CN107193806A - A kind of vocabulary justice former automatic prediction method and device - Google Patents

Abstract

The embodiment of the present invention discloses a method and device for automatically predicting a sememe of a vocabulary. The method includes: calculating the vector distance between each unknown sememe word and each known sememe word according to the word vector of each preset vocabulary; Each vector distance and distance threshold, select at least one target known sememe vocabulary as the candidate sememe set of each unknown sememe vocabulary; according to the sememe vector of each target known sememe vocabulary in the candidate sememe set, calculate The scores of each sememe of each unknown sememe vocabulary; according to the score threshold and the score of each sememe, the first sememe vector of each unknown sememe vocabulary is obtained. Determine the candidate sememe sets of each unknown sememe vocabulary by vector distance, further calculate the scores of each sememe in the candidate sememe set, and thus obtain the first sememe vector of each unknown sememe vocabulary, which can be automatically compared Accurate sememe prediction of unknown sememe words reduces the pressure of manual labeling and reduces the possible deviation of the results caused by different people's labeling.

Description

A method and device for automatic prediction of lexical sememe

technical field

The embodiments of the present invention relate to the field of computer technology, and in particular to a method and device for automatic prediction of lexical sememes.

Background technique

Sentences are composed of words to express different meanings, and different words have their particularities and similarities. HowNet is used to describe these characteristics of different words. HowNet is manually marked, and its sememes are marked for most of the common words, and the sememes are a smaller set than the words, which represent some more basic meanings of the words. Different sememe combinations can represent different Vocabulary, such as: the meaning of antique shop includes: place, business, buy, sell, treasure and past. The definition of an antique shop can be described by these meanings: a business place that buys and sells treasures from the past is an antique shop. Sememes have many good features, such as judging the similarity between two words based on the intersection of their sememes, which can be used to better generate word vectors for more tasks in natural language processing.

Although the meanings have many good properties, the labeling of the meanings is a very time-consuming and labor-intensive task. HowNet has been born for more than ten years. At first, many language experts took the lead in annotating. However, with the rapid development of information technology, the number of vocabulary is growing explosively. How to efficiently, quickly and accurately annotate these newly generated vocabulary? The original has become a problem that has to be solved. There is an urgent need for an automatic construction model of the sememe instead of manual labeling, which can not only ensure that the sememe has the same characteristics, but also avoid the deviation caused by human labeling.

Contents of the invention

Due to the above-mentioned problems in the prior art, an embodiment of the present invention proposes a method and device for automatic prediction of lexical sememes.

In the first aspect, the embodiment of the present invention proposes a method for automatic prediction of lexical sememes, including:

Calculate the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary;

According to each vector distance and the distance threshold, select at least one target known sememe vocabulary as a candidate sememe set for each unknown sememe vocabulary;

According to the sememe vectors of each target known sememe vocabulary in the candidate sememe set, calculate the scores of each sememe of each unknown sememe vocabulary;

According to the score threshold and the score of each sememe, the first sememe vector of each unknown sememe vocabulary is obtained;

Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary.

Optionally, the method also includes:

The preset sememe is obtained, and the word vector of each preset vocabulary is calculated according to the stochastic gradient descent method and the preset sememe.

Optionally, after obtaining the sememe vectors of each unknown sememe vocabulary according to the score threshold and the scores of each sememe, it also includes:

According to the preset sememe vector and the unknown sememe vocabulary vector, a sememe vocabulary matrix is obtained;

According to the sememe vocabulary matrix, calculate the co-occurrence matrix of the sememe vocabulary matrix;

Decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector;

calculating a target value according to the unknown sememe word vector and the second sememe vector;

calculating a target sememe vector according to the target value and the first sememe vector;

Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe.

Optionally, decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector, specifically includes:

Decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to a stochastic gradient descent method and a loss function to obtain a second sememe vector;

Wherein, the loss function is:

W is the unknown sememe vocabulary vector, S is the first pre-set sememe vector and the second pre-set sememe vector, S is the first pre-set sememe vector, λ is the pre-set coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector.

In the second aspect, the embodiment of the present invention also proposes an automatic prediction device for a vocabulary sememe, including:

The distance calculation module is used to calculate the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary;

A sememe set determination module is used to select at least one target known sememe vocabulary as an alternative sememe set for each unknown sememe vocabulary according to each vector distance and distance threshold;

The sememe score calculation module is used to calculate the scores of each sememe of each unknown sememe vocabulary according to the sememe vectors of each target known sememe vocabulary in the alternative sememe set;

The sememe vector determination module is used to obtain the first sememe vector of each unknown sememe vocabulary according to the score threshold and the scores of each sememe;

Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary.

Optionally, the device also includes:

The word vector calculation module is used to obtain the preset sememe, and calculate the word vector of each preset vocabulary according to the stochastic gradient descent method and the preset sememe.

Optionally, the device also includes:

The sememe vocabulary matrix acquisition module is used to obtain the sememe vocabulary matrix according to the preset sememe vector and the unknown sememe vocabulary vector;

A co-occurrence matrix calculation module, configured to calculate the co-occurrence matrix of the sememe vocabulary matrix according to the sememe vocabulary matrix;

A matrix decomposition module is used to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector;

A target value calculation module, configured to calculate a target value according to the unknown sememe word vector and the second sememe vector;

A target sememe vector calculation module, configured to calculate and obtain a target sememe vector according to the target value and the first sememe vector;

Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe.

Optionally, the matrix decomposition module is specifically configured to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to a stochastic gradient descent method and a loss function to obtain a second sememe vector;

Wherein, the loss function is:

L＝∑ _{w∈W，s∈S} (W·(S+S')+b _w +b _s -M _ws ) ² +λ∑ _s，t∈S (s·tC _st ) ²

W is the unknown sememe vocabulary vector, S is the first pre-set sememe vector and the second pre-set sememe vector, S is the first pre-set sememe vector, λ is the pre-set coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector.

It can be seen from the above technical solution that the embodiment of the present invention determines the candidate sememe sets of each unknown sememe word through the vector distance between the unknown sememe word and each known sememe word, and further calculates the sememe set in the candidate sememe set. , and thus get the first sememe vector of each unknown sememe vocabulary, which can automatically predict the exact sememe of unknown sememe words, reduce the pressure of manual labeling, and reduce the impact of labeling results by different people Possible deviation.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a schematic flow chart of a method for automatic prediction of a lexical sememe provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of the sememe of the vocabulary "antique shop" provided by an embodiment of the present invention;

Fig. 3 is a schematic diagram of the sememe of the word "apple" provided by an embodiment of the present invention;

Fig. 4 is a schematic flow diagram of selecting an alternative sememe set provided by an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of an automatic prediction device for a word sememe provided by an embodiment of the present invention.

detailed description

The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solution of the present invention more clearly, but not to limit the protection scope of the present invention.

Fig. 1 shows a schematic flow chart of a method for automatic prediction of a lexical sememe provided in this embodiment, including:

S101. Calculate the vector distance between each unknown sememe word and each known sememe word according to the word vector of each preset word.

Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary.

Specifically, first count the word frequency of each vocabulary and the hyponymy relationship between different vocabulary on a large corpus; then use stochastic gradient descent to achieve co-occurrence matrix decomposition to obtain the word vector of the vocabulary.

The co-occurrence matrix contains rich text information and the relationship between words. Through matrix decomposition and dimensionality reduction, the low-dimensional representation of vocabulary can still reflect the relationship between words and words.

The content of the sememe vector of each preset vocabulary is shown in Figures 2 and 3, Figure 2 shows the sememe included in the Chinese word "antique store", and Figure 3 shows the sememe included in the English word "apple".

S102. Select at least one target known sememe word as a candidate sememe set for each unknown sememe word according to the distance of each vector and the distance threshold.

For unknown sememe words, the selection of alternative sememe sets is shown in Figure 4. Find several known sememe words closest to it in the vector space, and use their sememes as the alternative sememe sets. And these sememes need to be scored according to the distance between words.

Specifically, firstly, for each new word, the distance between it and the words with known sememe vectors is calculated; and several nearest words are selected. For the nearest selected words, calculate the weights of their sememes to the new words.

Assuming that the closer a vocabulary is to a new word, then the sememe of this vocabulary is more likely to be the sememe of the new word, so for any sememe, the score it can get for a new word can be expressed as follows:

Among them, w represents a new word, s represents a sememe, W is the set of all known sememe words, M _vs represents whether the vocabulary v has a sememe s, if yes, it is 1, otherwise it is 0, and the obtained Pr(s|w ) is higher, the more likely s is the sememe of v.

In fact, it is much more complicated to actually calculate the above process. Since the distance between the normalized vectors is between (-1, 1), it is not possible to distinguish different sememes very well, so we make the closer the distance The sememe of the vocabulary has a greater weight. We introduce the hyperparameter p. For the k-th closest vocabulary, multiplying p ^k makes the difference between the sememes of different vocabulary greater, and multiplying it with an exponentially decaying coefficient ensures that If Pr(s|w) is within a certain range, it will not diverge.

S103. According to the sememe vectors of each target known sememe vocabulary in the candidate sememe set, calculate the scores of each sememe of each unknown sememe vocabulary.

S104. Obtain the first sememe vector of each unknown sememe vocabulary according to the score threshold and the scores of each sememe.

Specifically, a vocabulary often shares a sememe with its similar words. For example, both China and the United States share a sememe such as "special" and "country", but many words often have their own sememe, so the method provided in this embodiment is proposed , not only can learn the sememe from similar words, but also can learn the unique sememe.

In this embodiment, the candidate sememe sets of each unknown sememe word are determined by the vector distance between the unknown sememe word and each known sememe word, and the scores of each sememe in the candidate sememe set are further calculated, and thus obtained The first sememe vector of each unknown sememe word can automatically predict the unknown sememe word accurately, reduce the pressure of manual labeling, and reduce the possible deviation of the results caused by different people's labeling.

Further, on the basis of the above method embodiments, the method further includes:

S100. Obtain a preset sememe, and calculate a word vector of each preset vocabulary according to the stochastic gradient descent method and the preset sememe.

Among them, the default sememes include 1400 common sememes.

Further, on the basis of the above method embodiment, after obtaining the sememe vectors of each unknown sememe vocabulary according to the score threshold and the scores of each sememe, it also includes:

S105. Obtain a sememe vocabulary matrix according to the preset sememe vector and the unknown sememe vocabulary vector;

Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe.

S106. Calculate and obtain a co-occurrence matrix of the sememe vocabulary matrix according to the sememe vocabulary matrix;

The co-occurrence matrix of sememes and sememes contains rich relationships between sememes. Just as the co-occurrence matrix of words can be used to generate word vectors, the co-occurrence matrix of sememes can also assist in generating better sememe vectors.

S107. Decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector;

Specifically, first calculate the 01 matrix of vocabulary and sememe; then calculate the co-occurrence matrix of sememe and sememe; finally use the method of stochastic gradient descent to decompose the above two matrices to obtain the sememe vector.

S108. Calculate and obtain a target value according to the unknown sememe word vector and the second sememe vector;

S109. Calculate and obtain a target sememe vector according to the target value and the first sememe vector;

Further, on the basis of the above method embodiments, S107 specifically includes:

Decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to a stochastic gradient descent method and a loss function to obtain a second sememe vector;

Wherein, the loss function is:

L＝∑ _{w∈W，s∈S} (W·(S+S′)+b _w +b _s -M _ws ) ² +λ∑ _s，t∈S (s·tC _st ) ²

W is the unknown sememe word vector, S is S′ is the first preset sememe vector and the second preset sememe vector, λ is the preset coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector.

L is reduced by gradient descent to obtain a good sememe vector representation. Finally, the following function can be used to calculate the possible relationship between a new word and a sememe:

Pr(s|w)=∑ _v∈W cos(v, W)M _vs +λ cos(w, s)

At present, there is no automatic prediction model for sememes. The existing method is to manually label by humans, which is time-consuming and labor-intensive, and the labeling effect varies from person to person, which will have a great impact on the accuracy of sememes. In this embodiment, the existing tagging data can be used to automatically predict sememes, and a part of HowNet data is used as a test set for testing. It can be found that the results of this embodiment overlap to a large extent with manual tagging, and the accuracy is high. Moreover, the model of this embodiment can discover some recommended sememes that have not been marked in HowNet, and the results of these newly discovered candidate sememes are also quite correct.

Fig. 5 shows a schematic structural diagram of an automatic prediction device for a vocabulary sememe provided by this embodiment, the device includes: a distance calculation module 501, a sememe set determination module 502, a sememe score calculation module 503 and a sememe vector Determine module 504, wherein:

The distance calculation module 501 is used to calculate the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary;

The sememe set determination module 502 is used to select at least one target known sememe vocabulary as a candidate sememe set for each unknown sememe vocabulary according to each vector distance and distance threshold;

The sememe score calculation module 503 is used to calculate the scores of each sememe of each unknown sememe vocabulary according to the sememe vectors of each target known sememe vocabulary in the candidate sememe set;

The sememe vector determination module 504 is used to obtain the first sememe vector of each unknown sememe vocabulary according to the score threshold and the scores of each sememe;

Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary.

Specifically, the distance calculation module 501 calculates the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary; the sememe set determination module 502 calculates the vector distance between each unknown sememe vocabulary and each Distance threshold, select at least one target known sememe vocabulary as the candidate sememe set of each unknown sememe vocabulary; the sememe score calculation module 503 is based on the sememe of each target known sememe vocabulary in the candidate sememe set Vectors are calculated to obtain the scores of each sememe of each unknown sememe vocabulary; the sememe vector determination module 504 obtains the first sememe vector of each unknown sememe vocabulary according to the score threshold and the scores of each sememe.

In this embodiment, the candidate sememe sets of each unknown sememe word are determined by the vector distance between the unknown sememe word and each known sememe word, and the scores of each sememe in the candidate sememe set are further calculated, and thus obtained The first sememe vector of each unknown sememe word can automatically predict the unknown sememe word accurately, reduce the pressure of manual labeling, and reduce the possible deviation of the results caused by different people's labeling.

Further, on the basis of the above device embodiment, the device further includes:

The word vector calculation module is used to obtain the preset sememe, and calculate the word vector of each preset vocabulary according to the stochastic gradient descent method and the preset sememe.

Further, on the basis of the above device embodiment, the device further includes:

The sememe vocabulary matrix acquisition module is used to obtain the sememe vocabulary matrix according to the preset sememe vector and the unknown sememe vocabulary vector;

A co-occurrence matrix calculation module, configured to calculate the co-occurrence matrix of the sememe vocabulary matrix according to the sememe vocabulary matrix;

A matrix decomposition module is used to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector;

A target value calculation module, configured to calculate a target value according to the unknown sememe word vector and the second sememe vector;

A target sememe vector calculation module, configured to calculate and obtain a target sememe vector according to the target value and the first sememe vector;

Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe.

Further, on the basis of the above device embodiment, the matrix decomposition module is specifically used to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method and the loss function, to obtain the second sememe vector;

Wherein, the loss function is:

L＝Σ _{w∈W，s∈S} (W·(S+S')+b _w +b _s -M _ws ) ² +λΣ _s，t∈S (s·tC _st ) ²

W is the unknown sememe vocabulary vector, S is the first pre-set sememe vector and the second pre-set sememe vector, S is the first pre-set sememe vector, λ is the pre-set coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector.

The device for automatic prediction of word sememes described in this embodiment can be used to implement the above method embodiments, and its principles and technical effects are similar, and will not be repeated here.

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without any creative efforts.

Through the above description of the implementations, those skilled in the art can clearly understand that each implementation can be implemented by means of software plus a necessary general hardware platform, and of course also by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic discs, optical discs, etc., including several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) execute the methods described in various embodiments or some parts of the embodiments.

It should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it still can The technical solutions described in the foregoing embodiments are modified, or some of the technical features are replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (8)

1. an automatic prediction method of a vocabulary sememe, characterized in that, comprising: Calculate the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary; According to each vector distance and the distance threshold, select at least one target known sememe vocabulary as a candidate sememe set for each unknown sememe vocabulary; According to the sememe vectors of each target known sememe vocabulary in the candidate sememe set, calculate the scores of each sememe of each unknown sememe vocabulary; According to the score threshold and the score of each sememe, the first sememe vector of each unknown sememe vocabulary is obtained; Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary. 2. The method according to claim 1, characterized in that the method further comprises: The preset sememe is obtained, and the word vector of each preset vocabulary is calculated according to the stochastic gradient descent method and the preset sememe. 3. method according to claim 1, is characterized in that, described according to score threshold value and the score of each sememe, after obtaining the sememe vector of each unknown sememe vocabulary, also includes: According to the preset sememe vector and the unknown sememe vocabulary vector, a sememe vocabulary matrix is obtained; According to the sememe vocabulary matrix, calculate the co-occurrence matrix of the sememe vocabulary matrix; Decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector; calculating a target value according to the unknown sememe word vector and the second sememe vector; calculating a target sememe vector according to the target value and the first sememe vector; Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe. 4. The method according to claim 3, wherein said decomposing said sememe vocabulary matrix and said co-occurrence matrix respectively according to stochastic gradient descent method to obtain a second sememe vector, specifically comprising: Decomposing the sememe vocabulary matrix and the co-occurrence matrix respectively according to a stochastic gradient descent method and a loss function to obtain a second sememe vector; Wherein, the loss function is: L＝∑ _{w∈W，s∈S} (W·(S+S')+b _w +b _s -M _ws ) ² +λ∑ _s，t∈S (s·tC _st ) ² W is the unknown sememe vocabulary vector, S is the first pre-set sememe vector and the second pre-set sememe vector, S is the first pre-set sememe vector, λ is the pre-set coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector. 5. An automatic prediction device for a vocabulary sememe, characterized in that it comprises: The distance calculation module is used to calculate the vector distance between each unknown sememe vocabulary and each known sememe vocabulary according to the word vector of each preset vocabulary; A sememe set determination module is used to select at least one target known sememe vocabulary as an alternative sememe set for each unknown sememe vocabulary according to each vector distance and distance threshold; The sememe score calculation module is used to calculate the scores of each sememe of each unknown sememe vocabulary according to the sememe vectors of each target known sememe vocabulary in the alternative sememe set; The sememe vector determination module is used to obtain the first sememe vector of each unknown sememe vocabulary according to the score threshold and the scores of each sememe; Wherein, the preset vocabulary includes known sememe vocabulary and unknown sememe vocabulary. 6. The device according to claim 5, further comprising: The word vector calculation module is used to obtain the preset sememe, and calculate the word vector of each preset vocabulary according to the stochastic gradient descent method and the preset sememe. 7. The device according to claim 5, further comprising: The sememe vocabulary matrix acquisition module is used to obtain the sememe vocabulary matrix according to the preset sememe vector and the unknown sememe vocabulary vector; A co-occurrence matrix calculation module, configured to calculate the co-occurrence matrix of the sememe vocabulary matrix according to the sememe vocabulary matrix; A matrix decomposition module is used to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method to obtain a second sememe vector; A target value calculation module, configured to calculate a target value according to the unknown sememe word vector and the second sememe vector; A target sememe vector calculation module, configured to calculate and obtain a target sememe vector according to the target value and the first sememe vector; Wherein, the sememe vocabulary matrix is represented by 0 and 1, 1 indicates that the corresponding vocabulary includes the corresponding sememe, and 0 indicates that the corresponding vocabulary does not include the corresponding sememe. 8. The device according to claim 7, wherein the matrix decomposition module is specifically used to decompose the sememe vocabulary matrix and the co-occurrence matrix respectively according to the stochastic gradient descent method and the loss function, to obtain the first ambiguous original vector; Wherein, the loss function is: L＝∑ _{w∈W，s∈S} (W·(S+S')+b _w +b _s -M _ws ) ² +λ∑ _s，t∈S (s·tC _st ) ² W is the unknown sememe vocabulary vector, S is the first pre-set sememe vector and the second pre-set sememe vector, S is the first pre-set sememe vector, λ is the pre-set coefficient, M _ws , C _st , w, s are all Elements in the semantic original vocabulary matrix, the co-occurrence matrix, the unknown semantic original vocabulary vector and the first preset semantic original vector, b _w is the bias of the unknown semantic original vocabulary vector, and b _s is The offset of the first predefinition vector.