on a set of nodes and a path feature based on a series of labels of ancestor nodes from the inflection phrase (IP) node to the root node of the phrase structure tree or the complement indicator phrase (CP) node A feature vector creation unit for creating a feature vector, an empty category detection model for detecting an empty category label given to a refraction phrase (IP) node, which has been learned in advance, and the refraction phrase (IP) node Based on the feature vector of Te, it is configured to include a a sky category label detector for detecting the position and type of the air-categories in the input text.

  According to a second aspect of the present invention, there is provided an empty category detection method for detecting an empty category which is a noun phrase generated by omission or movement from input text including a feature vector creation unit and an empty category label detection unit. In the empty category detection device according to claim 1, wherein the feature vector creation unit is configured to include each phrase included in the phrase structure tree based on a phrase structure tree obtained by parsing the input text. A child node feature based on a set of labels of child nodes of the inflection phrase (IP) node, and a root node or a complement indicator phrase of the phrase structure tree from the inflection phrase (IP) node, for the inflection phrase (IP) node A feature vector is created using a path feature based on a series of labels of each of the ancestor nodes up to the CP) node, and the empty category label detection unit assigns it to a pre-learned infra-red phrase (IP) node Detect an empty category label And sky category detection model for, based on said feature vector of each of said refractive clause (IP) node, to detect the position and type of the air-categories in the input text.

  According to the first and second inventions, each inflection phrase (IP) node included in the phrase structure tree based on the phrase structure tree obtained by parsing the input text by the feature vector creation unit A child node feature based on a set of labels of child nodes of the inflection phrase (IP) node, and each of the ancestor nodes from the inflection phrase (IP) node to the root node of the phrase structure tree or the complement indicator phrase (CP) node An empty category detection for detecting an empty category label given to a refractive phrase (IP) node, which is prepared in advance by using an empty category label detection unit by creating a feature vector using a pass feature based on a series of labels Based on the model and the feature vector of each of the inflection phrase (IP) nodes, the position and type of the empty category in the input text is detected.

  Thus, based on the phrase structure tree obtained by parsing the input text, for each inflection phrase (IP) node included in the phrase structure tree, using the child node feature and the pass feature, Based on an empty category detection model for creating feature vectors and detecting an empty category label to be given to a refraction phrase (IP) node, which has been learned in advance, and a feature vector of each of the refraction phrase (IP) nodes By detecting the position and type of the empty category in the input text, it is possible to accurately detect the empty category for Japanese text.

  Further, in the first invention, the feature vector creation unit is configured to include, for each of the inflection phrase (IP) nodes, a foreword word feature based on a surface form of a foreword in the inflection phrase (IP) node, and the inflection phrase (IP The feature vector may be created by further using at least one of empty category labels based on empty category labels detected in a refraction phrase (IP) node that is an ancestor of the node.

  Also, in the first invention, the empty category detection model is created for each inflection phrase (IP) node included in the phrase structure tree of the text, for each of the plurality of texts to which the empty category label is attached. May be learned based on the feature vector.

  Further, in the first invention, for each inflection phrase (IP) node included in the phrase structure tree, a case relation between a distributed expression of a predicate as a leading word of the inflection phrase (IP) node and the predicate is specified. The distributed case frame creation unit may be further included which adds a vector obtained by normalizing the inner product of a possible predetermined distributed expression of words to the feature vector of the refractive phrase (IP) node as a distributed case frame feature. .

  Further, according to a third aspect of the present invention, there is provided an empty category detection model learning device, wherein, based on the phrase structure tree of the text, each of a plurality of texts to which blank category labels are given. A child node feature based on a set of labels of child nodes of the inflection phrase (IP) node for each inflection phrase (IP) node included in the phrase structure tree, and the inflection phrase tree from the inflection phrase (IP) node A feature vector creation unit for creating a feature vector using a sequence of labels of nodes of ancestor nodes up to a root node or a complement indicator phrase (CP) node of the above, and the empty category label of each of the texts And an empty category detection model for detecting the empty category label to be given to the inflection phrase (IP) node based on each of the text structure tree and the feature vector for each inflection phrase (IP) node included in each phrase structure tree of the text. Learn It is configured to include the model learning unit and, a.

  An empty category detection model learning method according to a fourth aspect of the present invention is the empty category detection model learning method in an empty category detection model learning device including a feature vector creating unit and a model learning unit, wherein the feature vector creating unit is For each of a plurality of texts with empty category labels, which are noun-like phrases generated by omission, or movement, per inflection phrase (IP) node included in the phrase structure tree based on the phrase structure tree of the text A child node feature based on a set of labels of child nodes of the inflection phrase (IP) node, and an ancestor from the inflection phrase (IP) node to a root node of the phrase structure tree or a complement indicator phrase (CP) node Creating a feature vector using a pass feature based on a series of labels of each of the nodes, and the model learning unit includes the empty category label of each of the texts and the phrase structure tree of each of the texts. And learn an air category detection model for detecting the air category label given to the inflection phrase (IP) node based on the feature vector of each inflection phrase (IP) node.

  According to the third and fourth inventions, each of the plurality of texts to which empty category labels, which are noun phrases generated by omission or movement, are given by the feature vector creation unit based on the phrase structure tree of the text. A child node feature based on a set of labels of child nodes of the inflection phrase (IP) node for each inflection phrase (IP) node included in the phrase structure tree, and a root node of the phrase structure tree from the inflection phrase (IP) node A feature vector is created using a path feature based on the series of labels of each node of ancestors up to a complement indicator phrase (CP) node, and the model learning unit determines each empty category label of the text and each of the text Based on the feature vector of each inflection phrase (IP) node included in the phrase structure tree, an empty category detection model for detecting an empty category label to be assigned to the inflection phrase (IP) node is trained.

  Thus, for each of a plurality of texts with empty category labels, which are noun-like phrases caused by omission or movement, inflected phrase (IP) nodes included in the phrase structure tree based on the phrase structure tree of the text Each time, a feature vector is created using child node features and pass features, each empty category label of the text, and a feature vector for each inflection phrase (IP) node included in each phrase structure tree of the text By learning an air category detection model for detecting air category labels given to inflection phrase (IP) nodes, it is possible to learn a model for detecting air category labels with high accuracy.

  Further, a program according to the present invention is a program for causing a computer to function as the above-described air category detection device or each unit constituting the air category detection model learning device.

  As described above, according to the empty category detection device, method, and program of the present invention, each refraction phrase included in the phrase structure tree is obtained based on the phrase structure tree obtained by parsing the input text. For the nodes, an empty category detection model for creating feature vectors using child node features and pass features and detecting an empty category label given to the inflection phrase node learned in advance, and an inflection phrase node By detecting the position and type of the empty category in the input text based on each feature vector, it is possible to accurately detect the empty category for Japanese text.

  Further, according to the empty category detection model learning apparatus, method, and program of the present invention, the phrase structure tree of text is provided for each of a plurality of texts with empty category labels, which are noun phrases generated by omission or movement. Based on, for each inflection node included in the phrase structure tree, create feature vectors using child node features and pass features, and create each empty category label of the text and each phrase structure tree of the text A model for detecting an empty category label with high accuracy by learning an empty category detection model for detecting an empty category label to be added to the inflection phrase node based on the feature vector for each included inflection phrase node I can learn.

It is a figure which shows an example of the syntactic tree of a persimmon tree bank. It is a figure which shows an example of the syntax tree after applying conversion. It is a figure which shows an example which grouped the feature into four types. It is a figure which shows an example of the precision of an empty category detection, and the result of a cutting test. It is a figure which shows an example of a pass feature. It is a figure which shows an example of a child node feature. It is a figure which shows an example of the combination of a pass feature and a head word feature. It is a figure which shows an example of the combination of a pass feature and a child node feature. It is a block diagram showing functional composition of an empty category detection model learning device concerning an embodiment of the present invention. It is a block diagram showing the functional composition of the air category detection device concerning the embodiment of the present invention. It is a flowchart figure showing the sky category detection model learning processing routine in the sky category detection model learning device according to the embodiment of the present invention. It is a flowchart figure which shows the air category detection processing routine in the air category detection device concerning an embodiment of the present invention. This is an example of the 欅 tree bank statistics. It is a figure which shows an example of comparison with the method in this embodiment, and the conventional method. It is a figure which shows an example of a resection test.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Overview of Embodiments of the Present Invention>
First, an outline of an embodiment of the present invention will be described. The present embodiment uses the keyaki tree bank described in Non-Patent Document 1. The Sakai Tree Bank is a tagged corpus in which phrase structures are added to Japanese sentences in accordance with the analysis rule of the pen passing corpus. In addition to the parallel structure, it uses a syntactic structure that is as flat as possible, such as not creating a verb phrase VP, and adds information on grammatical functions such as subjects and objects to noun phrases (NP) and clauses (IP, CP). There is a feature. FIG. 1 shows an example of a 欅 tree bank syntax tree for the sentence “I brought back my daughter who went home”.

  Here, IP is referred to as inflection phrase or temporal phrase, and CP (Complementizer Phrase) is referred to as complementary indicator phrase or complementary sentence. This is used to strictly distinguish "the degree of completeness as a sentence" with regard to one verb and the grouping of words that modify it, such as a "simple sentence", "a noun phrase", or a "clause" It is a term. Therefore, IP is a sentence that has missing elements like relative clauses, and CP is something that is almost completed as a sentence like an indirect interrogative sentence or citation clause (it is not wrong as an independent sentence as it is) is there.

  The 欅 tree bank is annotated as to the trace of movement (T) such as abbreviated pronouns (pro) and relative clauses. Although the annotation regarding the controlled empty element (PRO) is not given, it is also possible to give PRO automatically by the method described in Non-Patent Document 2. Although the sky category detection method for pro and T will be described later, sky tree detection for PRO can be realized in the same way by creating a tree bank with annotations for PRO.

  Here, techniques to be compared with the techniques used in the present embodiment will be described. First, the method of Non-Patent Document 7 which is considered to be the best method for detecting Chinese sky categories will be described. In the present embodiment, the phrase structure tree is a tree structure showing a hierarchical relationship with the types of words and phrases that constitute a sentence.

  In the method of Non-Patent Document 7, air category detection is formulated as a problem of classifying inflection phrase (IP) nodes of a phrase structure tree. For example, in FIG. 1, nodes corresponding to empty categories are deleted, and as shown in FIG. 2, an empty category label in which the type of empty category, grammar function, and position are encoded is added to the label of inflected phrase (IP) node .

For example, “IP-REL: ^* T ^* SBJ @ 0” means that the first child of the “IP-REL” (relational clause) node is a trace (T) and its grammar function is SBJ (subject) Represent. If no empty category exists, the empty category label is an empty string (it may be expressed by a special symbol such as Φ or NULL).

  Also, the problem of empty category detection is that empty category labels are removed from empty category labels, that is, the empty category label problem, since it is possible to uniquely decode empty category types, grammar functions, and locations from empty category labels attached to inflection phrase (IP) nodes. This results in the problem of detecting empty category labels to be assigned to inflection phrase (IP) nodes in conventional parsing results.

_Let T = t ₁ t ₂ ... T _n be a sequence in which nodes are traversed in the postorder from the root of a tree, and let an empty category label associated with t _i be e _i .

  The probability model of the method of Non Patent Literature 7 is formalized as a maximum entropy model (logistic regression model, log linear model) of the following equation (1).

  Here, φ is a feature vector, θ to φ is a weight vector, and Z is

Represents the normalization factor to be obtained as Here, ε represents a set of all empty category labels to be detected.

  In the method of Non-Patent Document 7, features are grouped into four types shown in FIG. They are tree label features, lexical features, empty category features and conjunction features.

  Since the feature of the method of Non-Patent Document 7 is directed to Chinese Penn Treebank, the following changes are made in this embodiment in order to apply the method to a 欅 tree bank.

  First, I changed the order of patrol from the postorder (bottom up) to the preorder (top down). Since PRO is not specified in the 欅 tree bank, air category detection in inflection phrase (IP) nodes depends on the result of air category detection of the upper level node and the left node. Empty category features are then calculated from ancestor nodes rather than descendent nodes.

  FIG. 4 shows the feature of the method of Non-Patent Document 7, the accuracy of detecting an empty category for a feature modified for Japanese, and the result of an ablation test for the feature after modification. From the results, it can be seen that the correction for Japanese language improves the accuracy of air category detection and that the combination feature is very effective.

  Next, features used in the present embodiment will be described. In this embodiment, the log linear model of the above equation (1) is used as the probability model, and the tree is traversed in descending order. As features of each inflection phrase (IP) node, there are three features of a head word feature, a child node feature, and an empty category feature and a path feature and Use a combination of

  Here, a path feature (PATH) is a sequence of non-terminal symbol labels for ancestor nodes from the current node to a root node or a complement indicator phrase (CP) node. For example, in FIG. 2, if the current node is "IP-REL", four paths shown in FIG. 5 are extracted.

  Also, the head word feature (HEAD) is a surface form of the word that is the lexical head of the current node. Child node feature (CHILD) is a set of labels of child nodes of the current node. This label is added with its surface shape if the rightmost terminal node is a function word. For example, in FIG. 1, if the current node is "IP-MAT", the head word feature is "follow" and the child node feature is as shown in FIG.

Also, an empty category feature (EC) is an empty category detected at a refractive phrase (IP) node that is an ancestor of the current node. For example, in FIG. 2, if the current node is "IP-REL", the empty category feature is ^* pro ^* .

  In the present embodiment, the head word feature, the child node feature, and the empty category feature and the pass feature are combined. For example, if the current node in FIG. 2 is "IP-MAT", the combination of the pass feature and the leading word feature is as shown in FIG. Further, the combination of the pass feature and the child node feature is as shown in FIG.

  Next, the distributed case frame used in the present embodiment will be described.

  Here, a case frame is a list of predicates, grammatical roles of their terms, and types of nouns that become terms. Also, a case frame dictionary (case frame lexicon) is a database in which case frames are described for each usage of a predicate.

  Also, since the case frame dictionary gives information on terms that can be taken by the predicate, it is clearly information effective for detecting empty categories, but in general, a large-scale and exhaustive case frame dictionary in the area to be processed is It is difficult to prepare.

  In the present embodiment, in a language such as Japanese in which a functional word (case particle) specifying a case relationship exists, an approximation of a case frame dictionary using a word distributed expression is used to detect an empty category. Use as a feature.

  Also, word embedding or distributed representation of words is a mapping from a word to a low dimensional real vector representing its meaning. In recent years, various methods for creating word embedding, such as dimensional reduction of neural networks and word co-occurrence matrices, have been proposed. Particularly, according to Non-Patent Document 5, the word embedding model Glove proposed by them is designed such that the dot product of two word embeddings approximates the logarithm of their co-occurrence frequency.

Therefore, if you use the property of Glove, you can easily create features that approximate the case frame of a predicate. Let w _{i be} a word distributed expression of a predicate, and q ₁ , q ₂ ,..., Q _N ∈Q, a word distributed expression of N case particles (and words corresponding to case particles). The distributed case frame (distributed case frame) is an N-dimensional vector obtained by normalizing the inner product of the word distributed expression of the predicate and the word distributed expression of each case particle, as in the following equations (2) and (3) It can be defined.

  Also, in the present embodiment, as a set Q of words corresponding to high frequency case particles and case particles, the words “but, also, of, to, from” are used. In addition, as a case particle equivalent word at the time of creating a distributed case frame, it is not restricted to the said example, You may use what is a word which has a possibility of demonstrating a case relationship in Japanese.

  Further, in this embodiment, in addition to the above-described feature, for each inflection phrase (IP) node, a distributed case frame of a predicate serving as a head word of the node is used as a feature.

  As described above, in the present embodiment, the method of Non-Patent Document 7 which is the most excellent in the prior art is reconsidered in the Chinese sky category detection which must detect the pronoun pro abbreviated like Japanese. Achieved the highest accuracy in Japanese.

  Specifically, instead of traversing the tree, the combination of three points of head word feature, child node feature, and empty category feature and pass feature is used as a feature, and instead of the case frame dictionary, The point of using a real number vector (distributed case frame) composed of the inner product of the word distributed expression of the particle equivalent term and the word distributed expression of the predicate as a feature is changed from the method of Non-Patent Document 7.

  Among the combinations of three of head word feature, child node feature, and empty category feature and pass feature, the most effective is the combination of pass feature and child node feature. In addition, since the distributed case frame (word distributed expression) can be easily calculated as compared with the case frame dictionary, field adaptation of the sky category detector is facilitated.

<Configuration of Air Category Detection Model Learning Device According to Embodiment of the Present Invention>
Next, the configuration of the air category detection model learning device according to the embodiment of the present invention will be described. As shown in FIG. 9, the air category detection model learning apparatus 100 according to the embodiment of the present invention stores a CPU, a RAM, a program for executing an air category model learning processing routine described later, and various data. And a ROM. This empty category detection model learning device 100 functionally includes an input unit 10, an operation unit 20, and an output unit 90, as shown in FIG.

  For example, as shown in FIG. 1, the input unit 10 receives a plurality of empty category-added phrase structure trees in an 欅 tree bank format, and stores the plurality of empty category added phrase structure tree data storage unit 22. Further, the input unit 10 receives, for example, a plurality of text data obtained by converting Japanese Wikipedia (registered trademark) into plain text, and stores the text data in the text data storage unit 24. In addition, the text data is not limited to what converted Japanese Wikipedia (registered trademark) into plain text, but may be any text, for example, Web crawl data, 30 years of newspaper articles, or all patent applications etc. You may use.

  Arithmetic unit 20 includes empty category added phrase structure tree data storage unit 22, text data storage unit 24, phrase structure tree conversion unit 26, empty category label storage unit 28, phrase structure tree data storage unit 30, word A distributed representation creating unit 32, a word distributed representation data storage unit 34, a distributed case frame creating unit 38, a feature vector creating unit 36, a model learning unit 40, and an empty category detection model storage unit 42 are provided.

  The empty category added phrase structure tree data storage unit 22 stores a plurality of empty category added phrase structure trees received in the input unit 10.

  The text data storage unit 24 stores a plurality of text data accepted by the input unit 10.

  The phrase structure tree converter 26 uses, for example, the phrase structure of FIG. 1 for each of the empty category phrase structure trees stored in the empty category phrase tree data storage unit 22 using the method of Non-Patent Document 7. In order to convert the tree into the phrase structure tree of FIG. 2, in the empty category added phrase structure tree, the nodes corresponding to the empty category are deleted, and in the label of each inflection phrase (IP) node, type of empty category, grammar Transform the phrase structure tree to add empty category labels that encode functions and locations.

  Also, the phrase structure tree conversion unit 26 stores each of the converted phrase structure trees in the phrase structure tree data storage unit 30, and is added to each inflection phrase (IP) node of the converted phrase structure tree. The empty category label is stored in the empty category label storage unit 28 in association with each inflection phrase (IP) node of the relevant phrase structure tree.

  Empty category labels are stored in the empty category label storage unit 28 in correspondence with each inflection phrase (IP) node of the transformed phrase structure tree.

  The phrase structure tree data storage unit 30 stores each of the converted phrase structure trees converted by the phrase structure tree conversion unit 26.

  The word distributed representation creating unit 32 creates a word distributed representation for each word according to the method of Non-Patent Document 5 based on each of the text data stored in the text data storage unit 24, and a word distributed representation data storage unit Store in 34.

  The word distributed expression data storage unit 34 stores word distributed expressions for each word.

  The feature vector creation unit 36 generates, for each inflection phrase (IP) node included in the phrase structure tree, a leading word feature, a child word feature, for each of the converted phrase structure trees stored in the phrase structure tree data storage unit 30. Leading word feature, child node feature, obtained for each inflection phrase (IP) node by acquiring three features of node feature and empty category feature, and also acquiring pass feature for each inflection phrase (IP) node A feature vector is created based on a combination of three features of sky category and pass feature, and is output to the distributed case frame creating unit 38.

  The distributed case frame creation unit 38 stores word distributed expression data for each of the inflection phrase (IP) nodes of the phrase structure tree for each of the transformed phrase structure trees stored in the phrase structure tree data storage unit 30. The above-described (2) is performed based on the word distributed representation of the predicate serving as the leading word in the inflected phrase (IP) node and the word distributed representation of N predetermined case particles, which are stored in the section 34. The distributed case frame is created as a feature according to equation (3) and equation (3) above. In addition, the distributed case frame creation unit 38 assigns the features of the distributed case frame of each inflection phrase (IP) node of the relevant phrase structure tree, created for each of the transformed phrase structure trees, to the corresponding inflection phrase (IP) node. It adds to the corresponding feature vector input from the feature vector creation unit 36 and outputs the feature vector to the model learning unit 40.

  The model learning unit 40 receives each of the feature vectors of each inflection phrase (IP) node included in the relevant phrase structure tree of each of the transformed phrase structure trees input from the distributed case frame creation unit 38 and the empty category label. An air category detection model for detecting an air category label is learned based on each of the air category labels corresponding to the inflection phrase (IP) nodes stored in the storage unit 28, and the air category detection model storage is performed. The data is stored in the unit 42 and output from the output unit 90.

<Configuration of Air Category Detection Device According to Embodiment of the Present Invention>
Next, the configuration of the air category detection device according to the embodiment of the present invention will be described. As shown in FIG. 10, the air category detection device 200 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM storing a program for executing an air category detection processing routine described later and various data. Can be configured on a computer. This air category detection device 200 functionally includes an input unit 210, an operation unit 220, and an output unit 290, as shown in FIG.

  The input unit 210 receives text to be detected as an empty category, and stores the text in the input text storage unit 222.

  The operation unit 220 includes an input text storage unit 222, a word distributed representation data storage unit 224, a syntax analysis unit 226, a phrase structure tree data storage unit 228, a feature vector creation unit 230, and a distributed case frame creation unit 232. , An empty category detection model storage unit 234, an empty category label detection unit 236, and an empty category label storage unit 238.

  The input text storage unit 222 stores the text to be detected received by the input unit 210.

  The word distributed expression data storage unit 224 stores word distributed expression data identical to the word distributed expression data stored in the word distributed expression data storage unit 34 of the empty category detection model learning device 100.

  The syntax analysis unit 226 parses the text to be detected, which is stored in the input text storage unit 222, acquires a phrase structure tree, and stores the phrase structure tree in the phrase structure tree data storage unit 228.

  The phrase structure tree data storage unit 228 stores a phrase structure tree obtained by parsing the text to be detected, which is input from the syntax analysis unit 226.

  For each phrase structure tree stored in the phrase structure tree data storage unit 228, the feature vector generation unit 230 generates, for each inflection phrase (IP) node included in the phrase structure tree, a leading word feature, a child node feature, And three features of sky category feature, and pass feature for each inflection phrase (IP) node, and three words of leading word feature, child node feature, and sky category feature and pass Create a feature vector based on the combination with the feature. The feature vector creation unit 230 creates feature vectors in order from the inflection phrase (IP) node closer to the root node of the phrase structure tree.

  The distributed case frame creation unit 232 stores, for each phrase structure tree stored in the phrase structure tree data storage unit 228, the word distributed expression data storage unit 224 for each inflection phrase (IP) node of the phrase structure tree. The above equation (2), based on the word distributed expression of the predicate that is the leading word in the inflected phrase (IP) node and the word distributed expression of N case particles defined in advance. A distributed case frame is created as a feature according to the above equation (3). Further, the distributed case frame creation unit 232 corresponds the feature of the distributed case frame of each inflection phrase (IP) node of the relevant phrase structure tree created for each phrase structure tree to the corresponding inflection phrase (IP) node. It is added to the feature vector input from the feature vector creation unit 230 and is output to the empty category label detection unit 236.

  In the sky category detection model storage unit 234, the same sky category detection model as the sky category detection model stored in the sky category detection model storage unit 42 of the sky category detection model learning device 100 is stored.

  The empty category label detection unit 236 stores the feature vector input from the distributed case frame creation unit 232, the empty category detection model stored in the empty category detection model storage unit 234, and the phrase structure tree data storage unit 228. An empty category label is detected for each of the inflection phrase (IP) nodes of the relevant phrase structure tree based on the existing phrase structure tree, and the phrase structure tree in which the detected result is attached to the corresponding node is the empty category The data is stored in the label storage unit 238 and output from the output unit 290.

<Operation of Air Category Detection Model Learning Device According to Embodiment of the Present Invention>
Next, the operation of the air category detection model learning device 100 according to the embodiment of the present invention will be described. First, the input unit 10 receives a plurality of 空 tree bank type empty phrase structure phrase trees in an 欅 tree bank format and stores the plurality of empty structure phrase structure tree data storage units 22. The input unit 10 receives a plurality of text data It is stored in the text data storage unit 24.

  Then, the CPU executes the program stored in the ROM of the air category detection model learning device 100, whereby the air category detection model learning processing routine shown in FIG. 11 is executed. It should be noted that before the sky category detection model learning processing routine shown in FIG. 11 is executed, the sky category detection model learning device 100 determines the non-patent document based on each of the text data stored in the text data storage unit 24. According to the method of 5, the word distributed expression for each word is created and stored in the word distributed expression data storage unit 34.

  First, in step S100, for each of the plurality of empty category-added phrase structure trees stored in the empty category-added phrase structure tree data storage unit 22, using the method of Non-Patent Document 7, the empty category added phrase structure In the tree, delete the node corresponding to the sky category, and add the phrase category tree to the label of each inflection phrase (IP) node by adding the sky category type, the grammar function, and the position encoded sky category label Convert.

  Next, in step S102, an empty category label is acquired from each of the converted phrase structure trees acquired in step S100, and is stored in the empty category label storage unit 28.

  Next, in step S104, for each of the transformed phrase structure trees acquired in step S100, for each inflection phrase (IP) node of the relevant phrase structure tree, 3 of the head word feature, the child node feature, and the empty category feature One feature is acquired, and for each inflection phrase (IP) node of the relevant phrase structure tree, a pass feature is acquired and a feature vector is created.

  Next, in step S106, the word distributed representation data storage unit 34 is stored for each of the inflection phrase (IP) nodes of the phrase structure tree, for each of the converted phrase structure trees acquired in step S100, The above equation (2) and the above equation (3) are based on the word distributed expression of the predicate that becomes the leading word in the inflected phrase (IP) node and each of the word distributed expressions of N predetermined case particles. Create a distributed case frame as a feature according to.

  Next, in step S108, the phrase structure tree obtained in step S106 is added to the feature vector for each inflection phrase (IP) node of the phrase structure tree for each of the converted phrase structure trees obtained in step S104. Add the features of the distributed case frame for each inflection clause (IP) node of, and create a feature vector.

  Next, in step S110, according to the feature vector for each inflection phrase (IP) node of the converted phrase structure tree acquired in step S108 and each of the empty category labels stored in the empty category label storage unit 28. Based on the empty category detection model for detecting the empty category label, the empty category detection model is learned and stored in the empty category detection model storage unit 42 and output from the output unit 90, and the empty category detection model learning processing routine is ended.

<Operation of the air category detection device according to the embodiment of the present invention>
Next, the operation of the air category detection device 200 according to the embodiment of the present invention will be described. First, the input unit 210 receives a text for which an empty category is to be detected, and stores it in the input text storage unit 222. Then, the CPU executes a program stored in the ROM of the empty category detection device 200. An empty category detection processing routine shown at 12 is executed.

  First, in step S200, a phrase structure tree to be parsed for the text to be detected stored in the input text storage unit 222 is acquired and stored in the phrase structure tree data storage unit 228.

  Next, in step S202, with regard to the phrase structure tree acquired in step S200, unprocessed inflection phrase (IP) nodes are processed in order from the one closest to the root of the phrase structure tree included in the phrase structure tree. For each inflection phrase (IP) node, select each of the three features of the prefix word feature, the child node feature, and the empty category feature acquired in step S104, and select the same as the inflection phrase (IP) node. The same feature as the pass feature obtained in step S104 is obtained for the inflection phrase (IP) node, and a feature vector is created based on each of the obtained features.

  Next, in step S204, for the inflection phrase (IP) node to be processed, a word variance representation of a predicate that is a leading word in the inflection phrase (IP) node stored in the word variance expression data storage unit 224; A distributed case frame is created as a feature according to the equation (2) and the equation (3) based on each of the word dispersed expressions of N case particles determined in advance.

  Next, in step S206, the feature vector of the distributed case frame of the processing target refractive phrase (IP) node obtained in step S204 is added to the feature vector of the processing target refractive phrase (IP) node obtained in step S202. , Create feature vector.

  Next, in step S208, the sky category detection model learned in step S110, which is stored in the sky category detection model storage unit 234, and the feature vector of the refraction phrase (IP) node to be processed acquired in step S206. And detect the sky category label of the inflection phrase (IP) node to be processed, based on the same sky category detection model.

  Next, in step S210, it is determined whether the processing from step S202 to step S208 has ended for all inflection phrase (IP) nodes included in the phrase structure tree acquired in step S200. If it is determined that the processing from step S202 to step S208 has ended for all inflection phrase (IP) nodes, the empty category label of each inflection phrase (IP) node acquired in step S208 is acquired in step S200 The phrase structure tree assigned to the phrase structure tree is stored in the empty category label storage unit 238 and output from the output unit 290, and the empty category label detection processing routine is ended. On the other hand, if the processing from step S202 to step S208 has not ended for all inflection phrase (IP) nodes, the process proceeds to step S202, and the inflection phrase (IP) node to be processed is changed, and step S202. The process from step S210 is repeated.

<Example of experiment>
The verification result about the case where the air category detection model learning apparatus and the air category detection apparatus which concern on this embodiment are used is demonstrated. The used persimmon tree bank (22,639 sentences) used the thing as of May 8, 2015.

  In order to balance the subject areas in the evaluation, a development set (1000 sentences) and a test set from three files of blog_KNB.psd (blog), spoken_CIAIR.psd (voice transcription), newswire_MAINICHI-1995.psd (newspaper article) (1003 sentences) was selected, and the remaining 20,646 sentences were set as a training set. FIG. 13 shows the number of characters, sentences, predicate numbers (number of inflection phrases (IP) nodes), abbreviated pronouns pro and marks of movement T according to grammatical roles with respect to the entire 欅 tree bank and test set.

  In addition, the word distributed expression used in the present experimental example was created by the method of Non-Patent Document 5 from Japanese Wikipedia (registered trademark) as of January 18, 2015 (about 6.6 million words, about 23.40 million sentences).

  Further, the accuracy of the air category detection apparatus according to the present embodiment was evaluated for the case where the correct answer of the syntax tree manually created is input and the case where the syntax tree output by the automatic parser is used. In the former, we used the tree tree's syntax tree with empty categories removed. In the latter case, in the output of the Berkeley parser model of the Japanese parser HARUNIWA (the Internet (URL: http://www.compling.jp/haruniwa/index.html)), immediately before the rule-based air category detection is executed. Used the syntax tree of

  As a rating scale, word-position-level identification metrics described in Non-Patent Document 7 were used.

  In order to evaluate the effectiveness of the distributed case frame used as a feature in the present embodiment, the Kyoto University Case Frame of Non-Patent Document 9 was used as an existing large-scale case frame dictionary. This case frame is automatically constructed from about 1.6 billion Japanese texts on the Web, and contains about 40,000 phrases.

  As a method of using the existing case frame, I tried three methods of BIN, SET and DIST. BIN expresses each case as a binary feature such as "ga" "to" "ni" for one predicate. SET takes as a feature a combination of cases of case frames such as "G", "G", "G", "G", and "G" to one predicate. DIST takes as its feature a real value obtained by normalizing the co-occurrence frequency of each case particle and predicate.

  The distributed case frame (DCF) used in the present invention can be considered as a smoothed DIST. FIG. 14 shows the accuracy of detecting the air category of the air category detecting device and the comparison method in the present embodiment under various conditions. In FIG. 14, “# non Z” indicates the number of features whose weight is not 0.

  When the correct syntax tree is used, the rule-based method (RULE) implemented by the parser HARUNIWA and the method of Non-Patent Document 7 modified for Japanese are respectively 62.6% in F-scale It was 68.6%. Among the proposed methods, the combination of path feature and child node feature (PATH CHILD) surpasses these two methods of becoming a baseline, achieving 73.2% when all features are used.

  When the parse tree outputs a parse tree, the F-scale of the proposed method drops from 73.2% to 54.7%. In most cases this is due to the parser failing to identify the inflection phrase (IP) node or to identify the grammatical function of the inflection phrase (IP) node.

  There is no significant difference in the three precisions of BIN, SET, and DIST that use the existing large-scale case frame as a feature, and the same improvement in precision as when using the distributed case frame is seen. That is, it is possible to replace a large case frame dictionary with a distributed case frame.

  FIG. 15 shows the result of the excision test of the feature of the proposed method. In this embodiment, a combination of three of head word feature, child node feature, and empty category feature and pass feature is used as the feature, and among them, the combination of pass feature and child node feature is the most effective. I know that there is.

  As described above, according to the empty category detection model learning device according to the embodiment of the present invention, the text of each of the plurality of texts to which the empty category label is assigned, which is the noun phrase generated by omission or movement Based on the phrase structure tree, for each inflected phrase node included in the phrase structure tree, a feature vector is created using a child node feature and a pass feature, and for each of the texts, a refraction included in the phrase structure tree of the text For each phrase node, a distributed case frame feature is added to the feature vector of the inflection phrase node, based on each empty category label of the text and the feature vector for each inflection node included in each phrase structure tree of the text The model can be learned with high accuracy by learning the empty category detection model for detecting the empty category label given to the inflection phrase node.

  Also, according to the empty category detection device according to the embodiment of the present invention, a child node is generated for each refraction phrase node included in the phrase structure tree based on the phrase structure tree obtained by parsing the input text. A feature vector is created using features and pass features, and a distributed case frame feature is added to the feature vector of the inflection phrase node for each inflection node included in the phrase structure tree, and the inflection phrase is learned in advance By detecting the position and type of the empty category in the input text based on the empty category detection model for detecting the empty category label given to the node and the feature vector of each of the refraction phrase nodes, Japanese It is possible to accurately detect the sky category for.

  In addition, Japanese sky categories can be detected with high accuracy compared to the conventional method. Furthermore, compared with the amount of text and the amount of calculation required to create a large case frame dictionary, the word distributed expression necessary to create a distributed case frame can be calculated with less amount of calculation from less text. It is easy to tune the model of sky category detection towards the area to be processed.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the scope of the present invention.

  For example, in the present embodiment, the case where the air category detection model learning device and the air category detection device are configured as separate devices has been described, but the present invention is not limited to this. For example, the air category detection model learning device and the air category detection device may be configured as one device.

  Further, in the present embodiment, a case is described in which a feature vector is created by using a combination of three of a head word feature, a child node feature, and an empty category feature and a pass feature in the feature vector creating unit as features. However, it is not limited thereto. For example, in the feature vector creation unit, a feature vector may be created by combining each of three arbitrary features of a head word feature, a child node feature, and an empty category feature with a pass feature as a feature.

  Moreover, in this embodiment, although the case where calculation of a distributed case frame was calculated for every process was demonstrated, it is not limited to this. For example, a distributed case frame may be calculated in advance for each predicate.

  Further, in the present embodiment, the case has been described where the feature vector created by adding the feature of the distributed case frame to the feature vector created by the feature vector creation unit is used, but the present invention is not limited to this. For example, feature vectors created by the feature vector creation unit may be used. In this case, the distributed case frame feature will not be used.

  Furthermore, although the present invention has been described as an embodiment in which the program is installed in advance, the program can be provided by being stored in a computer readable recording medium, and provided via a network. It is also possible.

10 input unit 20 operation unit 22 empty category added phrase structure tree data storage unit 24 text data storage unit 26 phrase structure tree conversion unit 28 empty category label storage unit 30 phrase structure tree data storage unit 32 word distributed expression creation unit 34 word distributed expression creation unit Data storage unit 36 Feature vector creation unit 38 Distributed case frame creation unit 40 Model learning unit 42 Air category detection model storage unit 90 Output unit 100 Air category detection model learning device 200 Air category detection device 210 Input unit 220 Operation unit 222 Input text storage Unit 224 word dispersed expression data storage unit 226 syntactic analysis unit 228 phrase structure tree data storage unit 230 feature vector creation unit 232 distributed case frame creation unit 234 empty category detection model storage unit 236 empty category label detection unit 238 empty category label storage unit 290 Output unit

Claims (9)

A sky category detection device for detecting a sky category which is a noun phrase generated by omission or movement from input text, comprising:
A child node feature based on a set of labels of child nodes of the inflection phrase node for each inflection node included in the phrase structure tree based on a phrase structure tree obtained by parsing the input text; A feature vector creation unit that creates a feature vector using a pass feature based on a series of labels of each of the ancestor nodes from the inflection phrase node to the root node of the phrase structure tree or the complement indicator phrase node;
An empty category detection model for detecting empty category labels to be given to inflected phrase nodes, which has been learned in advance, and the feature vector of the input text based on the feature vector of each of the inflected phrase nodes An air category label detection unit for detecting the position and type;
Air category detection device, including:   The feature vector creation unit adds, to each of the inflection phrase nodes, an affix word feature based on a surface form of a foreword of the inflection phrase node, and an empty category label detected in an inflection phrase node that is an ancestor of the inflection phrase node. The air category detection device according to claim 1, wherein at least one of the base category features is further used to create the feature vector. The air category detection model is
The blank according to claim 1 or 2, which is learned based on the feature vector created for each inflection node included in the phrase structure tree of the text for each of the plurality of texts with the blank category label attached. Category detection device.   For each inflection phrase node included in the phrase structure tree, a decentralized representation of a predicate that is the head word of the inflection phrase node and a decentralized representation of a predetermined word that may specify a case relationship with the predicate The air category detection device according to any one of claims 1 to 3, further comprising a distributed case frame creation unit that adds a vector obtained by normalizing an inner product as a distributed case frame feature to the feature vector of the refractive phrase node. For each of a plurality of refracted phrase nodes included in the phrase structure tree, based on the phrase structure tree of the text, for each of a plurality of texts with empty category labels that are noun phrases that are generated by omission or movement Child node features based on a set of labels of child nodes of phrase nodes and path features based on a series of labels of ancestor nodes from the inflection phrase node to a root node of the phrase structure tree or a complement indicator node A feature vector creation unit that creates feature vectors using
An empty for detecting the empty category label given to the inflection phrase node based on the empty category label of each of the text and the feature vector for each inflection node included in each phrase structure tree of each of the text A model learning unit for learning a category detection model;
An empty category detection model learning device, including: An empty category detection method in an empty category detection device for detecting an empty category which is a noun phrase generated by omission or movement from input text including a feature vector creation unit and an empty category label detection unit There,
The feature vector creation unit is configured to label child nodes of the inflection node with respect to each inflection node included in the phrase structure tree based on the phrase structure tree obtained by parsing the input text. A feature vector is created using a set-based child node feature and a pass feature based on the label sequence of each of the ancestor nodes from the inflection phrase node to the root node of the phrase structure tree or the complement indicator phrase node ,
The empty category label detection unit is configured to input the input based on an empty category detection model for detecting an empty category label to be given to the inflection phrase node, which has been learned in advance, and the feature vector of each of the inflection phrase nodes. An empty category detection method for detecting the position and type of the empty category in the specified text. An empty category detection model learning method in an empty category detection model learning apparatus including a feature vector creation unit and a model learning unit,
The feature vector creation unit is configured to, according to a phrase structure tree of the text, generate refractions included in the phrase structure tree for each of a plurality of texts with empty category labels, which are noun phrases generated by omission or movement. For each phrase node, a child node feature based on a set of labels of child nodes of the inflection phrase node, and a label of each node of an ancestor from the inflection phrase node to a root node of the phrase structure tree or a complement indicator phrase node Create feature vector using series based pass feature,
The model learning unit assigns the empty category label to the inflection phrase node based on the empty category label of each of the text and the feature vector for each inflection phrase node included in each phrase structure tree of the text. Learning the sky category detection model to detect the sky category detection model learning method. A program for a computer to function as an empty category detection equipment each part of any one of claims 1 to 4.   The program for functioning a computer as each part of the air category detection model learning apparatus of Claim 5.