CN108804512B - Apparatus, method and computer-readable storage medium for generating text classification model - Google Patents

Abstract

The invention discloses a text classification model generating device, comprising a memory and a processor. The memory stores a model generating program that can be run on the processor. When the program is executed by the processor, the following steps are implemented: obtaining word segmentation in the financial field. Dictionary and text corpus in the financial field; select candidate new words from the text corpus to add to the word segmentation dictionary; obtain a sample set and label the training samples in the sample set; use the preset word segmentation based on the word segmentation dictionary with candidate new words added The algorithm performs word segmentation on the training samples in the sample set and extracts word vectors. Based on the adaboost algorithm, the word vectors and labeled category information are input into multiple weak classifiers for training to obtain a text classification model. The invention also provides a method for generating a text classification model and a computer-readable storage medium. The invention solves the problem in the prior art that the emotional tendency classification of texts in the financial field cannot be realized.

Description

Apparatus, method and computer-readable storage medium for generating text classification model

technical field

The present invention relates to the technical field of text classification, and in particular, to an apparatus, method and computer-readable storage medium for generating a text classification model.

Background technique

With the development of the Internet and information technology, more and more institutions and individuals express their views, attitudes and positions on various things in various ways through the Internet, such as various news reviews, forums and social networking sites. These massive amounts of information have certain commercial value for e-commerce, market forecasting and other aspects, especially the financial industry, which is the industry with the fastest growth and the greatest impact on Internet information. In-depth research has gradually become an important topic.

Text sentiment orientation analysis is a part of text sentiment analysis. Through sentiment orientation analysis, we can grasp the positive and negative tendencies of this article. For the financial field, news public opinion reflects the prosperity of the market and industry and the enthusiasm of investors for trading. An important indicator, therefore, the analysis of the emotional tendencies of texts in the financial field has a significant impact on the study of financial duration. However, there is still a lack of solutions for classifying emotional tendencies of texts in the financial field in the prior art, which makes it impossible to achieve Sentiment-oriented classification of texts in the financial domain.

SUMMARY OF THE INVENTION

The present invention provides a text classification model generation device, method and computer-readable storage medium, the main purpose of which is to propose a text classification model generation device that can be used for emotional tendency classification of texts in the financial field, so as to solve the problem of the prior art In this paper, it is impossible to realize the classification of sentimental tendency of texts in the financial field.

In order to achieve the above object, the present invention provides a text classification model generation device, the device includes a memory and a processor, the memory stores a model generation program that can be run on the processor, and the model generation program is The processor implements the following steps when executing:

Obtain a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field;

According to a preset algorithm, candidate new words are selected from the text corpus and added to the word segmentation dictionary;

Obtain a sample set, and perform category labeling on the training samples in the sample set according to a preset emotional tendency classification mode;

Based on the word segmentation dictionary to which candidate new words are added, using a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set;

The word vector is extracted according to the word segmentation result. Based on the adaboost algorithm, the word vector corresponding to the training sample and the labeled category information are input into the preset multiple weak classifiers for training, and the multiple weak classifiers obtained by training are combined into the financial field. text classification model.

Optionally, the step of selecting candidate new words from the text corpus according to a preset algorithm, and adding them to the word segmentation dictionary includes:

Based on the word segmentation dictionary, the word segmentation algorithm is used to perform word segmentation processing on the text corpus, and a candidate word set is obtained according to the word segmentation result;

Calculate the information gain of each candidate word in the candidate word set, select a candidate word whose information gain is greater than a first preset threshold as the first candidate new word, and add the first candidate new word to the word segmentation dictionary;

Based on the word segmentation dictionary to which the first candidate new word is added, the word segmentation algorithm is used to segment the text corpus, and the word vector model is trained using the segmented text corpus;

Using the word vector model obtained by training to calculate the semantic similarity between the word in the word segmentation result and the first candidate new word;

A word with a semantic similarity greater than a second preset threshold is used as a second candidate new word, and the second candidate new word is added to the word segmentation dictionary.

Optionally, the processor can also be configured to execute the model generation program, so that the word whose semantic similarity is greater than the second preset threshold is used as the second candidate new word, and the second candidate new word is used as the second candidate new word. After the step of adding to the word dictionary, the following steps are also implemented:

Calculate the word frequency of the second candidate new word in the text corpus, and use the calculated word frequency as the weight of the second candidate new word in the word segmentation dictionary.

Optionally, in the acquisition of the sample set, the step of labeling the training samples in the sample set according to a preset emotional tendency classification mode includes:

Obtain a sample set, and obtain a plurality of labeling information obtained by labeling the training samples in the sample set by a plurality of labelers according to the preset sentiment tendency classification mode, and select the labeling information with the most occurrences as the corresponding labeling information from the plurality of labeling information The annotation results of the training samples.

Optionally, the weak classifier includes a classifier based on a convolutional neural network algorithm, a classifier based on a recurrent neural network algorithm, and a classifier based on a long short-term memory network algorithm.

In addition, in order to achieve the above purpose, the present invention also provides a method for generating a text classification model, the method comprising:

Obtain a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field;

According to a preset algorithm, candidate new words are selected from the text corpus and added to the word segmentation dictionary;

Obtain a sample set, and perform category labeling on the training samples in the sample set according to a preset emotional tendency classification mode;

Based on the word segmentation dictionary to which candidate new words are added, using a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set;

The word vector is extracted according to the word segmentation result. Based on the adaboost algorithm, the word vector corresponding to the training sample and the labeled category information are input into the preset multiple weak classifiers for training, and the multiple weak classifiers obtained by training are combined into the financial field. text classification model.

Optionally, the step of selecting candidate new words from the text corpus according to a preset algorithm, and adding them to the word segmentation dictionary includes:

Based on the word segmentation dictionary, the word segmentation algorithm is used to perform word segmentation processing on the text corpus, and a candidate word set is obtained according to the word segmentation result;

Calculate the information gain of each candidate word in the candidate word set, select a candidate word whose information gain is greater than a first preset threshold as the first candidate new word, and add the first candidate new word to the word segmentation dictionary;

Based on the word segmentation dictionary to which the first candidate new word is added, the word segmentation algorithm is used to segment the text corpus, and the word vector model is trained using the segmented text corpus;

Using the word vector model obtained by training to calculate the semantic similarity between the word in the word segmentation result and the first candidate new word;

A word with a semantic similarity greater than a second preset threshold is used as a second candidate new word, and the second candidate new word is added to the word segmentation dictionary.

Optionally, after the step of taking a word with a semantic similarity greater than a second preset threshold as a second candidate new word and adding the second candidate new word to the word dictionary, the method further includes the step of: :

Calculate the word frequency of the second candidate new word in the text corpus, and use the calculated word frequency as the weight of the second candidate new word in the word segmentation dictionary.

Optionally, in the acquisition of the sample set, the step of labeling the training samples in the sample set according to a preset emotional tendency classification mode includes:

Obtain a sample set, and obtain a plurality of labeling information obtained by labeling the training samples in the sample set by a plurality of labelers according to the preset sentiment tendency classification mode, and select the labeling information with the most occurrences as the corresponding labeling information from the plurality of labeling information The annotation results of the training samples.

In addition, in order to achieve the above object, the present invention also provides a computer-readable storage medium, where a model generation program is stored on the computer-readable storage medium, and the model generation program can be executed by one or more processors to achieve The steps of the generation method of the text classification model as described above.

The device, method and computer-readable storage medium for generating a text classification model proposed by the present invention construct a word segmentation dictionary in the financial field based on the collected vocabulary in the financial field, obtain a preset text corpus in the financial field, and obtain a candidate word set according to the text corpus, Select candidate new words from the candidate word set to add to the word segmentation dictionary. Obtain a sample set, and label the training samples in the sample set according to the preset sentiment tendency classification mode. Based on the word segmentation dictionary with the candidate new words added, the preset word segmentation algorithm is used to segment the training samples in the sample set, and word vectors are extracted according to the word segmentation results. Based on the adaboost algorithm, the word vectors corresponding to the training samples and the labeled category information are input Training is performed on multiple preset weak classifiers, and the multiple weak classifiers obtained by training are combined into a text classification model in the financial field. In the solution of the present invention, the candidate new words are selected and added to the word segmentation dictionary by mining the text corpus in the financial field, the word segmentation of the samples in the sample set is processed by the updated word segmentation dictionary, and the samples are classified according to the preset emotional tendency classification mode. The centralized sample data is classified into categories, and finally a text classification model is obtained by training, which can be applied to the classification of emotional tendencies in the financial field.

Description of drawings

1 is a schematic diagram of a preferred embodiment of a device for generating a text classification model according to the present invention;

2 is a schematic diagram of a program module of a model generation program in an embodiment of a text classification model generation device of the present invention;

FIG. 3 is a flowchart of a preferred embodiment of a method for generating a text classification model according to the present invention.

The realization, functional characteristics and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The present invention provides a device for generating a text classification model. Referring to FIG. 1 , it is a schematic diagram of a preferred embodiment of an apparatus for generating a text classification model according to the present invention.

In this embodiment, the device for generating the text classification model may be a PC (Personal Computer, personal computer), or may be a terminal device such as a smart phone, a tablet computer, or a portable computer. The generating apparatus 1 of the text classification model includes at least a memory 11 , a processor 12 , a communication bus 13 , and a network interface 14 .

The memory 11 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (eg, SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, and the like. In some embodiments, the memory 11 may be an internal storage unit of the apparatus 1 for generating a text classification model, such as a hard disk of the apparatus 1 for generating a text classification model. In other embodiments, the memory 11 may also be an external storage device of the text classification model generating apparatus 1, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a security Digital (Secure Digital, SD) card, flash memory card (Flash Card) and so on. Further, the memory 11 may also include both an internal storage unit of the text classification model generating apparatus 1 and an external storage device. The memory 11 can be used not only to store application software installed in the text classification model generating apparatus 1 and various types of data, such as the code of the model generating program 01 , but also to temporarily store data that has been output or will be output.

The processor 12 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor or other data processing chips in some embodiments, for running program codes or processing stored in the memory 11 Data, such as the execution model generation program 01, etc.

The communication bus 13 is used to realize the connection communication between these components.

The network interface 14 may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), and is usually used to generate a communication connection between the apparatus 1 and other electronic devices.

FIG. 1 only shows the generating apparatus 1 of the text classification model with components 11-14 and the model generating program 01, but it should be understood that it is not required to implement all the shown components, and more or less may be implemented instead. s component.

Optionally, the device 1 may further include a user interface, and the user interface may include a display (Display), an input unit such as a keyboard (Keyboard), and an optional user interface may also include a standard wired interface and a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, and the like. The display may also be appropriately called a display screen or a display unit, for displaying the information processed in the text classification model generating apparatus 1 and for displaying a visual user interface.

In the embodiment of the device 1 shown in FIG. 1 , the model generation program 01 is stored in the memory 11; the processor 12 implements the following steps when executing the model generation program 01 stored in the memory 11:

A1. Obtain a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field.

First, obtain a word segmentation dictionary in all fields, and on the basis of the word segmentation dictionary in all fields, add the collected vocabulary in the financial field to form a word segmentation dictionary in the financial field. Among them, the sources of vocabulary in the financial field mainly include the following three categories: professional terms in the financial field, such as "William's indicator", "moving average", "convertible bond", etc.; financial forum terms, such as some users in stock speculation forums commenting on A term for stocks; Internet terms and specific symbols applied to the financial sector, such as "junk stocks", etc.

A2. Select candidate new words from the text corpus according to a preset algorithm, and add them to the word segmentation dictionary.

On the basis of the above word segmentation dictionary, candidate new words are selected from the new text prediction and added to the current word segmentation dictionary. Specifically, step A2 includes:

A21. Based on the word segmentation dictionary, use the word segmentation algorithm to perform word segmentation on the text corpus, and obtain a candidate word set according to the word segmentation result; A22, calculate the information gain of each candidate word in the candidate word set, and select information A candidate word whose gain is greater than the first preset threshold is used as the first candidate new word, and the first candidate new word is added to the word segmentation dictionary; A23. Based on the word segmentation dictionary to which the first candidate new word is added, use The word segmentation algorithm performs word segmentation on the text corpus, and uses the word segmentation processed text corpus to train a word vector model; A24, using the word vector model obtained by training to calculate the semantics of the word in the word segmentation result and the first candidate new word Similarity; A25. Use a word whose semantic similarity is greater than a second preset threshold as a second candidate new word, and add the second candidate new word to the word segmentation dictionary.

Get the text corpus used to augment the word segmentation dictionary. Specifically, a large amount of financial news text information related to the financial topic to be analyzed is captured from a financial website by means of a web crawler to form a text corpus. Preprocess the crawled data, remove useless information such as garbled symbols and web escape symbols, and retain text data as text corpus. Next, classify the sentiment tendency of a large amount of text data in the text corpus by manual annotation, that is, add category annotation information to the text data.

Using the current word segmentation dictionary as the dictionary of the preset word segmentation algorithm, the word segmentation process is performed on the text corpus, and then the stop words in the word segmentation result are filtered according to the preset stop word list to remove irrelevant words in the result. A candidate word set is formed from the remaining word segmentation results. The category annotation information corresponding to the word segmentation result is consistent with the category annotation information of the corresponding text data.

Next, candidate new words are selected from the candidate word set according to the information gain. Information gain is an entropy-based evaluation method. When it is used for feature selection, it measures whether the appearance of a word is useful for judging a text. Whether it belongs to the amount of information provided by a certain class; it is defined as the difference between the amount of information before and after a feature value appears in the document, and the calculation formula is:

表示特征项t_i不出现的概率，

表示特征项t_i出现在不属于类别C_j的文档中的概率，|c|为类别的总数。其中，类别是指情感倾向的分类，特征项是候选词。上述概率值都可以通过对候选词在文本语料中的统计情况计算得到。In the above formula, P(C _j ) represents the probability that the category C _j appears in the data set, P(t _i ) represents the probability that the feature item t _i appears in the data set, and P(C _j |t _i ) represents the feature item t _i the probability of appearing in a document judged to be class C _j ,

represents the probability that the feature item t _i does not appear,

Represents the probability that the feature item t _i appears in the document that does not belong to the category C _j , |c| is the total number of categories. Among them, the category refers to the classification of emotional tendencies, and the feature items are candidate words. The above probability values can be obtained by calculating the statistics of the candidate words in the text corpus.

The usefulness of the candidate word is judged according to the calculated information gain. The larger the value of the information gain, the more useful it is for classification. A candidate word whose information gain is greater than the first preset threshold in the candidate word set is added to the current word segmentation dictionary as the first candidate new word to realize the expansion of the word segmentation dictionary.

Based on the above word segmentation dictionary after vocabulary expansion, use the same word segmentation algorithm to perform word segmentation processing on the same text corpus, obtain word segmentation results, use the word segmentation processed text corpus to train a word vector model, and use the trained word vector model to calculate word segmentation to get The word vector of each word is calculated according to the word vector, and the semantic similarity of the first candidate new word of the word obtained by the word segmentation process is calculated. If the semantic similarity is greater than the second preset threshold, it will be used as the second candidate new word. The second candidate word selected in the result is added to the word segmentation dictionary to realize the re-expansion of the word segmentation dictionary.

It is understandable that after the word segmentation algorithm is used to segment the text corpus, the stop words in the word segmentation result will be deleted through the stop word list, because these stop words are noisy and meaningless for text classification, so delete them. These words can improve the accuracy of text classification, while reducing the amount of calculation when selecting candidate new words.

After the above steps, three expansions of the word segmentation dictionary are actually realized. The first is to obtain the vocabulary in the financial field through manual collection for preliminary expansion. The second is to select new words by calculating the information gain. The third It is to calculate the semantic similarity through the word vector and select new words again. In addition, the second and third expansions are re-expands based on the word segmentation dictionary after the previous expansion. Through the above dynamic expansion method, as many new financial domain words as possible are selected from the corpus. The word segmentation algorithm of the expanded word segmentation dictionary is used for word segmentation of the training samples of the classification model. The richer the financial domain vocabulary in the word segmentation dictionary, the more accurate the word segmentation result of the financial domain text, and the higher the classification accuracy of the trained classification model. .

Optionally, as an embodiment, after the second candidate new word is selected and added to the word segmentation dictionary, the word frequency of the second candidate new word in the text corpus is calculated, and the word frequency is used as the second candidate new word in the word segmentation. Weights in the dictionary. For the first candidate new word, the word frequency can be calculated in the same way as its weight in the word segmentation dictionary.

A3. Obtain a sample set, and perform category labeling on the training samples in the sample set according to a preset emotional tendency classification mode.

Obtain a sample set for training the text classification model, and for each piece of training data in the sample set, obtain multiple labelers for each piece of training data according to the preset sentiment tendency classification mode, and select multiple labeling information. The annotation information with the most occurrences is used as the annotation result of the training data. Among them, the user can set the corresponding sentiment classification mode according to the financial issue to be analyzed, for example, the text in the stock forum is divided into hold, sell and buy; the stock discussion text in Weibo or the forum is divided into positive , Negative and Neutral; divide financial news text into positive, negative and neutral, etc.

A4. Based on the word segmentation dictionary to which candidate new words are added, use a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set.

A5. Extract the word vector according to the word segmentation result. Based on the adaboost algorithm, input the word vector corresponding to the training sample and the labeled category information into the preset multiple weak classifiers for training, and combine the multiple weak classifiers obtained by training as Text classification models in finance.

After completing the labeling of the training samples, based on the word segmentation dictionary that has been expanded many times, the preset word segmentation algorithm is used to process the word segmentation of the training samples, and the word vector of the word segmentation result is extracted by using the word vector model after training. It should be noted that the word segmentation algorithm used in the solution of this embodiment is always the same algorithm.

In this embodiment, in order to improve the accuracy of the text classification model, the word2vec model and the Glove (Global Vectors for word representation) model are respectively used to extract word vectors, and two word vectors are obtained for each word segmentation result. In addition, in this embodiment, the classifier based on the convolutional neural network algorithm, the classifier based on the cyclic neural network algorithm, and the classifier based on the long short-term memory network algorithm are used as weak classifiers. For each weak classifier, using the above two word vectors as input, six weak classification models can actually be constructed. Based on the adaboost algorithm, each weak classifier is trained using the samples in the sample set. In the training process, if a sample has been accurately classified, then in constructing the next sample set, reduce the weight of the sample; if a sample has not been accurately classified, in constructing the next sample set, increase the weight of the sample weight. The sample set with updated weights is used to train the next classifier, and the whole training process goes on iteratively. In addition, after the training process of each weak classifier is completed, the weight of the weak classifier with a small classification error rate is increased, so that it plays a larger decisive role in the final classification function, and the weak classification with a large classification error rate is reduced. The weight of the classifier so that it plays a less decisive role in the final classification function. Each weak classifier is iteratively trained as described above. The weak classifiers obtained from each training are fused as the final text classification model. The text classification model can be used to classify the sentimental tendency of texts in the financial field, and can be used to judge whether the stock discussion texts in the forum are negative, positive or neutral.

The device for generating a text classification model proposed in this embodiment selects as many new words in the financial field as possible from the corpus by mining the text corpus in the financial field, and adds them to the word segmentation dictionary, thereby realizing the expansion of the financial field word segmentation dictionary. And use the word segmentation dictionary with the expanded financial vocabulary to perform word segmentation processing on the training samples in the sample set, and label the sample data in the sample set according to the preset sentiment tendency classification mode, and finally train a text classification model, which can be applied to The problem of sentiment orientation classification in the financial domain.

Optionally, in other embodiments, the model generation program can also be divided into one or more modules, one or more modules are stored in the memory 11, and are processed by one or more processors (in this embodiment, The processor 12) is executed to complete the present invention. The module referred to in the present invention refers to a series of computer program instruction segments capable of accomplishing specific functions, and is used to describe the execution process of the model generation program in the text classification model generation device.

For example, referring to FIG. 2, it is a schematic diagram of program modules of a model generation program in an embodiment of a text classification model generation device of the present invention. In this embodiment, the model generation program can be divided into a data acquisition module 10, a new word selection module Module 20, sample labeling module 30, sample word segmentation module 40 and model training module 50, exemplarily:

The data acquisition module 10 is used for: acquiring a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field;

The new word selection module 20 is used to: select candidate new words from the text corpus according to a preset algorithm, and add them to the word segmentation dictionary;

The sample labeling module 30 is configured to: obtain a sample set, and perform category labeling on the training samples in the sample set according to a preset emotional tendency classification mode;

The sample word segmentation module 40 is configured to: use a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set based on the word segmentation dictionary to which candidate new words are added;

The model training module 50 is used for: extracting word vectors according to the word segmentation results, and based on the adaboost algorithm, inputting the word vectors corresponding to the training samples and the labeled category information into a plurality of preset weak classifiers for training, and training a plurality of weak classifiers obtained by training. Weak classifiers are combined into a text classification model in the financial domain.

The functions or operation steps realized when the program modules such as the data acquisition module 10, the new word selection module 20, the sample labeling module 30, the sample word segmentation module 40, and the model training module 50 are executed are substantially the same as those in the above-mentioned embodiment, and are not repeated here. Repeat.

In addition, the present invention also provides a method for generating a text classification model. Referring to FIG. 3 , it is a flowchart of a preferred embodiment of a method for generating a text classification model according to the present invention. The method may be performed by an apparatus, which may be implemented in software and/or hardware.

In this embodiment, the method for generating a text classification model includes:

Step S10, acquiring a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field.

First, obtain a word segmentation dictionary in all fields, and on the basis of the word segmentation dictionary in all fields, add the collected vocabulary in the financial field to form a word segmentation dictionary in the financial field. Among them, the sources of vocabulary in the financial field mainly include the following three categories: professional terms in the financial field, such as "William's indicator", "moving average", "convertible bond", etc.; financial forum terms, such as some users in stock speculation forums commenting on A term for stocks; Internet terms and specific symbols applied to the financial sector, such as "junk stocks", etc.

Step S20: Select candidate new words from the text corpus according to a preset algorithm and add them to the word segmentation dictionary.

On the basis of the above word segmentation dictionary, candidate new words are selected from the new text prediction and added to the current word segmentation dictionary. Specifically, step S20 includes: using the word segmentation algorithm to perform word segmentation on the text corpus based on the word segmentation dictionary, and obtaining a candidate word set according to the word segmentation result; calculating the information gain of each candidate word in the candidate word set , select the candidate word whose information gain is greater than the first preset threshold as the first candidate new word, and add the first candidate new word to the word segmentation dictionary; based on the word segmentation dictionary to which the first candidate new word is added, Use the word segmentation algorithm to segment the text corpus, and use the word segmentation processed text corpus to train a word vector model; use the word vector model obtained by training to calculate the word in the word segmentation result The semantic similarity of the first candidate new word degree; take the word whose semantic similarity is greater than the second preset threshold as the second candidate new word, and add the second candidate new word to the word segmentation dictionary.

Get the text corpus used to augment the word segmentation dictionary. Specifically, a large amount of financial news text information related to the financial topic to be analyzed is captured from a financial website by means of a web crawler to form a text corpus. Preprocess the crawled data, remove useless information such as garbled symbols and web escape symbols, and retain text data as text corpus. Next, classify the sentiment tendency of a large amount of text data in the text corpus by manual annotation, that is, add category annotation information to the text data.

Using the current word segmentation dictionary as the dictionary of the preset word segmentation algorithm, the word segmentation process is performed on the text corpus, and then the stop words in the word segmentation result are filtered according to the preset stop word list to remove irrelevant words in the result. A candidate word set is formed from the remaining word segmentation results. The category annotation information corresponding to the word segmentation result is consistent with the category annotation information of the corresponding text data.

Next, candidate new words are selected from the candidate word set according to the information gain. Information gain is an entropy-based evaluation method. When it is used for feature selection, it measures whether the appearance of a word is useful for judging a text. Whether it belongs to the amount of information provided by a certain class; it is defined as the difference between the amount of information before and after a feature value appears in the document, and the calculation formula is:

表示特征项t_i不出现的概率，

表示特征项t_i出现在不属于类别C_j的文档中的概率，|c|为类别的总数。其中，类别是指情感倾向的分类，特征项是候选词。上述概率值都可以通过对候选词在文本语料中的统计情况计算得到。In the above formula, P(C _j ) represents the probability that the category C _j appears in the data set, P(t _i ) represents the probability that the feature item t _i appears in the data set, and P(C _j |t _i ) represents the feature item t _i the probability of appearing in a document judged to be class C _j ,

represents the probability that the feature item t _i does not appear,

Represents the probability that the feature item t _i appears in the document that does not belong to the category C _j , |c| is the total number of categories. Among them, the category refers to the classification of emotional tendencies, and the feature items are candidate words. The above probability values can be obtained by calculating the statistics of the candidate words in the text corpus.

The usefulness of the candidate word is judged according to the calculated information gain. The larger the value of the information gain, the more useful it is for classification. A candidate word whose information gain is greater than the first preset threshold in the candidate word set is added to the current word segmentation dictionary as the first candidate new word to realize the expansion of the word segmentation dictionary.

Based on the above word segmentation dictionary after vocabulary expansion, use the same word segmentation algorithm to perform word segmentation processing on the same text corpus, obtain word segmentation results, use the word segmentation processed text corpus to train a word vector model, and use the trained word vector model to calculate word segmentation to get The word vector of each word is calculated according to the word vector, and the semantic similarity of the first candidate new word of the word obtained by the word segmentation process is calculated. If the semantic similarity is greater than the second preset threshold, it will be used as the second candidate new word. The second candidate word selected in the result is added to the word segmentation dictionary to realize the re-expansion of the word segmentation dictionary.

It is understandable that after the word segmentation algorithm is used to segment the text corpus, the stop words in the word segmentation result will be deleted through the stop word list, because these stop words are noisy and meaningless for text classification, so delete them. These words can improve the accuracy of text classification, while reducing the amount of calculation when selecting candidate new words.

After the above steps, three expansions of the word segmentation dictionary are actually realized. The first is to obtain the vocabulary in the financial field through manual collection for preliminary expansion. The second is to select new words by calculating the information gain. The third It is to calculate the semantic similarity through the word vector and select new words again. In addition, the second and third expansions are re-expands based on the word segmentation dictionary after the previous expansion. Through the above dynamic expansion method, as many new financial domain words as possible are selected from the corpus. The word segmentation algorithm of the expanded word segmentation dictionary is used for word segmentation of the training samples of the classification model. The richer the financial domain vocabulary in the word segmentation dictionary, the more accurate the word segmentation result of the financial domain text, and the higher the classification accuracy of the trained classification model. .

Optionally, as an embodiment, after the second candidate new word is selected and added to the word segmentation dictionary, the word frequency of the second candidate new word in the text corpus is calculated, and the word frequency is used as the second candidate new word in the word segmentation. Weights in the dictionary. For the first candidate new word, the word frequency can be calculated in the same way as its weight in the word segmentation dictionary.

In step S30, a sample set is obtained, and the training samples in the sample set are classified according to a preset emotional tendency classification mode.

Obtain a sample set for training the text classification model, and for each piece of training data in the sample set, obtain multiple labelers for each piece of training data according to the preset sentiment tendency classification mode, and select multiple labeling information. The annotation information with the most occurrences is used as the annotation result of the training data. Among them, the user can set the corresponding sentiment tendency classification mode according to the financial issue to be analyzed, for example, divide the text of the stock forum into hold, sell and buy; divide the text of the Weibo stock discussion into positive, negative and neutral; Divide financial news texts into positive, negative, neutral, etc.

Step S40, based on the word segmentation dictionary to which the candidate new words are added, using a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set.

Step S50, extracting word vectors according to the word segmentation results, and based on the adaboost algorithm, inputting the word vectors corresponding to the training samples and the labeled category information into a plurality of preset weak classifiers for training, and combining the plurality of weak classifiers obtained by training A text classification model for the financial domain.

After completing the labeling of the training samples, based on the word segmentation dictionary that has been expanded many times, the preset word segmentation algorithm is used to process the word segmentation of the training samples, and the word vector of the word segmentation result is extracted by using the word vector model after training. It should be noted that the word segmentation algorithm used in the solution of this embodiment is always the same algorithm.

In this embodiment, in order to improve the accuracy of the text classification model, the word2vec model and the Glove model are respectively used to extract word vectors, and two word vectors are obtained for each word segmentation result. In addition, in this embodiment, the classifier based on the convolutional neural network algorithm, the classifier based on the cyclic neural network algorithm, and the classifier based on the long short-term memory network algorithm are used as weak classifiers. For each weak classifier, using the above two word vectors as input, six weak classification models can actually be constructed. Based on the adaboost algorithm, each weak classifier is trained using the samples in the sample set. In the training process, if a sample has been accurately classified, then in constructing the next sample set, reduce the weight of the sample; if a sample has not been accurately classified, in constructing the next sample set, increase the weight of the sample weight. The sample set with updated weights is used to train the next classifier, and the whole training process goes on iteratively. In addition, after the training process of each weak classifier is completed, the weight of the weak classifier with a small classification error rate is increased, so that it plays a larger decisive role in the final classification function, and the weak classification with a large classification error rate is reduced. The weight of the classifier so that it plays a less decisive role in the final classification function. Each weak classifier is iteratively trained as described above. The weak classifiers obtained from each training are fused as the final text classification model. The text classification model can be used to classify the sentimental tendency of texts in the financial field, and can be used to judge whether the stock discussion texts in the forum are negative, positive or neutral.

The method for generating a text classification model proposed in this embodiment, by mining the text corpus in the financial field, selects as many new financial field words as possible from the corpus, and adds them to the word segmentation dictionary, so as to realize the expansion of the financial field word segmentation dictionary, And use the word segmentation dictionary with the expanded financial vocabulary to perform word segmentation processing on the training samples in the sample set, and label the sample data in the sample set according to the preset sentiment tendency classification mode, and finally train a text classification model, which can be applied to The problem of sentiment orientation classification in the financial domain.

In addition, an embodiment of the present invention also provides a computer-readable storage medium, where a model generation program is stored on the computer-readable storage medium, and the model generation program can be executed by one or more processors to achieve the following operations:

Obtain a word segmentation dictionary in the financial field constructed based on the collected vocabulary in the financial field, and a preset text corpus in the financial field;

According to a preset algorithm, candidate new words are selected from the text corpus and added to the word segmentation dictionary;

Obtain a sample set, and perform category labeling on the training samples in the sample set according to a preset emotional tendency classification mode;

Based on the word segmentation dictionary to which candidate new words are added, using a preset word segmentation algorithm to perform word segmentation processing on the training samples in the sample set;

The word vector is extracted according to the word segmentation result. Based on the adaboost algorithm, the word vector corresponding to the training sample and the labeled category information are input into the preset multiple weak classifiers for training, and the multiple weak classifiers obtained by training are combined into the financial field. text classification model.

The specific implementations of the computer-readable storage medium of the present invention are basically the same as the above-mentioned embodiments of the apparatus and method for generating a text classification model, and will not be described in detail here.

It should be noted that the above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages and disadvantages of the embodiments. And the terms "comprising", "comprising" or any other variation thereof herein are intended to encompass a non-exclusive inclusion such that a process, device, article or method comprising a list of elements includes not only those elements, but also includes no explicit Other elements listed, or those inherent to such a process, apparatus, article, or method are also included. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, apparatus, article, or method that includes the element.

From the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course hardware can also be used, but in many cases the former is better implementation. Based on such understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM) as described above. , magnetic disk, optical disc), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields , are similarly included in the scope of patent protection of the present invention.

Claims (8)

1. An apparatus for generating a text classification model, the apparatus comprising a memory and a processor, the memory having stored thereon a model generator executable on the processor, the model generator when executed by the processor implementing the steps of:

acquiring a word segmentation dictionary of the financial field constructed based on the collected financial field vocabularies and a preset text corpus of the financial field;

selecting new candidate words from the text corpus according to a preset algorithm, adding the new candidate words to the segmentation dictionary, selecting new candidate words from the text corpus according to the preset algorithm, and adding the new candidate words to the segmentation dictionary comprises the following steps:

based on the word segmentation dictionary, performing word segmentation processing on the text corpus by using the word segmentation algorithm, and acquiring a candidate word set according to the word segmentation result;

calculating information gain of each candidate word in the candidate word set, selecting a candidate word with the information gain larger than a first preset threshold value as a first candidate new word, and adding the first candidate new word into the word segmentation dictionary;

based on the word segmentation dictionary added with the first candidate new word, performing word segmentation on the text corpus by using the word segmentation algorithm, and training a word vector model by using the text corpus after word segmentation processing;

calculating semantic similarity between the words in the word segmentation result and the first candidate new words by using a word vector model obtained by training;

taking words with semantic similarity larger than a second preset threshold as second candidate new words, and adding the second candidate new words into the word segmentation dictionary; the information gain calculation formula of each candidate word is as follows:

in the above formula, P (C)_j) Represents class C_jProbability of occurrence in the dataset, P (t)_i) Representing a feature item t_iProbability of occurrence in the data set, P (C)_j|t_i) Representing a feature item t_iPresent in decision class C_jThe probability in the document of (a) is,

representing a feature item t_iThe probability of the non-occurrence of the event,

representing a feature item t_iPresent in a category other than C_jThe probability, | c | is the total number of categories, wherein the categories refer to the classification of emotional tendency, the feature items are candidate words, and the probability values can be obtained by calculating the statistical conditions of the candidate words in the text corpus;

acquiring a sample set, and carrying out class marking on training samples in the sample set according to a preset emotional tendency classification mode;

based on the word segmentation dictionary added with the candidate new words, performing word segmentation processing on training samples in the sample set by using a preset word segmentation algorithm;

extracting word vectors according to word segmentation results, inputting the word vectors corresponding to training samples and labeled class information into a plurality of preset weak classifiers for training based on an adaboost algorithm, and combining the plurality of weak classifiers obtained by training into a text classification model in the financial field.

2. The apparatus for generating a text classification model according to claim 1, wherein the processor is further operative to execute the model generation program to, after the step of taking the word having a semantic similarity greater than a second preset threshold as a second candidate new word and adding the second candidate new word to the word dictionary, further implement the steps of:

and calculating the word frequency of the second candidate new word in the text corpus, and taking the calculated word frequency as the weight of the second candidate new word in the word segmentation dictionary.

3. The apparatus for generating a text classification model according to any one of claims 1 or 2, wherein the step of obtaining a sample set and labeling training samples in the sample set according to a preset emotional tendency classification mode comprises:

the method comprises the steps of obtaining a sample set, obtaining a plurality of marking information obtained by marking training samples in the sample set by a plurality of marking persons according to a preset emotional tendency classification mode, and selecting the marking information with the largest occurrence frequency from the marking information as a marking result of the corresponding training sample.

4. The apparatus for generating a text classification model according to any one of claims 1 or 2, characterized in that the weak classifiers include a convolutional neural network algorithm-based classifier, a cyclic neural network algorithm-based classifier, and a long-short term memory network algorithm-based classifier.

5. A method for generating a text classification model, the method comprising:

acquiring a word segmentation dictionary of the financial field constructed based on the collected financial field vocabularies and a preset text corpus of the financial field;

selecting new candidate words from the text corpus according to a preset algorithm, adding the new candidate words to the segmentation dictionary, selecting new candidate words from the text corpus according to the preset algorithm, and adding the new candidate words to the segmentation dictionary comprises the following steps:

based on the word segmentation dictionary, performing word segmentation processing on the text corpus by using the word segmentation algorithm, and acquiring a candidate word set according to the word segmentation result;

calculating information gain of each candidate word in the candidate word set, selecting a candidate word with the information gain larger than a first preset threshold value as a first candidate new word, and adding the first candidate new word into the word segmentation dictionary;

based on the word segmentation dictionary added with the first candidate new word, performing word segmentation on the text corpus by using the word segmentation algorithm, and training a word vector model by using the text corpus after word segmentation processing;

calculating semantic similarity between the words in the word segmentation result and the first candidate new words by using a word vector model obtained by training;

taking words with semantic similarity larger than a second preset threshold as second candidate new words, and adding the second candidate new words into the word segmentation dictionary; the information gain calculation formula of each candidate word is as follows:

in the above formula, P (C)_j) Represents class C_jProbability of occurrence in the dataset, P (t)_i) Representing a feature item t_iProbability of occurrence in the data set, P (C)_j|t_i) Representing a feature item t_iPresent in decision class C_jThe probability in the document of (a) is,

representing a feature item t_iThe probability of the non-occurrence of the event,

representing a feature item t_iPresent in a category other than C_jThe probability, | c | is the total number of categories, wherein the categories refer to the classification of emotional tendency, the feature items are candidate words, and the probability values can be obtained by calculating the statistical conditions of the candidate words in the text corpus;

acquiring a sample set, and carrying out class marking on training samples in the sample set according to a preset emotional tendency classification mode;

based on the word segmentation dictionary added with the candidate new words, performing word segmentation processing on training samples in the sample set by using a preset word segmentation algorithm;

extracting word vectors according to word segmentation results, inputting the word vectors corresponding to training samples and labeled class information into a plurality of preset weak classifiers for training based on an adaboost algorithm, and combining the plurality of weak classifiers obtained by training into a text classification model in the financial field.

6. The method for generating a text classification model according to claim 5, wherein after the step of taking the word with semantic similarity greater than a second preset threshold as a second candidate new word and adding the second candidate new word to the word dictionary, the method further comprises the steps of:

and calculating the word frequency of the second candidate new word in the text corpus, and taking the calculated word frequency as the weight of the second candidate new word in the word segmentation dictionary.

7. The method for generating the text classification model according to any one of claims 5 or 6, wherein the step of obtaining the sample set and labeling the training samples in the sample set according to the preset emotional tendency classification mode comprises the following steps:

the method comprises the steps of obtaining a sample set, obtaining a plurality of marking information obtained by marking training samples in the sample set by a plurality of marking persons according to a preset emotional tendency classification mode, and selecting the marking information with the largest occurrence frequency from the marking information as a marking result of the corresponding training sample.

8. A computer-readable storage medium having stored thereon a model generation program executable by one or more processors to perform the steps of the method of generating a text classification model according to any one of claims 5 to 7.

Images