JP2015153261A - Dialogue sentence generating device, dialogue sentence generating method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method which generates dialogue sentences by using data automatically acquired from a dialogue log corpus.SOLUTION: A dialogue sentence generation device 1 includes storage parts 17, 18 which store a graph Laplacian which is calculated according to the degree of association between a word generated on the basis of dialogue sentences that are extracted from a dialogue log corpus 30 and a sentence pattern, and a sentence pattern transition matrix among sentence patterns of the dialogue log corpus. When sentences are inputted, the dialogue sentences generation device 1 uses a sentence pattern transition matrix and obtains a sentence pattern of a response sentence corresponding to the sentence pattern of the inputted sentence, and obtains a feature word from words that are extracted as specific expressions from the inputted sentence. With the sentence pattern of the response sentence and the feature word as a seed, Laplacian label propagation is performed on the graph Laplacian, a word of high degree of association with the feature word is obtained, the obtained word is inputted in the sentence pattern of the response sentence, and a response sentence is generated.

Description

  The present invention relates to dialogue sentence generation in dialogue processing.

  Dialog generation is a process in which a machine automatically generates a response so that a natural conversation can be made with respect to an utterance made by a human, and is used in a voice dialog system such as Siri. Conventionally, sentence templates created by human hands have been used to generate dialogue sentences (Patent Documents 1 and 2).

JP 2010-72578 A JP 2004-110524 A

  However, the cost of generating templates by humans is high, so the types of sentences that machines can generate are limited. Thus, a technique for automatically acquiring a sentence template using a large amount of dialog logs and generating a dialog sentence is desired.

  Therefore, the present invention proposes a method for solving the problem of manual sentence templates, acquiring sentence templates from a large-scale dialogue log, and automatically acquiring a word set that fills a slot of the template.

  The dialog sentence generation apparatus of the present invention includes a graph Laplacian calculated from the degree of association between a word and a sentence pattern generated based on a dialog sentence extracted from the dialog log corpus, a transition matrix between sentence patterns of the dialog log corpus, A storage unit that stores the sentence, a sentence input unit that receives a sentence input, a specific expression extraction unit that extracts a specific expression from the input sentence, a pattern determination unit that determines a sentence pattern of the input sentence based on the extracted specific expression, The transition matrix stored in the storage unit is read, and the pattern seed extraction unit that extracts the pattern seed of the response sentence corresponding to the sentence pattern of the input sentence using the transition matrix is extracted by the specific expression extraction unit. A feature word extraction unit that extracts a feature word from the specific expression, and reads the graph Laplacian from the storage unit, and on the graph Laplacian A Laplacian label propagation unit that performs Laplacian label propagation using the initial seed and the feature word as an initial value and obtains a word to be included in the response sentence, and a response sentence that includes the word obtained by the Laplacian label propagation unit in the response sentence pattern The response sentence generation part which produces | generates, and the output part which outputs the said response sentence are provided.

Conventionally, research using Laplacian label propagation for word semantic category acquisition has been known, but no attempt has been made to generate a dialog sentence using Laplacian label propagation. In the present invention, after determining the sentence pattern of the input sentence, the sentence pattern of the response sentence corresponding to the sentence pattern of the input sentence is extracted, and Laplacian label propagation is performed using the pattern seed of the response sentence and the feature word as initial values. Thus, a response sentence for the input sentence is generated. In this way, the input sentence is not used as it is, but is abstracted and aggregated into a sentence pattern so that the graph is not sparse and the word acquisition recall rate is increased. In addition, since the word class is assigned to the sentence pattern and the acquired word using proper expression extraction,
It is possible to replace an appropriate part in the sentence pattern with the acquired word. Here, since the graph Laplacian is generated based on the dialogue sentence extracted from the dialogue log corpus, the data necessary for generating the response sentence can be automatically generated without human intervention. Note that the graph Laplacian is a normalized matrix of the graph structure. In the present invention, since Laplacian label propagation using graph Laplacian is used, label propagation is unlikely to occur from a characteristic word that is an initial value to a word having low relevance, and semantic drift can be suppressed.

  In another aspect of the present invention, a dialog sentence generation apparatus includes a storage unit that stores a graph Laplacian calculated from the degree of association between a word and a sentence pattern generated based on a dialog sentence extracted from a dialog log corpus; A Laplacian label that reads the graph Laplacian from the storage unit and performs Laplacian label propagation with the word as an initial value on the graph Laplacian to obtain a sentence pattern and a word to be included in the sentence A propagation unit; a response sentence generation unit that generates a response sentence using a sentence pattern and a word obtained by the Laplacian label propagation unit; and an output unit that outputs the response sentence.

  Thus, the present invention can also be applied when generating a related sentence from a word. Thereby, a sentence including the concept of the input word can be generated.

  The dialog sentence generation method of the present invention includes a graph Laplacian calculated from the degree of association between a word generated based on a dialog sentence extracted from the dialog log corpus and a sentence pattern, a transition matrix between sentence patterns of the dialog log corpus, A dialogue sentence generation device comprising: a storage unit storing a statement, wherein the dialogue sentence generation device receives a sentence input; and the dialogue sentence generation device includes a specific expression from the input sentence. Extracting the sentence pattern of the input sentence based on the extracted unique expression, and reading the transition matrix stored in the storage unit, Extracting a response seed pattern seed corresponding to the sentence pattern of the input sentence using the transition matrix; and Extracting a feature word from the expression; and the dialog sentence generation device reads the graph Laplacian from the storage unit, and on the graph Laplacian, propagates a Laplacian label using the pattern seed of the response sentence and the feature word as initial values. Performing the step of obtaining a word to be included in the response sentence, generating the response sentence by inserting the word obtained by a Laplacian label propagation unit into the response sentence pattern, and the dialog sentence And a generation device that outputs the response sentence.

  According to another aspect of the present invention, there is provided a dialog sentence generation method including a storage unit that stores a graph Laplacian calculated from a degree of association between a word generated based on a dialog sentence extracted from a dialog log corpus and a sentence pattern. A method for generating a dialogue sentence by a dialogue sentence generation device, wherein the dialogue sentence generation device receives an input of a word, the dialogue sentence generation device reads the graph Laplacian from the storage unit, and the graph On the Laplacian, Laplacian label propagation is performed with the word as an initial value, and a sentence pattern and a word to be included in the sentence are obtained, and the dialogue sentence generation device generates a response sentence using the obtained sentence pattern and word And the dialog sentence generating device includes the step of outputting the response sentence.

  The program of the present invention is a program for causing a computer to execute the above-described dialog sentence generation method.

  According to the present invention, a sentence template that has been manually created so far can be automatically acquired by using a dialogue log. As a result, it is possible to generate various sentences that have been difficult in terms of cost by simply increasing the log data. In addition, word candidates for filling a pattern slot can also be acquired from log data without requiring language resources such as a synonym dictionary and a thesaurus.

It is a figure which shows the structure of the dialogue production | generation apparatus of 1st Embodiment. It is explanatory drawing of Laplacian label propagation. It is a figure which shows operation | movement of the dialogue production | generation apparatus of 1st Embodiment. It is a figure which shows operation | movement of the dialogue production | generation apparatus of 1st Embodiment. It is a figure which shows the structure of the dialogue production | generation apparatus of 2nd Embodiment. It is a figure which shows operation | movement of the dialogue production | generation apparatus of 2nd Embodiment.

Hereinafter, a dialogue generation apparatus according to an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram illustrating a configuration of a dialogue sentence generation device 1 according to an embodiment. The hardware of the dialog statement generating apparatus 1 is a computer including a CPU, RAM, ROM, hard disk, display, keyboard, mouse, communication interface, and the like. By causing a computer to execute a dialogue generation program, the dialogue generation device 1 shown in FIG. 1 is realized. Thus, the program which implement | achieves the dialog sentence production | generation apparatus 1 is also contained in the scope of the present invention.

  The dialogue sentence generation device 1 includes a pre-calculation unit 10 and a sentence generation unit 20. The pre-calculation unit 10 has a function of generating a sentence pattern transition matrix used for generating a response sentence and a graph Laplacian. The sentence pattern transition matrix and the graph Laplacian are stored in the sentence pattern transition matrix storage unit 17 and the graph Laplacian storage unit 18, respectively. The sentence generation unit 20 has a function of generating a response sentence that responds to an input sentence using a sentence pattern transition matrix and a graph Laplacian.

  A dialog log corpus 30 database is connected to the dialog statement generating apparatus 1 via a network or the like, and the dialog sentence generating apparatus 1 acquires dialog log data from the dialog log corpus 30. Note that the dialog sentence generation device 1 may acquire the dialog log from the dialog log corpus 30, for example, from an interactive article such as an interview on the Internet or data from a bulletin board.

  The pre-calculation unit 10 includes a dialogue log input unit 11, a specific expression extraction unit 12, a sentence pattern extraction unit 13, a sentence pattern transition matrix construction unit 14, a bipartite graph construction unit 15, a graph Laplacian calculation unit 16, have. The dialogue log input unit 11 has a function of accepting dialogue log input from the dialogue log corpus 30.

  The specific expression extraction unit 12 has a function of extracting a specific expression from the dialogue log. The named entity extraction is one of natural language processing techniques using a computer, and is a technique for extracting a proper noun (person name, place name, etc.), date, time expression, and the like. According to the definition of IREX (Information Retrieval and Extraction Exercise), specific names include organization name (ORGANIZATION), person name (PERSON), place name (LOCATION), date expression (DATE), time expression (TIME), monetary expression ( There are eight types, MONEY), ratio expression (PERCENT), and unique object name (ARTIFACT), and the dialog sentence generation apparatus 1 of this embodiment uses these.

  The sentence pattern extraction unit 13 also has a function of generating a sentence pattern by replacing the unique expression extracted from the dialogue log with a slot. For example, if we extracted a specific expression from the sentence "I saw Kaguya Hime today", "Today (DATE) (O) looked at (O) Hime Kaguya (O) (O). “Today” and “Princess Kaguya” are extracted as specific expressions. Note that “O” means not a unique expression (OUTSIDE). If you replace the specific expression of the above example with a slot, “(DATE) saw (ARTIFACT).” This is a sentence pattern. The sentence pattern extraction unit 13 generates a sentence pattern of the dialog log acquired from the dialog log corpus 30. At this time, the same sentence pattern may be generated from different dialogue sentences, and the sentence patterns are aggregated.

  The sentence pattern transition matrix construction unit 14 has a function of constructing a transition matrix that defines which sentence pattern of the response sentence has changed from the sentence pattern of the utterance sentence. The sentence pattern transition matrix construction unit 14 receives the sentence pattern generated by the sentence pattern extraction unit 13, and based on the transition from the sentence pattern of the utterance sentence to the sentence pattern of the response sentence, the sentence pattern of the utterance sentence and the response sentence Construct a transition matrix of sentence patterns. Specifically, all sentence patterns of utterance sentences are taken as rows, all sentence patterns of response sentences are taken as columns, and a matrix with the number of transitions from the sentence pattern of utterance sentences to the sentence pattern of response sentences as an element is generated. . By using such a transition matrix, the probability of transition from the sentence pattern of the spoken sentence to the sentence pattern of the response sentence can be obtained. The sentence pattern transition matrix construction unit 14 stores the constructed sentence pattern transition matrix in the sentence pattern transition matrix storage unit 17.

  The bipartite graph construction unit 15 has a function of generating a bipartite graph with the degree of association between words and sentence patterns as edge weights. A bipartite graph is a graph in which a vertex set can be divided into two subsets so that there are no edges between vertices in each set.

  FIG. 2 is a diagram illustrating an example of a bipartite graph constructed by the bipartite graph construction unit 15 of the present embodiment. As shown in FIG. 2, there is a set of words on the left side and a set of sentence patterns on the right side, and an edge is provided between the related word and the sentence pattern. This edge represents the strength of the relationship between the word and the sentence pattern.

  The bipartite graph construction unit 15 receives the word extracted as the specific expression by the specific expression extraction unit 12 and receives the sentence pattern extracted by the sentence pattern extraction unit 13. The bipartite graph construction unit 15 obtains the degree of association between the word and the sentence pattern based on the sentence pattern and the word appearing in the sentence pattern, and constructs the bipartite graph. In the dialogue generating apparatus 1, the structure of the bipartite graph is held in the form of a matrix. That is, a matrix having all words extracted by the unique expression extraction unit 12 as rows, all sentence patterns extracted by the sentence pattern extraction unit 13 as columns, and the number of occurrences of words in the sentence pattern as elements. Generate. In the following description, this matrix is referred to as “word-sentence pattern matrix W”.

The graph Laplacian calculator 16 calculates a graph Laplacian L from the bipartite graph. The graph Laplacian L is calculated by the following formula.
L = ID (A) ^-1/2 , A, D (A) ^-1/2
Here, A is an adjacency matrix, and A = W ^T · W. The adjacency matrix A is a matrix representing the co-occurrence frequency of words. D (A) is an order diagonal matrix, and its elements are represented by the following equations.

  Thus, the graph Laplacian sums the graph Laplacian L in the column direction and the row direction by subtracting the matrix obtained by multiplying the matrix obtained by dividing the diagonal matrix by the square root from before and after the adjacency matrix A from the unit matrix. Respectively become 0. Laplacian label propagation using the graph Laplacian normalized in this way has an effect of reducing the weight of the self-similarity in the graph, so that a high weight is not given to the generic pattern, and semantic drift hardly occurs. The graph Laplacian calculation unit 16 stores the obtained graph Laplacian data in the graph Laplacian storage unit 18.

[Configuration of sentence generator 20]
Next, the configuration of the sentence generation unit 20 will be described. The sentence generation unit 20 generates a response sentence that responds to the input dialogue sentence. The sentence generation unit 20 includes a sentence input unit 21, a specific expression extraction unit 22, an input sentence pattern determination unit 23, a response sentence pattern seed extraction unit 24, a feature word extraction unit 25, a Laplacian label propagation unit 26, The response sentence generation unit 27 and the response sentence output unit 28 are provided.

  The sentence input unit 21 has a function of receiving input of a dialog sentence. The specific expression extraction unit 22 has the same function as the specific expression extraction unit 12 included in the pre-calculation unit 10. That is, the specific expression extraction unit 22 extracts a specific expression from the input dialogue sentence.

  The input sentence pattern determination unit 23 generates a sentence pattern by replacing the specific expression included in the input dialogue sentence with a slot, and compares the input sentence pattern with an existing sentence pattern to determine which input sentence pattern is It has a function of determining whether it corresponds to a sentence pattern. The response sentence pattern seed extraction unit 24 reads out the sentence pattern transition matrix stored in the sentence pattern transition matrix storage unit 17. The response sentence pattern seed extracting unit 24 uses the sentence pattern transition matrix data to determine the sentence pattern of the response sentence that is likely to change from the input sentence pattern determined by the input sentence pattern determining unit 23. Have

  The feature word extraction unit 25 extracts a feature word to be used as a seed in Laplacian label propagation from the words extracted as the specific expression by the specific expression extraction unit 22. In the present embodiment, all of the extracted words are used as feature words. However, words characteristic of the input sentence pattern may be used as feature words. For example, the occurrence of semantic drift can be suppressed by eliminating common words such as “today” and “me” that appear regardless of context. Here, in order to determine whether a certain word is a general word, for example, a TF-IDF method can be used.

ここで、Ｆ（０）はシードのラベル、Ｆ（ｔ）はグラフ上で伝搬させた各ノードのラベルを表す。αはシードベクトルとグラフのどちらをどの程度重視するかの調整パラメータであり、ラプラシアンラベル伝搬をリスタート付きのランダムウォークと考えた場合では、αはリスタート確率と見なすことができる。 The Laplacian label propagation unit 26 reads graph Laplacian data from the graph Laplacian storage unit 18. Then, the Laplacian label propagation unit 26 performs Laplacian label propagation using the sentence pattern of the response sentence extracted by the response sentence pattern seed extraction unit 24 and the feature word extracted by the feature word extraction unit 25 as seeds. Laplacian label propagation is performed according to the following equation.

Here, F (0) represents the label of the seed, and F (t) represents the label of each node propagated on the graph. α is an adjustment parameter indicating how much importance is placed on the seed vector or the graph. When Laplacian label propagation is considered as a random walk with restart, α can be regarded as a restart probability.

  When the sentence pattern of the response sentence is used as a seed in addition to the feature word as in the present embodiment, the other seed vector is weighted when taking the product of the adjacency matrix and the seed vector. That is, the above formula (1) is expressed as follows.

グラフラプラシアンLを再掲する。
Ｌ＝Ｉ−Ｄ（Ａ）^-1/2・Ａ・Ｄ（Ａ）^-1/2

Reprint Graph Laplacian L.
L = ID (A) ^-1/2 , A, D (A) ^-1/2

上記の式の第二項、^TＷＤ（Ｗ^TＷ）^-1/2Ｆ（ｔ）は、文パターンのスコアベクトルとなる。
ここで、文パターンのスコアベクトルを

とおく。文パターンからのラベル伝搬にも、文パターンシードベクトルを導入し、確率的なリスタートを行うように上記を以下のように変更する。

これを（１）式に代入した、文パターンと特徴語の両方をシードとして利用可能なラベル伝搬の式を以下に示す。

ここで、Ｆｖ（０）は特徴語のシードのラベル、Ｆｐ（０）は文パターンシードのラベル、Ｆｖ（ｔ）は単語の各ノードのラベルを表す。αは単語のシードベクトルとグラフのどちらをどの程度重視するかの調整パラメータであり、βは文パターンのシードベクトルとグラフのどちらをどの程度重視するかの調整パラメータである。 Since this equation is a dense matrix, it cannot be expanded in memory when the data becomes large. Therefore, in practice, the first term of equation (1) is modified as follows and treated as sparse.

The second term of the above equation, ^T WD (W ^T W) ^−1/2 F (t), is a sentence pattern score vector.
Where the sentence pattern score vector

far. The sentence pattern seed vector is also introduced to the label propagation from the sentence pattern, and the above is changed as follows to perform a probabilistic restart.

An expression for label propagation that can be used with both sentence patterns and feature words as seeds, substituting this into expression (1), is shown below.

Here, Fv (0) represents the seed label of the feature word, Fp (0) represents the label of the sentence pattern seed, and Fv (t) represents the label of each node of the word. α is an adjustment parameter indicating how much importance is given to the word seed vector and the graph, and β is an adjustment parameter indicating how much importance is attached to the seed vector of the sentence pattern and the graph.

  Fv (t) is a vector whose number of dimensions is the total number of elements of the word, and the value of the i-th dimension of Fv (t) represents the degree to which the i-th word is related to the feature word. That is, Fv (t) is a score vector for the target feature word. Fv (0) given as an input is created by setting the dimension value of a feature word given as a seed to 1 and setting the others to 0. Similarly, Fp (0) is created by setting the dimension value of a sentence pattern given as a seed to 1 and setting the others to 0. The Laplacian label propagation unit 26 performs Laplacian label propagation on the graph Laplacian using Fv (0) and Fp (0) created in this way, and repeatedly calculates the above expression until convergence. As a result, finally converged Fv (t) is output. Fv (t) is a score vector of the word at the end of t step. The Laplacian label propagation unit 26 obtains a word having a high score as a word to be included in the response sentence.

  The response sentence generation unit 27 inputs the word obtained by the Laplacian label propagation unit 26 into the slot of the sentence pattern of the response sentence and generates a response sentence. The response sentence output unit 28 outputs the generated response sentence. The output of the response sentence may be displayed on a display, for example, or may be output as audio from a speaker.

[Operation of Dialog Generation Device 1]
Next, the operation of the dialogue sentence generation device 1 according to the present embodiment will be described.
FIG. 3 is a flowchart showing the preparatory operation for generating the dialog text by the dialog text generating apparatus 1. The dialog statement generating apparatus 1 first acquires a dialog log from the dialog log corpus 30 (S10). The dialogue sentence generation device 1 extracts a unique expression of each sentence included in the acquired dialogue log, and replaces the unique expression of each sentence with a slot to generate a sentence pattern (S11).

  Subsequently, the dialogue sentence generation device 1 constructs a sentence pattern transition matrix based on the sentence pattern transitions in the dialogue sentence (S12). Specifically, as described above, a matrix having the number of transitions from the sentence pattern of the utterance sentence to the sentence pattern of the response sentence as an element is generated. The dialogue sentence generation device 1 stores the constructed sentence pattern transition matrix in the sentence pattern transition matrix storage unit 17.

  Next, the dialog sentence generation apparatus 1 constructs a bipartite graph of words and sentence patterns (S13). A specific example of the bipartite graph is as shown in FIG. Subsequently, the dialog sentence generating apparatus 1 calculates a graph Laplacian from the bipartite graph (S14), and stores the graph Laplacian data in the graph Laplacian storage unit 18. With the above operation, a sentence pattern transition matrix and a graph Laplacian can be prepared using a large amount of dialogue logs acquired from the dialogue log corpus 30.

  FIG. 4 is a flowchart showing an operation of generating a response sentence according to the input sentence by the dialog sentence generating apparatus 1. When receiving the input of a sentence (S20), the dialog sentence generating apparatus 1 extracts a specific expression from the input sentence and generates a sentence pattern by replacing the specific expression with a slot (S21). Subsequently, the dialog sentence generating apparatus 1 extracts feature words used for Laplacian label propagation from the extracted unique expressions (S22). In the present embodiment, the dialog sentence generation apparatus 1 uses all words extracted as specific expressions as feature words.

  Next, the dialog sentence generating apparatus 1 determines the pattern of the input sentence (S23). Here, the input sentence pattern determination unit 23 generates a sentence pattern by replacing the specific expression extracted by the specific expression extraction unit 22 with a slot, and determines which existing sentence pattern the generated sentence pattern corresponds to. judge. The dialogue sentence generation device 1 reads the sentence pattern transition matrix from the sentence pattern transition matrix storage unit 17, and extracts the sentence pattern of the response sentence with respect to the sentence pattern of the input sentence using the read sentence pattern transition matrix (S24). The sentence pattern of the obtained response sentence is used as a pattern seed for Laplacian label propagation.

  The dialog sentence generation device 1 performs Laplacian label propagation using the feature words included in the input sentence and the sentence pattern of the response sentence as seeds (S25). As a result of Laplacian label propagation, the dialogue sentence generation device 1 puts a word having a high score in the sentence pattern slot of the response sentence, generates a response sentence, and outputs it (S26).

  Heretofore, the configuration and operation of the dialogue sentence generation device 1 according to the embodiment of the present invention have been described. Since the dialogue generation apparatus 1 according to the embodiment uses the dialogue log acquired from the dialogue log corpus 30 to generate a dialogue sentence using the automatically generated sentence pattern transition matrix and graph Laplacian, conventionally, Generation of various sentences, which was difficult in terms of aspect, can be realized simply by increasing the log data. In addition, word candidates for filling a slot of a pattern can be acquired from log data without requiring language resources such as a synonym dictionary and a thesaurus.

  The dialogue sentence generation device 1 according to the present embodiment performs Laplacian label propagation using a graph Laplacian by using a feature word included in the inputted sentence and a sentence pattern of a response sentence corresponding to the input sentence as a seed. The generation of an appropriate response sentence can be generated.

  As mentioned above, although the dialogue sentence generation device 1 of the present invention has been described in detail with reference to the embodiment, the present invention is not limited to the above-described embodiment.

  In the dialog sentence generation device 1 according to the above-described embodiment, an example of generating a response sentence that responds to an input sentence has been described. However, it is also possible to generate a sentence based on a word input. For example, a dialog sentence generation device 2 (a modification of the above embodiment) used when generating a dialog sentence including a word having the same meaning as a certain word is also included in the scope of the present invention.

  FIG. 5 is a diagram illustrating a configuration of the dialogue sentence generation device 2 according to the modification. The basic configuration of the dialog statement generation device 2 shown in FIG. 5 is the same as that of the dialog statement generation device 1 of the above-described embodiment, but in the dialog statement generation device 2 shown in FIG. The difference is that the pattern transition matrix construction unit 14 is not provided, and the sentence generation unit 20 does not include the input sentence pattern determination unit 23, the response sentence pattern seed extraction unit 24, and the feature word extraction unit 25. The sentence generation unit 20 includes a word input unit 29 instead of the sentence input unit 21. The operation in which the dialog statement generating device 2 obtains the graph Laplacian from the dialog log and makes the preparation in advance is the same as the operation of the dialog statement generating device 1 in the above-described embodiment.

  FIG. 6 is a diagram illustrating an operation of generating a dialogue sentence by the dialogue sentence generating apparatus 2 according to the modification. The dialog sentence generation device 2 receives an input of a word (S30). The dialog sentence generation device 2 reads graph Laplacian data from the graph Laplacian storage unit 18, and performs Laplacian label propagation using the input word as a seed (S31). As a result of Laplacian label propagation, the dialogue sentence generation device 2 obtains a sentence pattern and a word having a high vector score, generates a dialogue sentence by inserting the word into the slot of the obtained sentence pattern, and outputs it (S32).

  In this way, by performing Laplacian label propagation using a graph Laplacian including a word-sentence pattern matrix, a dialogue sentence can be generated only from the input word.

  According to the present invention, it is possible to automatically generate a graph Laplacian for generating a dialog sentence by using the dialog log acquired from the dialog log corpus, which is useful for a voice dialog system and the like. is there.

DESCRIPTION OF SYMBOLS 1, 2 Dialogue sentence production | generation apparatus 10 Prior calculation part 11 Dialogue log input part 12 Specific expression extraction part 13 Sentence pattern extraction part 14 Sentence pattern transition matrix construction part 15 Second part Graph construction part 16 Graph Laplacian calculation part 17 Sentence pattern transition matrix memory | storage Unit 18 Graph Laplacian Storage Unit 20 Sentence Generation Unit 21 Sentence Input Unit 22 Specific Expression Extraction Unit 23 Input Sentence Pattern Determination Unit 24 Response Sentence Pattern Seed Extraction Unit 25 Feature Word Extraction Unit 26 Laplacian Label Propagation Unit 27 Response Sentence Generation Unit 28 Response Sentence Output unit 29 Word input unit

Claims (6)

A storage unit that stores a graph Laplacian calculated from the degree of association between a word and a sentence pattern generated based on a dialog sentence extracted from the dialog log corpus, and a transition matrix between sentence patterns of the dialog log corpus;
A sentence input unit that accepts sentence input;
A specific expression extraction unit that extracts a specific expression from an input sentence;
A pattern determination unit that determines a sentence pattern of an input sentence based on the extracted specific expression;
A pattern seed extraction unit that reads the transition matrix stored in the storage unit and extracts a pattern seed of a response sentence corresponding to a sentence pattern of an input sentence using the transition matrix;
A feature word extraction unit for extracting a feature word from the specific expression extracted by the specific expression extraction unit;
Reading the graph Laplacian from the storage unit, performing Laplacian label propagation using the pattern seed of the response sentence and the feature word as initial values on the graph Laplacian, and obtaining a word to be included in the response sentence; a Laplacian label propagation unit;
A dialog sentence generation unit that generates a response sentence by putting the word obtained by the Laplacian label propagation unit into the pattern of the response sentence;
An output unit for outputting the response sentence;
A dialogue sentence generation device comprising: A storage unit storing a graph Laplacian calculated from the degree of association between a word generated based on a dialogue sentence extracted from a dialogue log corpus and a sentence pattern;
An input unit that accepts input of words;
Reading the graph Laplacian from the storage unit, performing Laplacian label propagation using the word as an initial value on the graph Laplacian, a Laplacian label propagation unit for obtaining a sentence pattern and a word to be included in the sentence,
A dialog sentence generation unit that generates a dialog sentence using the sentence pattern and the word obtained by the Laplacian label propagation unit;
An output unit for outputting the dialogue sentence;
A dialogue sentence generation device comprising: Dialogue comprising a graph Laplacian calculated from the degree of association between a word and a sentence pattern generated based on a dialogue sentence extracted from the dialogue log corpus, and a storage unit storing a transition matrix between sentence patterns of the dialogue log corpus A method for generating a dialogue sentence by a sentence generator,
The interactive sentence generation device accepting an input of a sentence;
The dialog sentence generation device extracting a specific expression from an input sentence;
The dialog sentence generation device determines a sentence pattern of an input sentence based on the extracted specific expression;
The dialogue sentence generation device reads a transition matrix stored in the storage unit, and extracts a pattern seed of a response sentence corresponding to a sentence pattern of an input sentence using the transition matrix;
The dialog sentence generating device extracting a feature word from a specific expression extracted from an input sentence;
The dialogue sentence generation device reads the graph Laplacian from the storage unit, performs Laplacian label propagation using the pattern seed of the response sentence and the feature word as initial values on the graph Laplacian, and selects a word to be included in the response sentence. Seeking steps,
The dialogue sentence generating device generating a response sentence by putting a word obtained by a Laplacian label propagation unit in the response sentence pattern;
The dialog sentence generating device outputting the response sentence;
A dialog sentence generation method comprising: A method for generating a dialogue sentence by a dialogue sentence generation device including a storage unit storing a graph Laplacian calculated from a degree of association between a word generated based on a dialogue sentence extracted from a dialogue log corpus and a sentence pattern. ,
The interactive sentence generating device accepting an input of a word;
The dialogue sentence generation device reads the graph Laplacian from the storage unit, performs Laplacian label propagation with the word as an initial value on the graph Laplacian, and obtains a sentence pattern and a word to be included in the sentence;
The interactive sentence generation device generates a response sentence using the obtained sentence pattern and word;
The dialog sentence generating device outputting the response sentence;
A dialog sentence generation method comprising: A computer comprising a storage unit that stores a graph Laplacian calculated from the degree of association between a word generated based on a dialogue sentence extracted from a dialogue log corpus and a sentence pattern, and a transition matrix between sentence patterns of the dialogue log corpus A program for generating a dialogue sentence, wherein the computer
Receiving a sentence input;
Extracting a specific expression from the input sentence;
Determining a sentence pattern of the input sentence based on the extracted specific expression;
Reading the transition matrix stored in the storage unit, and using the transition matrix to extract the pattern seed of the response sentence corresponding to the sentence pattern of the input sentence;
Extracting a feature word from the specific expression extracted from the input sentence;
Reading the graph Laplacian from the storage unit, performing Laplacian label propagation with the pattern seed of the response sentence and the feature word as initial values on the graph Laplacian, and obtaining a word to be included in the response sentence;
Generating a response sentence by putting the word obtained by the Laplacian label propagation unit into the response sentence pattern;
Outputting the response sentence;
A program that executes A program that causes a computer including a storage unit that stores a graph Laplacian calculated from the degree of association between a word generated based on a dialog sentence extracted from a dialog log corpus and a sentence pattern to generate a dialog sentence,
Receiving a word input;
Reading the graph Laplacian from the storage unit, performing Laplacian label propagation with the word as an initial value on the graph Laplacian, obtaining a sentence pattern and a word to be included in the sentence;
Generating a response sentence using the obtained sentence pattern and word;
Outputting the response sentence;
A program that executes

Images