CN111738791A

CN111738791A - Text processing method, device, equipment and storage medium

Info

Publication number: CN111738791A
Application number: CN202010065630.2A
Authority: CN
Inventors: 徐松; 李浩然; 袁鹏
Original assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Current assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Priority date: 2020-01-20
Filing date: 2020-01-20
Publication date: 2020-10-02

Abstract

The embodiment of the invention discloses a text processing method, a text processing device, text processing equipment and a storage medium, wherein the method comprises the following steps: acquiring a target title text for describing a target object; determining a target sequence labeling result corresponding to the target title text based on a preset sequence labeling network model, wherein the target sequence labeling result is used for representing the position of a selling point keyword in the target title text; determining a target selling point keyword in a target title text according to a target sequence labeling result; the preset sequence labeling network model is obtained in advance according to sample data training, and the sample data comprises a sample title text and actual article conversion rates corresponding to the preset selling point keywords. By the technical scheme of the embodiment of the invention, higher-quality selling point keywords can be extracted, and the conversion rate of the articles is further improved.

Description

Text processing method, device, equipment and storage medium

Technical Field

The present invention relates to computer technologies, and in particular, to a text processing method, apparatus, device, and storage medium.

Background

With the rapid development of computer technology, more and more users prefer to purchase items online in order to save time and physical strength.

Generally, a title text and an item picture of each item are shown in a browsing page of the item so that a user can roughly understand item information. The title text of the article is often longer, so that the browsing speed is slower, the title text of the article needs to be compressed and extracted, the shorter title text is obtained, and the shorter title text is displayed, so that the user can quickly know the characteristics of the article, the purchasing behavior of the user is promoted, and the conversion rate of the article is improved.

At present, when the selling point keywords for describing the characteristics of the article in the title text are extracted, the selling point keywords with the highest use frequency in the title text are generally determined by using a rule matching mode.

However, in the process of implementing the present invention, the inventor finds that at least the following problems exist in the prior art:

although the selling point keywords in the title text extracted in the existing mode are some high-frequency words, displaying the high-frequency words does not mean that the conversion rate of the article is high, and many of the selling point keywords with high conversion rates are usually some low-frequency words, such as long-tail selling point keywords, so that the selling point keywords extracted in the existing mode are not really wanted selling point keywords, and the extraction accuracy is reduced.

Disclosure of Invention

The embodiment of the invention provides a text processing method, a text processing device, text processing equipment and a storage medium, which are used for extracting selling point keywords with higher quality, improving the extraction accuracy and further improving the conversion rate of articles.

In a first aspect, an embodiment of the present invention provides a text processing method, including:

acquiring a target title text for describing a target object;

determining a target sequence labeling result corresponding to the target title text based on a preset sequence labeling network model, wherein the target sequence labeling result is used for representing the position of a selling point keyword in the target title text;

determining a target selling point keyword in the target title text according to the target sequence labeling result;

the preset sequence labeling network model is obtained in advance according to sample data training, and the sample data comprises a sample title text and actual article conversion rates corresponding to the preset selling point keywords.

In a second aspect, an embodiment of the present invention further provides a text processing apparatus, including:

the target title text acquisition module is used for acquiring a target title text for describing a target object;

a target sequence labeling result determining module, configured to determine a target sequence labeling result corresponding to the target title text based on a preset sequence labeling network model, where the target sequence labeling result is used to represent a position of a selling point keyword in the target title text;

the target selling point keyword determining module is used for determining a target selling point keyword in the target title text according to the target sequence labeling result;

In a third aspect, an embodiment of the present invention further provides an apparatus, where the apparatus includes:

one or more processors;

a memory for storing one or more programs;

when executed by the one or more processors, cause the one or more processors to implement a text processing method as provided by any of the embodiments of the invention.

In a fourth aspect, the embodiments of the present invention further provide a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the text processing method provided in any embodiment of the present invention.

The embodiment of the invention has the following advantages or beneficial effects:

the preset sequence labeling network model is obtained by training in advance based on the actual article conversion rate corresponding to the sample title text and each preset selling point keyword, so that the article conversion rate of an output result is also trained in the process of training the preset sequence labeling network model, and a target sequence result corresponding to a target title text determined by the preset sequence labeling network model obtained after training is the best labeling result taking the article conversion rate as an index into consideration, so that the determined target selling point keyword is a higher-quality selling point keyword, the extraction accuracy is improved, the purchasing behavior of a user can be further promoted by displaying the higher-quality target selling point keyword, and the article conversion rate is improved.

Drawings

Fig. 1 is a flowchart of a text processing method according to an embodiment of the present invention;

fig. 2 is a flowchart of a text processing method according to a second embodiment of the present invention;

FIG. 3 is an example of a pre-set sequence tagging network model obtained by training according to a second embodiment of the present invention;

fig. 4 is a flowchart of a training process of a preset sequence tagging network model according to a third embodiment of the present invention;

fig. 5 is an example of a network model labeled with a preset sequence in a training process according to a third embodiment of the present invention;

fig. 6 is a flowchart of a text processing method according to a fourth embodiment of the present invention;

FIG. 7 is an example of a mapping between candidate brand keywords and standard brand keywords according to a fourth embodiment of the present invention;

fig. 8 is a schematic structural diagram of a text processing apparatus according to a fifth embodiment of the present invention;

fig. 9 is a schematic structural diagram of an apparatus according to a sixth embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.

Example one

Fig. 1 is a flowchart of a text processing method according to an embodiment of the present invention, which is applicable to a case where a title text describing an article is extracted to obtain a high-quality selling point keyword. The method can be executed by a text processing device, which can be implemented by software and/or hardware, integrated in a device with data processing function. The method specifically comprises the following steps:

and S110, acquiring a target title text for describing the target object.

The target item may refer to any item displayed in the item browsing page. The target title text may be text for describing contents of category information and property information of the target item. For example, when the target item is a headset, the corresponding target heading text may be: the built-in bone vocal print sensor can accurately acquire bone vocal print information of a speaker when speaking, and clear voice communication is brought. The mobile phone is taken out, unlocking is not needed, only one sentence of 'WeChat payment' is needed, the two-dimensional code interface is opened, and quick payment is achieved.

S120, determining a target sequence labeling result corresponding to the target title text based on the preset sequence labeling network model, wherein the target sequence labeling result is used for representing the position of the selling point keyword in the target title text.

The selling point keyword may refer to a keyword used for describing characteristics of the target item in the target title text. One or more sell point keywords may be included in the target title text. The preset sequence labeling network model may be a network model for labeling a tag corresponding to each heading word in the heading text sequence and for named entity recognition. The preset sequence labeling network model can be obtained in advance according to sample data training, and the sample data can comprise a sample title text and actual article conversion rates corresponding to the preset selling point keywords. The preset selling point keywords can be preset for all articles and can be used as vocabularies of the selling point keywords. The actual conversion rate of the articles corresponding to the preset selling point keyword may be an actual conversion rate of the articles purchased after the user browses the articles after the preset selling point keyword is displayed as a final selling point keyword, and may be determined based on the user browsing frequency and the user purchasing frequency which are counted in advance.

Specifically, the embodiment can convert the task of extracting the selling point keywords into the sequence labeling task, so that the selling point keywords in the target title text can be determined more accurately and quickly by using the preset sequence labeling network model. When the preset sequence labeling network model is trained on the basis of the actual article conversion rate corresponding to the sample title text and each preset selling point keyword, the article conversion rate of an output result can be trained, so that the target sequence result corresponding to the target title text determined by the preset sequence labeling network model obtained after training is the optimal labeling result taking the article conversion rate as an index into consideration, and the target sequence labeling result is more accurate.

And S130, determining a target selling point keyword in the target title text according to the target sequence labeling result.

The target selling point keyword can be a selling point keyword which is extracted from the target title text and is finally displayed. Specifically, according to the positions of the selling point keywords in the target title text represented by the target sequence labeling result, all candidate selling point keywords in the target title text can be accurately determined, and the best target selling point keywords are determined from all candidate selling point keywords, so that the extraction accuracy is improved, and the purchasing behavior of the user can be further promoted by displaying the better target selling point keywords, and further the conversion rate of the articles is improved.

Illustratively, the target sequence annotation result may include a target label corresponding to each heading word in the target heading text; the target label may include a first preset label, a second preset label or a third preset label; the first preset label can be used for representing a beginning word of a selling point keyword in a target title text; the second preset label can be used for representing intermediate words or final words except for the first word in the selling point keywords in the target title text; the third preset label can be used for representing other subject words except the selling point key words in the target subject text.

Wherein each of the point keywords in the target title text may include at least one title word. When the selling point keyword only comprises one entry word, determining a target label corresponding to the entry word as a first preset label; when the selling point keyword includes at least two entry words, the target tag corresponding to a first entry word (i.e., a beginning word) of the at least two entry words may be determined as a first preset tag, and the target tag corresponding to each entry word (i.e., an intermediate word or an ending word) of the at least two entry words except the first entry word may be determined as a second preset tag. And determining the target label corresponding to each other entry word except for each selling point keyword in the target title text as a third preset label, so that each selling point keyword in the target title text can be determined based on the target label corresponding to each entry word in the target title text. For example, the target labels in this embodiment may adopt a labeling manner of { B, I, O }, that is, the first preset label is B, the second preset label is I, and the third preset label is O, and assuming that the target sequence labeling result corresponding to the target title text is obiioooo, it indicates that the target title text includes a selling point keyword, and the candidate selling point keyword is composed of the second, third, and fourth title words in the target title text.

It should be noted that the embodiment may use the first preset tag to distinguish different selling point keywords in the target title text. The number of the selling point keywords contained in the target title text is equal to the number of the first preset labels contained in the target sequence labeling result.

Illustratively, S130 may include: determining candidate selling point keywords in the target title text according to the entry word corresponding to the first preset label and the entry word corresponding to the second preset label in the target sequence labeling result, wherein the candidate selling point keywords consist of at least one entry word; if only one candidate selling point keyword exists, determining the candidate selling point keyword as a target selling point keyword in the target title text; and if at least two candidate selling point keywords exist, determining the target selling point keywords in the target title text according to the target label prediction probability corresponding to each candidate selling point keyword in each target sequence labeling result.

Specifically, in the target sequence labeling result, if there is no adjacent second preset tag after the first preset tag, the entry word corresponding to the first preset tag may be determined as a candidate selling point keyword; if an adjacent second preset label exists behind the first preset label, the entry word corresponding to the first preset label and the entry word corresponding to each adjacent and continuous second preset label behind the first preset label can form a candidate selling point keyword, so that each candidate selling point keyword in the target title text can be determined. If only one candidate selling point keyword exists in the target title text, the candidate selling point keyword can be directly determined as the target selling point keyword in the target title text. If at least two candidate selling point keywords exist in the target title text, the best candidate selling point keyword in the candidate selling point keywords can be determined as the target selling point keyword according to the target label prediction probability corresponding to each candidate selling point keyword in the target sequence labeling result.

Illustratively, determining the target selling point keyword in the target title text according to the target label prediction probability corresponding to each candidate selling point keyword in each target sequence labeling result may include: determining a comprehensive prediction probability corresponding to each candidate selling point keyword according to a target label prediction probability corresponding to each entry word in each candidate selling point keyword in a target sequence labeling result; and determining the candidate selling point key words with the highest comprehensive prediction probability as the target selling point key words in the target title text.

The target label prediction probability may be a probability that a label corresponding to a header word is a target label predicted based on a preset sequence labeling network model.

Specifically, the embodiment may obtain the target label prediction probability corresponding to each heading word in the target heading text based on the output of the preset sequence tagging network model. For each candidate selling point keyword, determining a comprehensive prediction probability corresponding to the candidate selling point keyword according to the target label prediction probability corresponding to each entry word forming the candidate selling point keyword. For example, the target label prediction probabilities corresponding to the respective heading words that constitute the candidate selling point keyword may be multiplied, and the obtained multiplication result may be determined as the comprehensive prediction probability corresponding to the candidate selling point keyword. When the comprehensive prediction probability is higher, the predicted candidate selling point keywords are more accurate, and the quality is higher, so that the candidate selling point keywords with the highest comprehensive prediction probability can be determined as the target selling point keywords, and the best target selling point keywords in the target title text are obtained.

In the technical scheme of this embodiment, the preset sequence labeling network model is obtained by training in advance based on the sample title text and the actual article conversion rate corresponding to each preset selling point keyword, so that the article conversion rate of the output result is also trained in the process of training the preset sequence labeling network model, and further, the target sequence result corresponding to the target title text determined by using the preset sequence labeling network model obtained after training is the best labeling result in consideration of the article conversion rate, so that the determined target selling point keyword is a higher-quality selling point keyword, thereby improving the extraction accuracy, and by displaying the higher-quality target selling point keyword, the purchasing behavior of the user can be further promoted, and further the article conversion rate is improved.

Example two

Fig. 2 is a flowchart of a text processing method according to a second embodiment of the present invention, where in this embodiment, on the basis of the foregoing embodiment, the presetting of a sequence tagging network model may include: and on the basis of the preset feature extraction submodel and the preset discriminant probability submodel, optimizing the determining mode of the target sequence labeling result corresponding to the target title text. Wherein explanations of the same or corresponding terms as those of the above-described embodiments are omitted.

Referring to fig. 2, the text processing method provided in this embodiment includes the following steps:

and S210, acquiring a target title text for describing the target object.

S220, extracting the sub-models based on the preset features, and determining a target feature vector corresponding to each title word in the target title text.

The preset feature extraction submodel may be a network submodel for extracting features corresponding to each entry word in a preset sequence labeling network model to obtain a corresponding target feature vector. For example, the preset feature extraction submodel may include: a preset language processing submodel and a preset recurrent neural network submodel. The preset language processing submodel can perform language processing on the title text so as to obtain data of inputtable types of the preset recurrent neural network submodel, namely to obtain a title word vector corresponding to each title word in the title text. The preset cyclic neural network submodel can be a neural network model for extracting feature vectors corresponding to the entry words based on a deep learning mode. For example, the predetermined language processing submodel may be, but is not limited to, a BERT (bidirectional Encoder retrieval from transforms) bi-directional Encoder model; the preset recurrent neural network submodel may be, but is not limited to, a Bi-directional Long-short term Memory (Bi-LSTM) bidirectional Long-short term neural network model.

Specifically, in this embodiment, word segmentation processing may be performed on the target title text based on the word segmentation dictionary, each entry word in the target title text is determined, a title word sequence is obtained, the title word sequence is input into the preset feature extraction submodel, and a target feature vector corresponding to each entry word may be obtained according to the output of the preset feature extraction submodel, so that a target feature vector corresponding to each entry word may be effectively extracted in a deep learning manner.

Illustratively, S220 may include: performing word segmentation processing on the target title text, and determining each title word in the target title text; inputting each entry word into a preset language processing submodel, and determining an entry word vector corresponding to each entry word according to the output of the preset language processing submodel; and inputting each entry word vector into a preset cyclic neural network submodel, and determining a target characteristic vector corresponding to each entry word according to the output of the preset cyclic neural network submodel.

Specifically, when the word segmentation is performed on the target title text, each word in the target title text can be used as one entry word, and each entry word under the word granularity is obtained, so that the problem that the labeled boundary is influenced due to errors Of a word segmentation dictionary can be avoided, the problem Of data sparsity can be solved, and the appearance Of Out-Of-Vocabulary (OOV) outside a thesaurus is reduced. The method includes the steps that firstly, each entry word (namely, entry word sequence) is input into a preset language processing submodel, and entry word vectors corresponding to the entry words are obtained according to the output of the preset language processing submodel; inputting each entry word vector (namely entry word vector sequence) into a preset cyclic neural network submodel, and obtaining a target feature vector corresponding to each entry word according to the output of the preset cyclic neural network submodel.

And S230, determining a target sequence labeling result corresponding to the target title text based on each target feature vector and the preset discriminant probability submodel.

The preset discriminant probability submodel may be a network submodel for predicting a label corresponding to each word in a preset sequence labeling network model. For example, the predetermined discriminant probability submodel may be, but is not limited to, a CRF (conditional random fields) conditional random field model. The conditional random field model may determine, based on the target feature vector corresponding to the entry word and the target feature vectors corresponding to other entry words adjacent to the entry word, a prediction probability that the entry word is labeled as each preset label, and determine, based on each prediction probability, a target label corresponding to the entry word, for example, using the preset label with the highest prediction probability as the target label corresponding to the entry word, thereby obtaining a target sequence labeling result.

Specifically, each target feature vector (i.e., a target feature vector sequence) may be input into the preset discriminant probability submodel, and a target sequence labeling result corresponding to the target title text may be obtained according to the output of the preset discriminant probability submodel. The target sequence labeling result may include a target label corresponding to each of the entry words in the target entry text and a prediction probability that the entry word is labeled as each of the preset labels.

S240, determining a target selling point keyword in the target title text according to the target sequence labeling result.

Illustratively, fig. 3 shows an example of a network model for training the obtained preset sequence labels. As shown in FIG. 3, the preset sequence labeling network model may include a BERT bidirectional encoder model, a Bi-LSTM bidirectional long-and-short time memory cyclic neural network model and a CRF conditional random field model. As shown in fig. 3, after training a BERT bidirectional encoder model, a Bi-LSTM bidirectional long-and-short memory cyclic neural network model and a CRF conditional random field model based on sample data, word segmentation processing may be performed on a target title text to obtain each title word in the target title text; inputting each entry word into a BERT bidirectional encoder model for encoding to obtain an entry word vector corresponding to each entry word; inputting each entry word vector into a Bi-LSTM bidirectional long-and-short time memory cyclic neural network model for feature extraction, and obtaining a target feature vector corresponding to each entry word; and inputting each target feature vector into a CRF conditional random field model for sequence labeling, and obtaining a target sequence labeling result corresponding to a target title text, namely a target label corresponding to each title word in the target title text according to the output of the CRF conditional random field model, so that the best quality target selling point keyword can be accurately determined based on the target sequence labeling result.

According to the technical scheme of the embodiment, the target sequence labeling result after the index of the item conversion rate is considered can be more accurately determined by extracting the submodel based on the preset features and the preset discriminant probability submodel, so that the extraction accuracy of the key words of the selling points is further improved, and the conversion rate of the items is further improved.

EXAMPLE III

Fig. 4 is a flowchart of a training process of a preset sequence tagging network model according to a third embodiment of the present invention, where on the basis of the foregoing embodiments, the training process of the preset sequence tagging network model is described in detail based on sample data before the preset sequence tagging network model is used. Wherein explanations of the same or corresponding terms as those of the above embodiments are omitted.

Referring to fig. 4, the training process of the preset sequence labeling network model provided in this embodiment specifically includes the following steps:

s410, determining a sample sequence labeling result corresponding to the sample title text based on the preset sequence labeling network model.

Specifically, the network structure of the preset sequence tagging network model can be constructed based on business requirements. For example, the preset sequence labeling network model may include a preset feature extraction sub-model and a preset discriminant probability sub-model. The preset feature extraction submodel may include a preset language processing submodel and a preset recurrent neural network submodel. For example, the preset language processing submodel may be, but is not limited to, a BERT bidirectional encoder model; the preset recurrent neural network submodel may be, but is not limited to, a Bi-LSTM Bi-directional long-and-short memory recurrent neural network model. The pre-defined discriminative probability submodel may be, but is not limited to, a CRF conditional random field model. After the preset sequence labeling network model is constructed, the sample title text is subjected to word segmentation processing based on the word segmentation processing mode of the target title text in the embodiment to obtain each sample word in the sample title text, each sample word is input into the preset sequence labeling network model, and a sample sequence labeling result corresponding to the sample title text is obtained according to the output of the preset sequence labeling network model. The sample sequence labeling result may include a sample label corresponding to each sample word in the sample title text and a prediction probability that the sample word is labeled as each preset label.

Illustratively, fig. 5 shows an example of a network model labeled by a preset sequence in a training process. As shown in FIG. 5, the preset sequence labeling network model may be composed of a BERT bidirectional encoder model, a Bi-LSTM bidirectional long-and-short time memory cyclic neural network model, and a CRF conditional random field model. As shown in fig. 5, in the training process, word segmentation processing may be performed on the sample title text to obtain each sample word in the sample title text; inputting each sample word into a BERT bidirectional encoder model for encoding to obtain a sample word vector corresponding to each sample word; inputting each sample word vector into a Bi-LSTM bidirectional long-time memory cyclic neural network model for feature extraction, and obtaining a sample feature vector corresponding to each sample word; and inputting each sample feature vector into a CRF (conditional random field) model for sequence labeling, and obtaining a sample sequence labeling result corresponding to the sample title text according to the output of the CRF model, namely a sample label corresponding to each title word in the sample title text, so that each selling point keyword in the sample title text can be determined based on the sample sequence labeling result.

And S420, determining a first training error corresponding to the sample title text according to the preset linear regression model, the sample sequence labeling result, the actual article conversion rate corresponding to each preset selling point keyword and the sample feature vector corresponding to each sample word in the sample title text.

The preset linear regression model can be a network model which is constructed based on a linear regression algorithm and used for predicting the item conversion rate corresponding to the selling point keywords. The actual conversion rate of the articles corresponding to each preset selling point keyword is an actual conversion rate of the articles purchased by the user after the user browses the articles after the preset selling point keyword is used as a final selling point keyword for display, and may be determined based on the user browsing times and the user purchasing times which are counted in advance, for example, the ratio of the user purchasing times to the user browsing times is used as the actual conversion rate corresponding to the preset selling point keyword. The sample feature vector corresponding to each sample word in the sample title text can be obtained through a preset feature extraction submodel in a preset sequence labeling network model, and specifically can be obtained through a preset recurrent neural network submodel. For example, in fig. 5, a sample feature vector corresponding to each sample word in the sample title text may be obtained according to the output of the Bi-LSTM bidirectional long-and-short memory recurrent neural network model. The first training error may be used to characterize a deviation of a predicted item conversion rate corresponding to a selling point keyword in the determined sample heading text from an actual item conversion rate.

Illustratively, S420 may include the following steps S421-S425 to implement a function of determining a first training error corresponding to the sample title text:

and S421, determining each sample selling point keyword in the sample title text according to the sample sequence labeling result.

The sample sequence labeling result may include a sample label corresponding to each sample word in the sample title text; the sample label comprises a first preset label, a second preset label or a third preset label; the first preset label is used for representing a beginning word of a selling point keyword in a sample title text; the second preset label is used for representing intermediate words or final words except for the first word in the selling point keywords in the sample title text; the third preset label is used for representing other sample words except the selling point key words in the sample title text. One or more sample sell point keywords may be included in the sample title text. Each sample point keyword may consist of at least one sample word.

Specifically, each sample selling point keyword in the sample title text may be determined based on the manner in which each candidate selling point keyword in the target title text is determined in the above embodiment. For example, the sample selling point keyword in the target title text can be determined according to the sample word corresponding to the first preset label and the sample word corresponding to the second preset label in the sample sequence labeling result. The number of the sample selling point keywords contained in the sample title text is equal to the number of the first preset labels contained in the sample sequence labeling result.

S422, determining sample feature vectors corresponding to all sample words in each sample selling point keyword according to the sample feature vectors corresponding to all sample words in the sample title text.

Specifically, a sub-model can be extracted based on preset features in a preset sequence labeling network model, and a sample feature vector corresponding to each sample word in a sample title text is determined, for example, a sample feature vector corresponding to each sample word in the sample title text can be obtained according to the output of a Bi-LSTM bidirectional long-and-short memory recurrent neural network model, and then a sample feature vector corresponding to each sample word included in each sample selling point keyword can be obtained.

And S423, determining the predicted article conversion rate corresponding to each sample selling point keyword according to each sample feature vector corresponding to each sample selling point keyword based on a preset linear regression model.

Specifically, as shown in fig. 5, for each sample selling point keyword, each sample feature vector corresponding to each sample word included in the sample selling point keyword may be input into a preset linear regression model to predict the article conversion rate, and according to the output of the preset linear regression model, the predicted article conversion rate cvr corresponding to the sample selling point keyword may be obtained. Similarly, the predicted article conversion rate corresponding to each sample selling point keyword can be determined, and the function of predicting the article conversion rate is realized.

S424, determining the actual article conversion rate corresponding to each sample selling point keyword according to the actual article conversion rate corresponding to each preset selling point keyword.

Specifically, each sample selling point keyword may be subjected to keyword matching with each preset selling point keyword, so that the actual article conversion rate corresponding to each sample selling point keyword may be determined based on the matching result.

It should be noted that the execution sequence of step S424 is not limited in this embodiment, for example, step S424 may be executed sequentially after step S423, may be executed before step S423, and may also be executed simultaneously with step S423.

And S425, determining a first training error corresponding to the sample title text according to the actual article conversion rate and the predicted article conversion rate corresponding to each sample selling point keyword.

Specifically, a selling point keyword error corresponding to each sample selling point keyword may be determined according to an actual article conversion rate and a predicted article conversion rate corresponding to each sample selling point keyword, for example, an absolute value of a difference between the actual article conversion rate and the predicted article conversion rate corresponding to the sample selling point keyword may be determined as the selling point keyword error corresponding to the sample selling point keyword, and a square value of a difference between the actual article conversion rate and the predicted article conversion rate corresponding to the sample selling point keyword may also be determined as the selling point keyword error corresponding to the sample selling point keyword. And adding the selling point keyword errors corresponding to the sample selling point keywords, wherein the addition result can be determined as a first training error corresponding to the sample title text, so that the article conversion rate condition corresponding to the sample selling point keywords can be measured.

And S430, determining a second training error corresponding to the sample title text according to the sample sequence labeling result and the standard sequence labeling result corresponding to the sample title text based on a preset loss function.

The preset loss function may be preset and used for measuring the accuracy of the extracted selling point keyword. The present embodiment may set the preset loss function based on KL (Kullback-Leibler) divergence, i.e., relative entropy. For example, the preset loss function may be set as follows:

wherein, y_iAnd the prediction probability p (y | x) of the sample label corresponding to the ith sample word in the sample sequence labeling result being the standard label is shown. y'_iAnd the standard probability p (y' | x) of the sample label corresponding to the ith sample word in the standard sequence labeling result being the standard label is shown. N represents the number of sample words in the sample title text. Standard probability y 'corresponding to each sample word in the embodiment'_iAre all 1, so the preset loss function can be abbreviated as:

the standard sequence labeling result corresponding to the sample title text may be a standard sequence obtained by labeling in advance based on the position of the actual selling point keyword in the sample title text. Before determining the second training error corresponding to the sample title text, the method may further include: and determining a standard sequence labeling result corresponding to the sample title text according to the actual article conversion rate corresponding to the sample title text and each preset selling point keyword. For example, the preset selling point keyword with the highest actual article conversion rate in the sample title text can be determined as the actual selling point keyword in the sample title text in advance based on the actual article conversion rate corresponding to each preset selling point keyword, so that model training is performed by using the preset selling point keyword with the highest article conversion rate, the training speed is further increased, and the trained preset sequence labeling network model can output the selling point keyword with the higher article conversion rate.

Specifically, the standard label corresponding to each sample word in the sample title text can be obtained according to the standard sequence labeling result corresponding to the sample title text, so that the prediction probability that the sample label corresponding to each sample word is the standard label can be obtained based on the sample sequence labeling result. For example, if the standard sequence labeling result is BIOOOO, the obtained prediction probabilities that the sample label corresponding to each sample word is the standard label are respectively: y is₁＝P(y₁＝B)＝0.9；y₂＝P(y₂＝I)＝0.8；y₃＝P(y₃＝O)＝0.2；y₄＝P(y₄＝O)＝0.3；y₅＝p(y₅＝O)＝0.9；y₆＝p(y₆＝O)＝0.9；y₇＝p(y₇O) ═ 0.8, it can be determined that the second training error corresponding to the sample title text is: - (log0.9+ log0.8+ log0.2+ log0.3+ log0.9+ log0.9+ log 0.8).

S440, taking the convergence condition of the first training error and the second training error as a training target, and training a preset sequence labeling network model and a preset linear regression model.

Specifically, it may be detected whether a convergence condition is currently reached based on the first training error and the second training error, such as whether a sum of errors of the first training error and the second training error is smaller than a preset error or whether an error change tends to be stable, or whether the current iteration number is equal to a preset number. If the convergence condition is detected to be currently achieved, for example, the sum of the errors of the first training error and the second training error is smaller than the preset error or the error change tends to be stable, or the current iteration number is equal to the preset number, which indicates that the training of the preset sequence labeling network model is completed, the iterative training may be stopped at this time. If the convergence condition is not reached currently, the errors of the first training error and the second training error can be reversely propagated to the preset sequence labeling network model and the preset linear regression model, and the network parameters in the preset sequence labeling network model and the preset linear regression model are adjusted until the convergence condition is reached.

According to the technical scheme of the embodiment, the preset linear regression model is used for predicting and training the article conversion rate of the output result in the training process of the preset sequence labeling network model, so that the target sequence labeling result output by the trained preset sequence labeling network model is the best result considering the article conversion rate as an index through a multi-task learning mode of the selling point keyword prediction and the article conversion rate prediction, the selling point keyword with high article conversion rate can be accurately extracted, and the article conversion rate is improved.

Example four

Fig. 6 is a flowchart of a text processing method according to a fourth embodiment of the present invention, where in this embodiment, a target brand keyword corresponding to a target article is determined based on the foregoing embodiments, and based on the target selling point keyword and the target brand keyword, a compressed title text obtained by compressing a target title text is determined and displayed. Wherein explanations of the same or corresponding terms as those of the above embodiments are omitted.

Referring to fig. 6, the text processing method provided in this embodiment specifically includes the following steps:

s610, obtaining a target title text for describing the target object.

S620, determining a target sequence labeling result corresponding to the target title text based on the preset sequence labeling network model, wherein the target sequence labeling result is used for representing the position of the selling point keyword in the target title text.

S630, determining the target selling point keywords in the target title text according to the target sequence labeling result.

And S640, acquiring a current brand keyword for describing brand information of the target object.

The current brand keyword may refer to a brand name to which the target item belongs, such as amantai. Specifically, the current brand keyword corresponding to the target item may be obtained based on brand information input by the creator when creating the target item information. If the target title text contains the current brand key words, the preset brand key words which are successfully matched in the target title text can be determined in a mode of matching with the preset brand key words, and the preset brand key words are used as the current brand key words.

S650, matching the current brand keywords with the candidate brand keywords according to the preset mapping relation between the candidate brand keywords and the standard brand keywords, and determining target brand keywords corresponding to the target object.

Because the brand keywords corresponding to the same article can exist in various expression forms, such as the brand keywords in the form of Chinese, the brand keywords in the form of English, the abbreviated brand keywords, the full-name brand keywords, and the like. The candidate brand keyword may refer to various expression forms of brand keywords corresponding to the same item, which may be determined based on various brand keywords input by the creator. The standard brand keyword may refer to a uniform brand keyword selected from among various candidate brand keywords based on business needs. For example, the candidate brand keyword with the highest frequency of use may be taken as the standard brand keyword; the candidate brand keyword with the shortest character length can also be used as the standard brand keyword so as to save the display space. FIG. 7 provides an example of a mapping between candidate brand keywords and standard brand keywords, as shown in FIG. 7, which may be represented by: amani, ARMANI and EMPORIO ARMANI, all mapped as standard brand keywords Amani.

Specifically, the current brand keyword may be matched with each candidate brand keyword, and the standard brand keyword corresponding to the candidate brand keyword that is successfully matched may be determined as the target brand keyword corresponding to the target article, so that the normalization of the brand keyword may be achieved. If direct regard as target brand keyword with current brand keyword, then the condition of spelling irregularity appears probably, the wrong problem of spelling appears even to this embodiment can map the candidate brand keyword of irregularity to standard brand keyword through the mapping mode of brand keyword, and then can solve the problem that brand keyword is irregular.

And S660, determining a title compressed text after the target title text is compressed according to the target selling point key words and the target brand key words.

Specifically, based on a preset compression template, a compressed title compressed text may be generated according to the target selling point keyword and the target brand keyword, where the compressed title compressed text includes: the target brand keywords and the target selling point keywords are adopted, so that the target title text is effectively compressed.

Exemplarily, the present embodiment may further include: determining candidate category keywords in the target title text for describing the item categories of the target item; if only one candidate category keyword exists, determining the candidate category keyword as a target category keyword corresponding to the target article; and if at least two candidate category keywords exist, determining the target category keywords corresponding to the target object according to the character length corresponding to each candidate category keyword.

The candidate category keyword may refer to a name of a category to which the target item belongs, such as a shoe, a headset, and the like. Specifically, each preset category keyword successfully matched in the target title text can be used as a candidate category keyword in a manner of matching each keyword in the target title text with a preset category keyword. When a candidate category keyword exists in the target title text, the candidate category keyword can be directly used as the target category keyword. When at least two candidate category keywords exist in the target title text, the character length corresponding to each candidate category keyword can be compared, and the candidate category keyword with the largest character length is determined as the target category keyword, namely the long-tail category keyword can be used as the target category keyword. For example, if the obtained candidate category keywords are: "shoes" and "basketball shoes", then "basketball shoes" can be determined as the target category keyword to improve the description accuracy of the article.

Accordingly, S660 may include: and determining a title compressed text after the target title text is compressed according to the target selling point key words, the target brand key words and the target category key words. For example, based on a preset compression template, a compressed title compressed text may be generated according to the target selling point keyword, the target brand keyword, and the target category keyword, where the compressed title compressed text includes: the target brand keywords and the target selling point keywords and the target category keywords are adopted, so that the target title text is effectively compressed.

And S670, displaying the title compressed text on the display interface.

Specifically, by displaying the compressed title compressed text on the display interface, the user can quickly know the characteristics of the article, the purchasing behavior of the user is promoted, and the conversion rate of the article is further improved.

According to the technical scheme, the current brand keywords are matched with the candidate brand keywords according to the mapping relation between the preset candidate brand keywords and the standard brand keywords, and the target brand keywords corresponding to the target object are determined, so that the normalization of the brand keywords can be realized, and the problem that the brand keywords are irregular can be solved. By displaying the compressed title compressed text on the display interface, the user can quickly know the characteristics of the article, the purchasing behavior of the user is promoted, and the conversion rate of the article is further improved.

The following is an embodiment of a text processing apparatus according to an embodiment of the present invention, which belongs to the same inventive concept as the text processing methods of the above embodiments, and reference may be made to the above embodiment of the text processing method for details that are not described in detail in the embodiment of the text processing apparatus.

EXAMPLE five

Fig. 8 is a schematic structural diagram of a text processing apparatus according to a fifth embodiment of the present invention, which is applicable to a case where a title text describing an article is extracted to obtain a high-quality selling point keyword, and the apparatus specifically includes: a target title text acquisition module 810, a target sequence annotation result determination module 820, and a target selling point keyword determination module 830.

The target title text obtaining module 810 is configured to obtain a target title text for describing a target item; a target sequence labeling result determining module 820, configured to determine a target sequence labeling result corresponding to the target title text based on a preset sequence labeling network model, where the target sequence labeling result is used to represent a location of a selling point keyword in the target title text; a target selling point keyword determining module 830, configured to determine a target selling point keyword in the target title text according to the target sequence labeling result; the preset sequence labeling network model is obtained in advance according to sample data training, and the sample data comprises a sample title text and actual article conversion rates corresponding to the preset selling point keywords.

Optionally, the preset sequence tagging network model includes: presetting a feature extraction submodel and a discriminant probability submodel; accordingly, the target sequence labeling result determining module 820 includes:

the target feature vector determining unit is used for extracting the sub-model based on the preset features and determining a target feature vector corresponding to each title word in the target title text;

and the target sequence labeling result determining unit is used for determining a target sequence labeling result corresponding to the target title text based on each target feature vector and the preset discriminant probability submodel.

Optionally, the preset feature extraction submodel includes: presetting a language processing submodel and a recurrent neural network submodel; correspondingly, the target feature vector determination unit is specifically configured to: performing word segmentation processing on the target title text, and determining each title word in the target title text; inputting each entry word into a preset language processing submodel, and determining an entry word vector corresponding to each entry word according to the output of the preset language processing submodel; and inputting each entry word vector into a preset cyclic neural network submodel, and determining a target characteristic vector corresponding to each entry word according to the output of the preset cyclic neural network submodel.

Optionally, the target sequence labeling result includes a target label corresponding to each heading word in the target heading text; the target label comprises a first preset label, a second preset label or a third preset label; the first preset label is used for representing a beginning word of a selling point keyword in the target title text; the second preset label is used for representing intermediate words or final words except for the first word in the selling point keywords in the target title text; the third preset label is used for representing other heading words except the selling point key words in the target heading text.

Optionally, the target selling point keyword determining module 830 includes:

the candidate selling point keyword determining unit is used for determining candidate selling point keywords in the target title text according to the entry word corresponding to the first preset label and the entry word corresponding to the second preset label in the target sequence labeling result, wherein the candidate selling point keywords consist of at least one entry word;

the first target selling point keyword determining unit is used for determining a candidate selling point keyword as a target selling point keyword in a target title text if only one candidate selling point keyword exists;

and the second target selling point keyword determining unit is used for determining the target selling point keywords in the target title text according to the target label prediction probability corresponding to each candidate selling point keyword in each target sequence labeling result if at least two candidate selling point keywords exist.

Optionally, the second target selling point keyword determining unit is specifically configured to: determining a comprehensive prediction probability corresponding to each candidate selling point keyword according to a target label prediction probability corresponding to each entry word in each candidate selling point keyword in a target sequence labeling result; and determining the candidate selling point key words with the highest comprehensive prediction probability as the target selling point key words in the target title text.

Optionally, the apparatus further includes a preset sequence labeling network model training module, including:

the sample sequence labeling result determining unit is used for determining a sample sequence labeling result corresponding to the sample title text based on a preset sequence labeling network model;

the first training error determining unit is used for determining a first training error corresponding to the sample title text according to a preset linear regression model, a sample sequence labeling result, the actual article conversion rate corresponding to each preset selling point keyword and a sample feature vector corresponding to each sample word in the sample title text;

the second training error determining unit is used for determining a second training error corresponding to the sample title text according to the sample sequence marking result and the standard sequence marking result corresponding to the sample title text based on a preset loss function;

and the model training unit is used for training the preset sequence labeling network model and the preset linear regression model by taking the convergence condition of the first training error and the second training error as a training target.

Optionally, the first training error determination unit is specifically configured to: determining each sample selling point keyword in the sample title text according to the sample sequence labeling result; determining a sample feature vector corresponding to each sample word in each sample selling point keyword according to the sample feature vector corresponding to each sample word in the sample title text; determining a predicted article conversion rate corresponding to each sample selling point keyword according to each sample feature vector corresponding to each sample selling point keyword based on a preset linear regression model; determining the actual article conversion rate corresponding to each sample selling point keyword according to the actual article conversion rate corresponding to each preset selling point keyword; and determining a first training error corresponding to the sample title text according to the actual article conversion rate and the predicted article conversion rate corresponding to each sample selling point keyword.

Optionally, the apparatus further comprises: and the standard sequence labeling result determining module is used for determining a standard sequence labeling result corresponding to the sample title text according to the sample title text and the actual article conversion rate corresponding to each preset selling point keyword before determining a second training error corresponding to the sample title text.

Optionally, the preset language processing sub-model is a BERT bidirectional encoder model; presetting a cyclic neural network sub-model as a bidirectional long-time memory cyclic neural network model; and presetting the discriminant probability submodel as a conditional random field model.

Optionally, the apparatus further comprises: the current brand keyword acquisition module is used for acquiring a current brand keyword for describing brand information of the target object; and the target brand keyword determining module is used for matching the current brand keyword with the candidate brand keyword according to the preset mapping relation between the candidate brand keyword and the standard brand keyword, and determining the target brand keyword corresponding to the target article.

Optionally, the apparatus further comprises: the title compressed text determination module is used for determining a title compressed text after the target title text is compressed according to the target selling point key words and the target brand key words; and the title compressed text display module is used for displaying the title compressed text on the display interface.

The text processing device provided by the embodiment of the invention can execute the text processing method provided by any embodiment of the invention, and has corresponding functional modules and beneficial effects of the execution method.

It should be noted that, in the embodiment of the text processing apparatus, the included modules are merely divided according to the functional logic, but are not limited to the above division as long as the corresponding functions can be implemented; in addition, the specific names of the functional modules are only for convenience of distinguishing from each other and are not used for limiting the protection scope of the present invention.

EXAMPLE six

Fig. 9 is a schematic structural diagram of an apparatus according to a sixth embodiment of the present invention. FIG. 9 illustrates a block diagram of an exemplary device 12 suitable for use in implementing embodiments of the present invention. The device 12 shown in fig. 9 is only an example and should not bring any limitation to the function and scope of use of the embodiments of the present invention.

As shown in FIG. 9, device 12 is in the form of a general purpose computing device. The components of device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 that couples various system components including the system memory 28 and the processing unit 16.

Bus 18 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example, such architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, micro-channel architecture (MAC) bus, enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

Device 12 typically includes a variety of computer system readable media. Such media may be any available media that is accessible by device 12 and includes both volatile and nonvolatile media, removable and non-removable media.

The system memory 28 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM)30 and/or cache memory 32. Device 12 may further include other removable/non-removable, volatile/nonvolatile computer system storage media. By way of example only, storage system 34 may be used to read from and write to non-removable, nonvolatile magnetic media (not shown in FIG. 9, and commonly referred to as a "hard drive"). Although not shown in FIG. 9, a magnetic disk drive for reading from and writing to a removable, nonvolatile magnetic disk (e.g., a "floppy disk") and an optical disk drive for reading from or writing to a removable, nonvolatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 by one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the invention.

A program/utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28, such program modules 42 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each of which examples or some combination thereof may comprise an implementation of a network environment. Program modules 42 generally carry out the functions and/or methodologies of the described embodiments of the invention.

Device 12 may also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), with one or more devices that enable a user to interact with device 12, and/or with any devices (e.g., network card, modem, etc.) that enable device 12 to communicate with one or more other computing devices. Such communication may be through an input/output (I/O) interface 22. Also, the device 12 may communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network, such as the Internet) via the network adapter 20. As shown, the network adapter 20 communicates with the other modules of the device 12 via the bus 18. It should be understood that although not shown in the figures, other hardware and/or software modules may be used in conjunction with device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

The processing unit 16 executes various functional applications and data processing by executing programs stored in the system memory 28, for example, to implement a text processing method provided by the embodiment of the present invention, the method including:

acquiring a target title text for describing a target object;

determining a target selling point keyword in a target title text according to a target sequence labeling result;

Of course, those skilled in the art can understand that the processor can also implement the technical solution of the text processing method provided in any embodiment of the present invention.

EXAMPLE seven

The present embodiment provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the text processing method steps as provided in any of the embodiments of the invention, the method comprising:

acquiring a target title text for describing a target object;

Computer storage media for embodiments of the invention may employ any combination of one or more computer-readable media. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer-readable storage medium may be, for example but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

A computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.

Computer program code for carrying out operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).

It will be understood by those skilled in the art that the modules or steps of the invention described above may be implemented by a general purpose computing device, they may be centralized on a single computing device or distributed across a network of computing devices, and optionally they may be implemented by program code executable by a computing device, such that it may be stored in a memory device and executed by a computing device, or it may be separately fabricated into various integrated circuit modules, or it may be fabricated by fabricating a plurality of modules or steps thereof into a single integrated circuit module. Thus, the present invention is not limited to any specific combination of hardware and software.

It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method of text processing, comprising:

acquiring a target title text for describing a target object;

2. The method of claim 1, wherein the pre-set sequence tagging network model comprises: presetting a feature extraction submodel and a discriminant probability submodel;

correspondingly, determining a target sequence labeling result corresponding to the target title text based on a preset sequence labeling network model, including:

determining a target feature vector corresponding to each title word in the target title text based on a preset feature extraction submodel;

and determining a target sequence labeling result corresponding to the target title text based on each target feature vector and a preset discriminant probability submodel.

3. The method of claim 2, wherein the preset feature extraction submodel comprises: presetting a language processing submodel and a recurrent neural network submodel;

correspondingly, the step of determining a target feature vector corresponding to each heading word in the target heading text based on the preset feature extraction submodel comprises the following steps:

performing word segmentation processing on the target title text, and determining each title word in the target title text;

inputting each entry word into a preset language processing submodel, and determining an entry word vector corresponding to each entry word according to the output of the preset language processing submodel;

and inputting each title word vector into a preset cyclic neural network submodel, and determining a target characteristic vector corresponding to each title word according to the output of the preset cyclic neural network submodel.

4. The method of claim 1, wherein the target sequence labeling result comprises a target label corresponding to each heading word in the target heading text; the target label comprises a first preset label, a second preset label or a third preset label;

the first preset label is used for representing a beginning word of a selling point keyword in the target title text;

the second preset label is used for representing intermediate words or final words except the beginning word in the selling point keywords in the target title text;

the third preset label is used for representing other subject words except the selling point key words in the target subject text.

5. The method of claim 4, wherein determining the target selling point keyword in the target title text according to the target sequence labeling result comprises:

determining candidate selling point keywords in the target title text according to the entry word corresponding to the first preset label and the entry word corresponding to the second preset label in the target sequence labeling result, wherein the candidate selling point keywords consist of at least one entry word;

if only one candidate selling point keyword exists, determining the candidate selling point keyword as a target selling point keyword in the target title text;

and if at least two candidate selling point keywords exist, determining the target selling point keywords in the target title text according to the target label prediction probability corresponding to each candidate selling point keyword in each target sequence labeling result.

6. The method according to claim 5, wherein determining the target selling point keyword in the target caption text according to the target label prediction probability corresponding to each candidate selling point keyword in the target sequence labeling result comprises:

determining a comprehensive prediction probability corresponding to each candidate selling point keyword according to a target label prediction probability corresponding to each entry word in each candidate selling point keyword in the target sequence labeling result;

and determining the candidate selling point keyword with the highest comprehensive prediction probability as a target selling point keyword in the target title text.

7. The method according to any one of claims 1 to 6, wherein the training process of the network model is labeled by the preset sequence, and comprises:

determining a sample sequence labeling result corresponding to the sample title text based on a preset sequence labeling network model;

determining a first training error corresponding to the sample title text according to a preset linear regression model, the sample sequence labeling result, the actual article conversion rate corresponding to each preset selling point keyword and a sample feature vector corresponding to each sample word in the sample title text;

determining a second training error corresponding to the sample title text according to the sample sequence marking result and a standard sequence marking result corresponding to the sample title text based on a preset loss function;

and taking the convergence condition of the first training error and the second training error as a training target, and training the preset sequence labeling network model and the preset linear regression model.

8. The method according to claim 7, wherein determining a first training error corresponding to the sample heading text according to a preset linear regression model, the sample sequence labeling result, the actual article conversion rate corresponding to each preset selling point keyword, and the sample feature vector corresponding to each sample word in the sample heading text comprises:

determining each sample selling point keyword in the sample title text according to the sample sequence labeling result;

determining a sample feature vector corresponding to each sample word in each sample selling point keyword according to the sample feature vector corresponding to each sample word in the sample title text;

based on a preset linear regression model, determining a predicted article conversion rate corresponding to each sample selling point keyword according to each sample feature vector corresponding to each sample selling point keyword;

determining the actual article conversion rate corresponding to each sample selling point keyword according to the actual article conversion rate corresponding to each preset selling point keyword;

and determining a first training error corresponding to the sample title text according to the actual article conversion rate and the predicted article conversion rate corresponding to each sample selling point keyword.

9. The method of claim 7, further comprising, prior to determining a second training error for the sample header text:

and determining a standard sequence labeling result corresponding to the sample title text according to the actual article conversion rate corresponding to the sample title text and each preset selling point keyword.

10. The method of claim 3, wherein the pre-set language processing sub-model is a BERT two-way coder model; the preset cyclic neural network submodel is a bidirectional long-time memory cyclic neural network model; the preset discriminant probability submodel is a conditional random field model.

11. The method according to any one of claims 1-6, further comprising:

acquiring a current brand keyword for describing brand information of the target item;

and matching the current brand key words with the candidate brand key words according to a preset mapping relation between the candidate brand key words and standard brand key words, and determining target brand key words corresponding to the target articles.

12. The method of claim 11, further comprising:

determining a title compressed text after the target title text is compressed according to the target selling point keyword and the target brand keyword;

and displaying the title compressed text on a display interface.

13. A text processing apparatus, comprising:

14. An apparatus, characterized in that the apparatus comprises:

one or more processors;

a memory for storing one or more programs;

when executed by the one or more processors, cause the one or more processors to implement a text processing method as recited in any of claims 1-12.

15. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out a text processing method according to any one of claims 1 to 12.