System and method for automatically generating knowledge-driven content planning

Abstract

A system and method for obtaining content planning. The system includes a computing device configured to: provide a label for each clause of each sentence of each product description of each product category, wherein the label includes an introductory label corresponding to a product attribute and a descriptive label corresponding to the product description; determine label exclusivity, label similarity, label distance, and label causality between labels; generate a candidate label sequence for the product description; and filter the candidate label sequence using label exclusivity, label similarity, label distance, and label causality to obtain a content planning.

Description

[0001] Cross-references

[0002] Several references are cited and discussed in the description of the present disclosure, which may include patents, patent applications, and various publications. Citation and / or discussion of such references are provided solely to clarify the description of the present disclosure, and no admission is made that any such reference is "prior art" to the disclosure described herein. All references cited and discussed in this specification are incorporated herein by reference in their entirety to the same extent as if each reference were incorporated by reference individually. Technical Field

[0003] The present disclosure relates generally to the field of content planning, and more particularly to a system and method for knowledge-driven content planning for automatically generating coherent product descriptions in e-commerce. Background Art

[0004] The background description provided here is to generally present the context of the present disclosure. Within the scope of the description in this background section, the work of the inventors, as well as the description that may not be in the prior art at the time of application, is neither explicitly nor implicitly admitted as prior art for the present disclosure.

[0005] Major e-commerce websites operate multiple product channels, such as recommendation texts. However, the generation of these content-based channels requires a lot of human involvement.

[0006] Therefore, there exists a need in the art to address the above-mentioned unresolved deficiencies and inadequacies. Summary of the invention

[0007] In some aspects, the present disclosure relates to a system for evaluating the text content of a product. In some embodiments, the system includes a computing device. The computing device includes a processor and a storage device for storing computer executable code, and the computer executable code is configured to:

[0008] Providing a label for each clause of each sentence of each product description of each product category, wherein the label includes a plurality of introductory labels and a plurality of descriptive labels, the plurality of introductory labels corresponding to industry attributes of the product category, and the plurality of descriptive labels corresponding to a plurality of descriptions of the product category;

[0009] Determining tag exclusivity between any two of the tags, wherein when the two tags are both the introductory tags, the two tags are defined as having the tag exclusivity;

[0010] Determine the tag similarity between any two tags based on the co-occurrence rate of the two tags in the product description;

[0011] determining a tag distance between any two tags in the product description based on the number of the clauses between the two tags, wherein the two tags co-occur;

[0012] Determining tag causality between any two tags in the product description based on a causal relationship between the two tags, wherein the two tags appear together;

[0013] generating a candidate tag sequence from each of the product descriptions using the tags in each of the product descriptions; and

[0014] The candidate tag sequence is filtered using the tag exclusivity, the tag similarity, the tag distance, and the tag causality to obtain the content plan.

[0015] In some embodiments, the product categories are large product categories, such as clothing, consumer electronics, home appliances, computers and office, sports and outdoor, health and beauty, toys and shoes. In some embodiments, the product categories are small product categories, such as refrigerators, air conditioners, televisions and washing machines.

[0016] In some embodiments, the product category is clothing, and the industry attributes include size, color, and material. In some embodiments, the product category is clothing, and the description includes function, style, feel, potential audience, and scenario.

[0017] In some embodiments, the computer executable code is configured to determine the tag similarity using Fisher's exact test. In some embodiments, the computer executable code is configured to determine the tag distance between the two tags by averaging the tag distances between the two tags in the product descriptions, where the two tags co-occur. In some embodiments, the computer executable code is configured to determine that the tag causality exists between the two tags when at least one of the terms "results from", "arise from", "because" and "due to" exists between the two tags in any of the product descriptions, where the two tags co-occur.

[0018] In some embodiments, the computer executable code is configured to generate the candidate tag sequence using N-gram. In some embodiments, N is a positive integer in the range of 3 to 10. In some embodiments, N is 5.

[0019] In some embodiments, the computer executable code is configured to filter the candidate tag sequences by:

[0020] removing candidate tag sequences containing two of the introductory tags; and

[0021] At least a tag sequence whose tag distance is less than a predefined tag distance threshold, a tag sequence whose tag similarity is greater than a predefined tag similarity threshold, or a tag sequence having the tag causality is retained.

[0022] In some embodiments, the computer executable code is further configured to generate a new product description based on the content plan.

[0023] In some aspects, the present disclosure relates to a method for generating a content plan. In some embodiments, the method includes:

[0024] Providing, by a computing device, a label for each clause of each sentence of each product description of each product category, wherein the label comprises a plurality of introductory labels and a plurality of descriptive labels, the plurality of introductory labels corresponding to industry attributes of the product category, and the plurality of descriptive labels corresponding to a plurality of descriptions of the product category;

[0025] Determining, by the computing device, the tag exclusivity between any two of the tags, wherein when the two tags are both the introductory tags, the two tags are defined as having the tag exclusivity;

[0026] Determining, by the computing device, a tag similarity between any two tags in the product description based on a common occurrence rate of the two tags;

[0027] determining, by the computing device, a tag distance between any two tags in the product description based on the number of the clauses between the two tags, wherein the two tags appear together;

[0028] Determining, by the computing device, tag causality between any two tags in the product description based on a causal relationship between the two tags, wherein the two tags appear together;

[0029] generating, by the computing device, a candidate tag sequence from each of the product descriptions using the tags in the each of the product descriptions; and

[0030] The computing device filters the candidate tag sequence using the tag exclusivity, the tag similarity, the tag distance, and the tag causality to obtain the content plan.

[0031] In some embodiments, the product categories include clothing (or apparel), refrigerators, air conditioners, televisions, and washing machines. In some embodiments, when the product category is clothing, the industry attributes include size, color, and material, and the description includes function, style, feel, potential audience, and scene.

[0032] In certain embodiments, the tag similarity is determined using Fisher's exact test; the tag distance between the two tags is determined by averaging the tag distance between the two tags in the product descriptions, wherein the two tags co-occur; and when at least one of the terms "results from", "originates from", "because" and "due to" exists between the two tags in any of the product descriptions, it is determined that tag causality exists between the two tags, wherein the two tags co-occur.

[0033] In some embodiments, the candidate tag sequence is generated using N-gram. In some embodiments, N is a positive integer in the range of 3 to 10. In some embodiments, N is 5.

[0034] In some embodiments, filtering the candidate tag sequences includes removing candidate tag sequences containing two of the introductory tags; and retaining at least tag sequences whose tag distance is less than a predefined tag distance threshold, tag sequences whose tag similarity is greater than a predefined tag similarity threshold, or tag sequences with the tag causality.

[0035] In some embodiments, the method further includes generating a new product description based on the content plan.

[0036] In some aspects, the present disclosure relates to a non-transitory computer-readable medium for storing computer-executable code. In some embodiments, when executed at a processor of a computing device, the computer-executable code is configured to perform the above method.

[0037] These and other aspects of the disclosure will become apparent from the following description of the preferred embodiments taken in conjunction with the following drawings and their headings, although changes and modifications may be made therein without departing from the spirit and scope of the novel concepts of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The content of the present disclosure can be more fully understood through the detailed description and the accompanying drawings. These drawings illustrate one or more embodiments of the present disclosure and are used together with the written description to explain the principles of the present disclosure. Wherever possible, the same reference numerals are used throughout the drawings to indicate the same or similar elements in one embodiment, wherein:

[0039] Figure 1 A system for acquiring content planning for a product category according to certain embodiments of the present disclosure is schematically depicted.

[0040] Figure 2 A knowledge learning module according to certain embodiments of the present disclosure is schematically depicted.

[0041] Figure 3A labeled product description according to certain embodiments of the present disclosure is schematically depicted.

[0042] Figure 4 Schematically depicts learned relationships between tags in a tagged product description according to certain embodiments of the present disclosure.

[0043] Figure 5 The method of generating a 2-gram to 5-gram tag sequence using a tag sequence having five consecutive tags according to some embodiments of the present disclosure is schematically depicted.

[0044] Figure 6 The generation of candidate tag sequences for product categories according to certain embodiments of the present disclosure is schematically depicted.

[0045] Fig. 7A Schematically depicting the use of learned knowledge to filter candidate tag sequences according to certain embodiments of the present disclosure.

[0046] Figure 7B Schematically depicting the use of learned knowledge to filter candidate tag sequences according to certain embodiments of the present disclosure.

[0047] Figure 7C Schematically depicting the use of learned knowledge to filter candidate tag sequences according to certain embodiments of the present disclosure.

[0048] Figure 8 Schematically depicting the use of learned knowledge to filter candidate tag sequences according to certain embodiments of the present disclosure.

[0049] Fig. 9 The architecture of a knowledge-driven content planning system according to some embodiments of the present disclosure is schematically depicted.

[0050] Fig.10 A method for generating a content plan according to some embodiments of the present disclosure is schematically depicted.

[0051] Fig.11 A comparison between content plans generated by a knowledge-based approach and a RNN-based approach according to the present disclosure is schematically depicted. DETAILED DESCRIPTION

[0052] The present disclosure is described in more detail in the following examples, which are intended to be illustrative only, as many modifications and variations therein will be apparent to those skilled in the art. Various embodiments of the present disclosure are now described in detail. With reference to the accompanying drawings, throughout the views, the same numerals (if any) indicate the same components. Unless the context clearly states otherwise, the meanings of "one", "an" and "said" used in the description herein and throughout the subsequent claims include the plural. In addition, as used in the description and claims herein, unless the context clearly states otherwise, the meaning of "in" includes "in" and "on". Moreover, titles or subtitles may be used in the specification for the convenience of the reader, which does not affect the scope of the present disclosure. In addition, some of the terms used in this specification are more specifically defined below.

[0053] The terms used in this specification generally have their ordinary meanings in the art, in the context of this disclosure, and in the specific context in which each term is used. Certain terms used to describe the present disclosure are discussed below or elsewhere in the specification to provide practitioners with additional guidance on the description of the present disclosure. For convenience, certain terms may be highlighted, such as using italics and / or quotation marks. The use of highlighting has no effect on the scope and meaning of the term; in the same context, the scope and meaning of a term are the same whether it is highlighted or not. It is understood that the same thing can be expressed in more than one way. Therefore, alternative language and synonyms can be used for any one or more terms discussed herein, and there is no special meaning for whether the term is elaborated or discussed in detail here. The present disclosure provides synonyms for certain terms. The use of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification, including examples of any term discussed herein, is illustrative only and in no way limits the scope and meaning of the present disclosure or any exemplary term. Similarly, the present disclosure is not limited to the various embodiments given in this specification.

[0054] Unless otherwise defined, all technical and scientific terms used in this disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. In the event of a conflict, the present document, including definitions, will control.

[0055] As used herein, the terms "includes," "comprising," "carrying," "having," "containing," "involving," etc. should be understood as open-ended, meaning including but not limited to.

[0056] As used herein, the phrase at least one of A, B, and C should be construed to mean a logical (A or B or C), using a non-exclusive logical OR. It should be understood that one or more steps within a method can be executed in different order (or concurrently) without altering the principles of the present disclosure.

[0057] As used herein, the term "module" or "unit" may indicate a component that belongs to or includes an application specific integrated circuit (ASIC), an electronic circuit, a combinational logic circuit, a field programmable gate array (FPGA), a processor (shared, dedicated, or group) that executes code, other suitable hardware components that provide the described functionality, or a combination of some or all of the above, such as in a system on a chip. The term module or unit may include a memory (shared, dedicated, or group) that stores code executed by a processor.

[0058] The term "code" used herein may include software, firmware and / or microcode, and may refer to a program, a routine, a function, a class and / or an object. The term shared used above means that a single (shared) processor may be used to execute some or all of the code from multiple modules. In addition, some or all of the code from multiple modules may be stored in a single (shared) memory. The term group used above means that a group of processors may be used to execute some or all of the code from a single module. In addition, a group of memories may be used to store some or all of the code from a single module.

[0059] As described herein, the term "interface" generally refers to a communication tool or device used to perform data communication between components at an interaction point between components. Generally speaking, an interface can be applicable at both the hardware and software levels and can be a unidirectional or bidirectional interface. Examples of physical hardware interfaces can include electrical connectors, buses, ports, cables, terminals, and other input / output (I / O) devices or components. The components that communicate with the interface can be, for example, multiple components or peripheral devices of a computer system.

[0060] The present disclosure relates to computer systems. As shown in the accompanying drawings, computer components may include physical hardware components, which are shown as solid line blocks, and virtual software components, which are shown as dashed line blocks. It will be understood by those skilled in the art that, unless otherwise specified, these computer components may be implemented in the form of, but not limited to, software, firmware, or hardware components, or a combination thereof.

[0061] The devices, systems, and methods described herein may be implemented by one or more computer programs executed by one or more processors. The computer programs include processor-executable instructions stored on a non-transitory tangible computer-readable medium. The computer programs may also include stored data. Non-limiting examples of non-transitory tangible computer-readable media are non-volatile memory, magnetic storage, and optical storage.

[0062] The present disclosure will now be described more fully hereinafter with reference to the accompanying drawings, in which embodiments of the present disclosure are shown. However, the present disclosure may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; on the contrary, these embodiments are provided to make the present disclosure thorough and complete and to fully convey the scope of the present disclosure to those skilled in the art.

[0063] As mentioned above, the generation of product descriptions is important for customer shopping experience, but it is challenging to manually generate a large number of product descriptions. To solve this problem, in some embodiments, the present disclosure provides a method for automatically generating product descriptions. The product description is usually a simple paragraph describing the appearance, function, potential audience, scene, etc. of the product. The paragraph may be based on the industry attributes of the product, such as the capacity of the refrigerator (118L), door design (asymmetric up and down), color (black or white), compressor (frequency conversion). In some embodiments, the automatic generation of product descriptions includes two stages, content planning and sentence planning. Content planning is to design theme extensions from product attributes. For example, "black" or "white" can be extended to "beautiful", "asymmetric up and down" can be extended to "fashionable", "frequency conversion" can be extended to "energy saving", and "118L" can be extended to "suitable for small families". Sentence planning is to make sentences for each theme. The generation of product descriptions needs to consider content planning and sentence planning. Content planning determines the theme relationship of each sentence and ensures the coherence of the entire paragraph, while sentence planning determines the structure and vocabulary content of each sentence.

[0064] In certain aspects, the present disclosure provides a system and method for automatically generating a coherent content plan for product descriptions given product attributes and a dataset of product descriptions manually written by authors. Figure 1 A system for generating content planning for product descriptions according to certain embodiments of the present disclosure is schematically depicted. Figure 1 As shown, system 100 includes computing device 110. In some embodiments, computing device 110 can be a server computer, a cluster, a cloud computer, a general-purpose computer, a headless computer, or a special-purpose computer that generates content planning. Computing device 110 may include, but is not limited to, a processor 112, a memory 114, and a storage device 116. In some embodiments, computing device 110 may include other hardware components and software components (not shown) to perform their corresponding tasks. Examples of these hardware and software components may include, but are not limited to, other required memories, interfaces, buses, input / output (I / O) modules or devices, network interfaces, and peripherals.

[0065] Processor 112 may be a central processing unit (CPU) configured to control the operation of computing device 110. In some embodiments, processor 112 may execute an operating system (OS) or other application of computing device 110. In some embodiments, computing device 110 may have more than one CPU as a processor, such as two CPUs, four CPUs, eight CPUs, or any suitable number of CPUs. Memory 114 may be a volatile memory, such as a random access memory (RAM), for storing data and information during operation of computing device 110. In some embodiments, memory 114 may be a volatile memory array. In some embodiments, computing device 110 may run on more than one processor 112 and / or more than one memory 114. Storage device 116 is a non-volatile data storage medium or device. Examples of storage device 116 may include flash memory, memory card, USB drive, solid state drive, or other types of non-volatile storage devices, such as hard disk drives, floppy disks, optical drives, or any other types of data storage devices. In some embodiments, computing device 110 may have more than one storage device 116. In certain embodiments, computing device 110 may also include a remote storage device 116 .

[0066] The storage device 116 stores computer executable code. The computer executable code includes a content planning application 118, a tag list database 180, and optionally a product attribute database 182 and a product description database 184. The content planning application 118 includes code or instructions that generate a content plan when executed at the processor 112. In some embodiments, the content planning application 118 may not be an executable code, but may be in the form of circuits corresponding to the functions of the executable code. By providing circuits instead of executable code, the operating speed of the content planning application 118 is greatly improved. In some embodiments, such as Figure 1 As shown, the content planning application 118 includes a data preparation module 120, a knowledge learning module 130, a candidate tag sequence generation module 140, a knowledge filter 150, a function module 160, a user interface 170, and the like.

[0067] The data preparation module 120 is configured to prepare product samples using the tag list database 180, the product database 182, and the product description database 184, and send the product samples to the knowledge preparation module 120 and the candidate tag sequence generation module 140. In some embodiments, after defining the tag list 180 based on the product database 182 and the product description database 184, and labeling the product descriptions in the product description database 184 using the tag list in the tag list database 180, the data preparation module 120 may only need the labeled descriptions from the product description database 184 for data preparation.

[0068] In some embodiments, the tag list database 180 includes multiple tag lists corresponding to different product categories. Different categories of products include small product categories such as air conditioners, televisions, washing machines, refrigerators, or large product categories such as clothing, home appliances, mobile phones and accessories, consumer electronics, computers, and offices. Each category has a corresponding tag list. In some embodiments, the tags in the tag list are predefined, and based on the product category, the number of tags in each tag list can be about hundreds to thousands. Tags include introductory tags and descriptive tags. Introductory tags correspond to industry attributes of product categories. For categories of clothing or apparel, introductory tags may include different colors, such as black, white, red, yellow, green, silver, mixed colors, different necklines and collars, such as high collars, boat collars, stand-up collars, V-necks, and different fabrics, such as cotton, silk, leather. Descriptive tags correspond to features that are described only in the product description, rather than the industry attributes of the product. For example, the tags "casual" or "cotton" may appear together in the product database and product description, but they are defined as introductory tags because they are industry attributes that appear in the product database 182 and are regarded as basic information or basic attributes of the product. For the category of clothing or apparel, descriptive labels can include different functions such as warmth, skin-friendly, breathable, stretchy, durable, wrinkle-resistant, and waterproof, different styles such as retro, casual, classic, and bohemian, different feelings such as comfort, lightness, sexy, luxury, and elegance, different audiences such as women, men, neutrals, and couples, and different scenes such as business, sports, office, commuting, outdoor, and travel. For the category of refrigerators, the label list can include introductory labels such as capacity, door design, color, and compressor, as well as descriptive labels such as simplicity, beauty, charm, fashion, and freshness.

[0069] In some embodiments, the product database 182 includes products provided by the e-commerce server and basic information about the products. Please note that the product database 182 provides a basis for defining introductory tags, however, when a tag list of a product category has been defined in the tag list database 180, it is not necessary to directly provide the product industry attributes from the product database 182 to the data preparation module 120.

[0070] In some embodiments, the product description database 184 includes a product description for each product in the product database 182. Each product description is a tagged text, and the tagging of the product description is performed based on a defined tag list from the tag list database 180. Figure 3 Schematically illustrates a product description of a label according to some embodiments of the present disclosure. Figure 3As shown, the product description is a paragraph of text. The text paragraph includes several sentences such as sentence 1, sentence 2, sentence 3, etc. Each sentence is divided into several clauses. Each clause can be marked with a label, which is one selected from the label list of the corresponding product category, where the label list is defined in the label list database 180. Please note that some clauses, such as clauses 22 and 33, may not contain much substantive information about the product, or may not correspond to any label in the label list, so there is no label. For the convenience of data processing, clauses 22 and 33 can also be marked with a pseudo-label, such as label 0. Each of labels 1 to 6 can be an introductory label that also exists in the product database 182 or a descriptive label that only exists in the product description. In some embodiments, the product database 182 and the product description database 184 are stored in a computing device other than the computing device 110, as long as they can be accessed by the computing device 110.

[0071] The knowledge learning module 130 is configured to learn knowledge from the marked product descriptions after receiving the marked product descriptions of the product categories prepared by the data preparation module 120. Figure 2 The knowledge learning module 130 includes a tag exclusivity module 132, a tag similarity module 134, a tag distance module 136, and a tag causality module 138. The modules 132 to 138 are configured to learn tag exclusivity, tag similarity, tag distance, and tag causality between tags in a tag list of a product category.

[0072] The tag exclusivity module 132 is configured to, after receiving a product description marked for a product category, learn the tag exclusivity between tags in the tag list of the category, and provide the learned tag exclusivity to the knowledge filter 150. This knowledge can be obtained from the type of tag or label. As described above, tags are divided into two types: introductory tags and descriptive tags. Introductory tags can be industry attributes of products, and these introductory tags should be exactly the same as the products. Introductory tags can be used to infer another tag, but cannot be inferred by other tags. All introductory tags are mutually exclusive. Descriptive tags can represent the feeling, function and scenario of the product. Descriptive tags can be inferred from introductory tags. Descriptive tags are inclusive. In other words, if two tags are introductory tags, the two tags have an exclusive relationship; if at least one of the two tags is a descriptive tag, the two tags will have an inclusive relationship. Figure 4 Schematically shows Figure 3 The learned relationships between tags in product descriptions. Figure 4 As shown, both label 1 and label 5 are introductory labels, and thus these two labels are mutually exclusive.

[0073] The tag similarity module 134 is configured to, after receiving a product description marked for a product category, learn the tag similarity between tags in the tag list of the category, and provide the learned tag similarity to the knowledge filter 150. The tag similarity module 134 is configured to acquire similarity knowledge based on the co-occurrence rate of tags in the data set. In some embodiments, because the product description needs to be presented very accurately, the co-occurrence association between tags is measured by Fisher's Exact Test. For example, "cotton" and "skin-friendly" co-occur, and they are considered similar tags. This similarity measurement using co-occurrence is better than the similarity measurement using word embeddings and is more suitable for the scenario of the present disclosure. Word embeddings measure similarity based on context. However, words with similar contexts may have different meanings. For example, "spring" and "winter" may have similar contexts and are considered similar to each other if word embeddings are used, but "spring" and "winter" are unlikely to co-occur in product descriptions and are not considered similar to each other if co-occurrence is used as the criterion. Figure 4 As shown, in some embodiments, the similarity between label 2 and label 3 is 0.4, which is calculated by Fisher's exact test, and the similarity value is the p-value of Fisher's exact test.

[0074] The tag distance module 136 is configured to, upon receiving a tagged product description for a product category, learn the tag distances between tags in the tag list of the category and provide the learned tag distances to the knowledge filter 150. The distance knowledge takes into account the pairwise tag distances in the directional tag sequence. Figure 4 As shown, the distance between label 1 and label 2 is 1, because clause 11 and clause 12 are adjacent to each other, the distance between label 5 and label 6 is 1, because clause 31 and clause 32 are adjacent to each other, and the distance between label 3 and label 4 is 2, because there is a clause 22 between clause 21 and clause 23. In some embodiments, the present disclosure is configured to calculate the arithmetic mean of the label distances between a pair of labels in all product descriptions of a product category. For example, for a total of 100 product descriptions in a product category, label 1 and label 2 appear together in 5 of the product descriptions, and the distances are 1, 1, 3, 2, and 1, respectively. Then the distance between label 1 and label 2 will be 1.6, (1+1+3+2+1) / 5=1.6. The distance captures the local semantic relationship between two labels. The smaller the distance between two labels, the closer the relationship between the two labels.

[0075] The tag causality module 138 is configured to, after receiving a tagged product description for a product category, learn the tag causality between tags in the category tag list, and provide the learned tag causality to the knowledge filter 150. The present disclosure regards causality as the most reasonable and strict tag order in a paragraph. In some embodiments, causality is mined through syntactic and semantic analysis in product descriptions. In some embodiments, causal conjunctions and verbs are used as confidence signals for mining causal relations, for example, "leads to", "originates from", "because", and "due to" are causal verbs and conjunctions. Figure 4 As shown, there is a causal relationship between tags 3 and 4, wherein any of clauses 21, 22, and 23 may contain causal verbs and conjunctions. In some embodiments, a causal relationship can also be determined when any other terms such as "due to", "therefore", "as long as", "by virtue of", "because", "consider", "for", "in view of", etc. exist between two tags in the product description, where the two tags appear together. In some embodiments, as long as two tags have a causal relationship in one of the product descriptions in which the two tags appear together, the two tags are defined to have a causal relationship. In some embodiments, a causal relationship is determined when causality exists in two product descriptions in which the two tags appear together.

[0076] The candidate tag sequence generation module 140 is configured to, after receiving the marked product description, generate a tag sequence from each product description, combine those tag sequences from different product descriptions to obtain a candidate tag sequence, and provide the candidate tag sequence to the knowledge filter 150. In some embodiments, the candidate tag sequence generation module 140 is configured to use an N-gram model to generate a candidate tag sequence, where N is the maximum length of the tag sequence. Because product descriptions usually have a word limit, such as 100 words, a tag sequence that is too long may produce a product description that is too long, resulting in a violation of the limit. Therefore, in some embodiments, N can be set to a positive integer from 3 to 10. In some embodiments, N is set to 5, and the candidate tag sequence generation module 140 is used to generate a tag sequence where N is 2 to 5. Figure 5 As shown, the product description includes five tags, namely "8kg", "white", "roller", "fixed frequency", and "home". The 10 tag sequences generated include 4 2-gram tag sequences, 3 3-gram tag sequences, 2 4-gram sequences, and 1 5-gram sequence. The tag sequences of all product descriptions in a product category are candidate tag sequences for the product category (e.g., refrigerator category or clothing category). In some embodiments, the candidate tag sequence generation module 140 is also used to remove redundant tag sequences from the combined tag sequence to obtain a candidate tag sequence.

[0077] The candidate tag sequence generation module 140 may be configured to implement the generation of candidate tag sequences in different ways. Figure 6 The method of generating a candidate tag sequence of a product category according to some embodiments of the present disclosure is schematically shown. Figure 6 As shown, multiple product descriptions of a product category are named as a data set. Each product description is divided into clauses, and the clauses are marked with topics or tags. The tag list of the product category from the tag list database 180 may include hundreds to thousands of tags depending on the product category. For each specific tag in the tag list, a corresponding 2-gram tag sequence, 3-gram tag sequence, 4-gram tag sequence and 5-gram tag sequence are generated. For example, for a 2-gram tag sequence of a specific tag or topic, the candidate tag sequence generation module 140 is configured to scan product descriptions with a specific topic. If the specific topic is not the last one in the product description, there will be a 2-gram tag sequence in the product description, including the specific topic and the topic next to it. If the 2-gram tag sequence is not in the candidate tag sequence, the candidate tag sequence generation module 140 will add the 2-gram tag sequence to the candidate tag sequence. Afterwards, 3-gram, 4-gram and 5-gram tag sequences can also be determined in a similar manner for a specific topic. The candidate tag sequence generation module 140 is configured to perform the same 2-gram to 5-gram scan on each topic in the tag list, and finally obtain the final candidate tag sequence. In a different manner of generating a candidate tag sequence, the candidate tag sequence generating module 140 may also be configured to directly use the tag list of the product category extracted from the tag list database 180 to generate N-grams.

[0078] The knowledge filter 150 is configured to, after receiving the learned knowledge from the knowledge learning module 130 and the candidate tag sequence from the candidate tag sequence generation module 140, filter the candidate tag sequence using the learned knowledge to obtain a filtered tag sequence, and provide the filtered tag sequence to the functional module 160. In some embodiments, the knowledge filter 150 is configured to filter the candidate tag sequence using tag exclusivity, and then further filter the candidate tag sequence using tag similarity, tag distance, and tag causality. The filtered tag sequence is essentially coherent. In some embodiments, the algorithm first uses exclusive knowledge for filtering, and then uses the other three types of knowledge respectively. Finally, the minimum intersection is obtained as the final result. Figure 7A to Figure 7C The implementation of filtering the candidate tag sequence according to some embodiments of the present disclosure is schematically shown, wherein the smiley face portion corresponds to the filtered tag sequence. Fig. 7AAs shown, for all candidate tag sequences 700 generated using N-gram, candidate sequences with two exclusive tags are removed to obtain inclusive tag sequences 702, wherein each inclusive tag sequence has at most one introductory tag.

[0079] The similarity tag sequence (circle) 704 includes candidate tag sequences in the inclusive tag sequence 702 whose similarity values ​​are greater than the similarity threshold. In some embodiments, for a tag sequence with two tags, the similarity value is the similarity learned by the tag similarity module 134 for the two tags, for example, by Fisher's exact test. For a tag sequence with more than two tags, the similarity value of each adjacent two tags should be greater than the similarity threshold. For example, if a candidate tag sequence has five consecutive tags ABCDE, then the similarity value of each of AB, BC, CD, and DE should be greater than the similarity threshold. In some embodiments, the similarity threshold is set to approximately 0.2 to 0.8. In some embodiments, the similarity threshold is set to approximately 0.5. The distance tag sequence (circle) 706 includes candidate sequences in the inclusive tag sequence 702 whose distance values ​​are equal to or less than the distance threshold. In some embodiments, for a tag sequence with two tags, the distance value is the distance learned by the tag distance module 136 for the two tags. If a tag sequence has more than two tags, the distance value of each adjacent two tags should be equal to or less than the distance threshold. For example, if a candidate tag sequence has five consecutive tags ABCDE, then the distance value of each of AB, BC, CD and DE should be equal to or less than the distance threshold. In some embodiments, the distance threshold is set to approximately 1 to 8. In some embodiments, the threshold is set to 4. The causal tag sequence (circle) 708 includes candidate sequences in the inclusive tag sequence 702 that have a causal relationship. In some embodiments, if any one of two consecutive tags in the tag sequence has a causal relationship, it is determined that the tag sequence has a causal relationship. The similarity threshold and the distance threshold can depend largely on the category of the product, the N value of the N-gram, the number of data sets in the category, and the number of candidate tag sequences.

[0080] like Fig. 7A As shown, the tag sequence that exists in the similarity tag sequence 704, the distance tag sequence 706, and the causality tag sequence 708 is selected as the filtered tag sequence. If a sufficient number of filtered tag sequences are obtained in this way, the obtained filtered tag sequence is the final content planning. The sufficient number can be, for example, greater than 50 to 100. If the number of filtered tag sequences obtained is insufficient or no filtered tag sequence is obtained, the present disclosure can further adopt Figure 7B Standards shown.

[0081] like Figure 7BAs shown, when there is no overlap between the similarity tag sequence 704, the distance tag sequence 706, and the causality tag sequence 708 (or the number of tag sequences overlapping the three circles 704, 706, and 708 is insufficient), the tag sequences appearing in any two of the three circles 704, 706, and 708 will be selected. In addition, if the circles 704, 706, and 708 do not overlap, as shown in FIG. Figure 7C As shown, the knowledge filter 150 may be configured to select the smallest circle, such as the causal label sequence 708 , which has a smaller number of candidate label sequences than the similarity label sequence 704 and the distance label sequence 706 .

[0082] Figure 8 Schematically illustrates a filtering algorithm according to some embodiments of the present disclosure. Figure 8 As shown, when generating candidate label sequences, the candidate label sequences are first exclusively filtered to remove label sequences with two exclusive labels. Then, a set of overlapping label sequences is determined, wherein the SDC set includes candidate label sequences that appear in the similarity label sequence 706, the distance label sequence 708, and the causal label sequence 710, the SD set includes the intersection between the similarity label 706 and the distance label sequence 708, the SC set includes the intersection between the similarity label sequence 706 and the causal label sequence 710, and the DC set includes the intersection between the distance label sequence 708 and the causal label sequence 710. In some embodiments, when the SDC set is not empty, the final label sequence is SDC; if SD, SC, and DC are all empty, the final label sequence is the one with the least number of label sequences among S, D, and C; when at least one of SD, SC, and DC is not empty, the final label sequence is the one with the least number of label sequences among S, D, C, SD, SC, and DC. In some embodiments, the knowledge filter 150 can also be used Figure 8 Other variations of the method shown are used to obtain the final tag sequence or the filtered tag sequence. In some embodiments, the knowledge filter 150 can also sort the filtered tag sequence based on their similarity, distance and causal relationship. The filtered tag sequence is also called a coherent tag sequence because the tags in each filtered tag sequence are coherent with each other. The filtered tag sequence may include, for example, ("enough movie sources", "teaching content"), ("home theater", "enjoyment"), ("excellent sound quality", "enjoyment"), ("simple", "fashionable and beautiful"), ("bright colors", "natural pictures"), ("easy to control", "suitable for all ages"), ("natural pictures", "eye protection / anti-fatigue"), ("home theater", "home"), etc. After obtaining the filtered candidate tag sequence, the knowledge filter 150 is also configured to provide the filtered candidate tag sequence to the functional module 160.

[0083] Functional module 160 may be stored in computing device 110 or any other computing device in communication with computing device 110. Functional module 160 is configured to perform certain functions using the filtered candidate tag sequence. In some embodiments, its function is to generate a new text description, using the filtered candidate tag sequence as a content plan, so that functional module 160 can use the content plan to generate sentences, and the sentences constitute a text description of the product. In some embodiments, in order to generate a new product description text, functional module 160 is configured to select multiple industry attributes of the product, obtain introductory tags or inferred tags using the industry attributes, select one or more tag sequences using the introductory tags or inferred tags in the coherent tag sequence corresponding to the product category, generate sentences from each selected coherent tag sequence, and arrange the sentences to form a new description of the product. In some embodiments, the selection of coherent tag sequences may use a standard different from or in addition to the introductory tag, such as the planning style of the text to be generated.

[0084] User interface 170 is configured to provide a user interface or graphical user interface in computing device 110. In certain embodiments, a user or administrator of the system can use user interface 170 to configure parameters of computing device 110, particularly parameters used in content planning application 118.

[0085] Fig. 9 The architecture of the content planning system according to some embodiments of the present disclosure is schematically depicted. Fig. 9 As shown, the product attribute database or product database includes product industry attributes, and the product description database includes labeled product descriptions. The labeled product descriptions can be used to generate candidate tag sequences using N-gram. In addition, four types of knowledge can be learned from the product's industry attributes and product descriptions. The learned knowledge can be used to filter candidate tag sequences to obtain coherent tag sequences. The coherent tag sequence of a product category is a good content plan and can be used to generate product descriptions for any product in the category.

[0086] Fig.10 Schematically depicts a method for content planning according to certain embodiments of the present disclosure. In certain embodiments, Fig.10 The method 1000 shown may be performed as follows. Figure 1 It is implemented on the computing device 110 shown. It should be noted that, unless otherwise specified in the present disclosure, the steps of the above method can be arranged in different orders, and are therefore not limited to Fig.10 in the order shown.

[0087] In step 1002, the data preparation module 120 provides the labeled product descriptions of the product category to the knowledge learning module 130 and the candidate label sequence generation module 140. Each product description contains multiple sentences, each sentence is divided into one or several clauses, and each clause is labeled (some clauses may have no labels or have a virtual label 0). The labels can be introductory labels representing product industry attributes and descriptive labels extracted from the product description instead of introductory labels.

[0088] In step 1004, for each tagged product description, the knowledge learning module 130 determines the tag exclusivity between any two tags based on the tag type. When both tags are introductory tags, the two tags are exclusive, otherwise the two tags are inclusive. In some embodiments, the knowledge learning module 130 can also determine the tag exclusivity between any two tags in the tag list based on the tag type.

[0089] In step 1006, for all tagged product descriptions, the knowledge learning module 130 determines the tag similarity between any two tags based on whether the two tags co-occur in any tagged product description. The co-occurrence rate of the two tags in the tagged product description is the similarity value between the two tags. In some embodiments, the similarity value between any two tags is determined using Fisher's exact test.

[0090] In step 1008, for all tagged product descriptions, the knowledge learning module 130 determines the tag distance between any two tags based on whether the two tags co-appear in any tagged product description, and the distance between the two tags is measured by the number of clauses between the two tags in the tagged product descriptions where the two tags co-appear. If the two tags are adjacent to each other, the distance between the two tags is 1, and if there is another clause between the two tags, the distance between the two tags is 2. When two tags co-appear in multiple product descriptions, the distance values ​​of the two tags are averaged. For example, if the two tags co-appear in 20 product descriptions out of a total of 100 product descriptions, the distance in 10 product descriptions is 1, the distance in 6 product descriptions is 2, and the distance in 4 product descriptions is 3, then the distance between the two tags is (10×1+6×2+4×3) / 20=1.7.

[0091] In step 1010, for all tagged product descriptions, the knowledge learning module 130 determines the tag causality between any two tags based on whether the two tags co-appear in the tagged product descriptions and whether there is a causal term between two adjacent tags in the tagged product descriptions where the two tags co-appear. If the two tags co-appear in at least one product description and there is a causal term such as "cause" or "because" or "due to" between the two tags in the at least one product description, it is determined that the two tags have a causal relationship. In some embodiments, when the causal relationship of two tags is displayed in at least two product descriptions, the two tags are determined to have causality.

[0092] In step 1012, when the labeled product descriptions of the product category are received from the data preparation module 120, the candidate label sequence generation module 140 generates a label sequence for each labeled product description using N-gram, combines the label sequences generated by all labeled product descriptions to form a candidate label sequence, and sends the candidate label sequence to the knowledge filter 150. For example, when N is 5, if the first product description has two labels, label 1 and label 2, the number of generated label sequences is one, i.e., label 1-label 2. If the second product description has three labels, label 3 to label 5, the number of generated label sequences will be three, i.e., label 3-label 4, label 4-label 5, label 3-label 4-label 5. If the third product description has six tags, tag 6 to tag 11, then the number of generated tag sequences is 14, namely tag 6-tag 7, tag 7-tag 8, tag 8-tag 9, tag 9-tag 10, tag 10-tag 11, tag 6-tag 7-tag 8, tag 7-tag 8-tag 9, tag 8-tag 9-tag 10, tag 9-tag 10-tag 11, tag 6-tag 7-tag 8-tag 9, tag 7-tag 8-tag 9-tag 10, tag 8-tag 9-tag 10-tag 11, tag 6-tag 7-tag 8-tag 9, tag 7-tag 8-tag 9-tag 10, tag 8-tag 9-tag 10-tag 11, tag 6-tag 7-tag 8-tag 9-tag 10 and tag 7-tag 8-tag 9-tag 10-tag 11. The tag sequences generated in different product descriptions are combined and redundant tag sequences are removed to obtain candidate tag sequences.

[0093] In step 1014, after receiving the knowledge from the knowledge learning module 130 and the candidate tag sequence from the candidate tag sequence generation module 140, the knowledge filter 150 uses the knowledge to filter the candidate tag sequence and sends the filtered tag sequence to the function module 160. The filtering is to first remove the candidate tag sequence with two exclusive tags, and then select the candidate tag sequence with high similarity, close distance, and causal relationship as much as possible. In some embodiments, the filtering is to use Fig. 7A , Figure 7B , Figure 7C and Figure 8 Or any of the variations of these methods. The obtained filtered tag sequence is also called content planning.

[0094] In step 1016, after receiving the content plan from the knowledge filter 150, the function module 160 can use the content plan to generate a new product description, for example, using the content plan and at least one of the following: the industry attribute of the product, the product description style preferred by the user, and the product description of the old version. In some embodiments, the function module 160 selects several tag sequences from the content plan based on important product industry attributes, arranges the content sequences based on the style selected by the user, and generates a sentence from each tag sequence to form a new product description.

[0095] In some aspects, the present disclosure relates to a non-transitory computer-readable medium for storing computer executable code. When executed at the processor 112 of the computing device 110, the code may perform the method 1000 described above. In some embodiments, the non-transitory computer-readable medium may include, but is not limited to, any physical or virtual storage medium. In some embodiments, the non-transitory computer-readable medium may be implemented as Figure 1 A storage device 116 of the computing device 110 is shown.

[0096] Some embodiments of the present disclosure are compared with a method based on a recurrent neural network (RNN). For example, the RNN method may be a TopicRNN. Fig.11 A comparison between certain embodiments of the present disclosure and RNN-based methods is shown. Fig.11 As shown, for the categories of clothing, air conditioner, television, washing machine, and refrigerator, the knowledge-based method according to the present disclosure is significantly better than the RNN-based method.

[0097] In summary, all relevant methods in this field have coherence problems, which are manifested as high false positives and high false negatives. In contrast, the present disclosure provides a system and method for automatically generating coherent content planning for product descriptions in e-commerce, reducing false positives and false negatives. The system combines a knowledge construction component, a candidate tag sequence generation algorithm, and a content planning filtering algorithm. Knowledge-driven content planning methods rely on mining common tag knowledge from data sets and product attributes: tag similarity, tag distance, tag causality, and tag exclusivity. Here, "tag" refers to the topic words representing the main idea of ​​each sentence, and each product attribute can be regarded as a tag. The present disclosure assumes that each sentence has one or more tags to represent its main idea (theme). Therefore, the relationship between tags is a microcosm of the relationship between sentences in the product description. This knowledge about tags represents various aspects in the product description, including not only the distribution of topics, local (syntactic) dependencies, and global (semantic) dependencies, but also the semantic role of natural language, and even the characteristics of causality and product attributes. Therefore, compared with existing methods, using this knowledge can improve the coherence of content planning. Specifically, the knowledge of label exclusivity, label similarity, and label distance can simultaneously capture the global / local semantic relations and word exclusivity in a paragraph, thereby reducing false positives in content planning; the knowledge of label distance and label causality can accurately locate low-frequency topic relations, thereby reducing false negatives in content planning.

[0098] The foregoing description of exemplary embodiments of the present disclosure has been presented for purposes of illustration and description only and is not intended to be exhaustive or to limit the disclosure to the precise form disclosed.Many modifications and variations are possible in light of the above teachings.

[0099] The embodiments are selected and described in order to explain the principles of the present disclosure and its practical application, so that other persons skilled in the art can utilize the present disclosure and various embodiments and various modifications suitable for the intended specific use. Alternative embodiments will become apparent to those skilled in the art to which the present disclosure belongs without departing from the spirit and scope of the present disclosure. Therefore, the scope of the present disclosure is limited by the appended claims rather than the foregoing description and the exemplary embodiments described therein.

Claims

1. A system for obtaining content planning, wherein: The system includes a computing device including a processor and a storage device for storing computer executable code, the computer executable code when executed at the processor is configured to: Providing a label for each clause of each sentence of each product description of each product category, wherein the label includes a plurality of introductory labels and a plurality of descriptive labels, the plurality of introductory labels corresponding to industry attributes of the product category, and the plurality of descriptive labels corresponding to a plurality of descriptions of the product category; Determining tag exclusivity between any two of the tags, wherein when the two tags are both the introductory tags, the two tags are defined as having the tag exclusivity; Determine the tag similarity between any two tags based on the co-occurrence rate of the two tags in the product description; determining a tag distance between any two tags in the product description based on the number of the clauses between the two tags, wherein the two tags co-occur; Based on a causal relationship between any two tags in the product description, determining tag causality for characterizing the causal relationship between the two tags, wherein the two tags appear together; generating a candidate tag sequence from each of the product descriptions using the tags in each of the product descriptions; and The candidate tag sequence is filtered using the tag exclusivity, the tag similarity, the tag distance, and the overlapping relationship between the tag causality to obtain the content plan.

2. The system according to claim 1, wherein: The product categories include refrigerators, air conditioners, televisions and washing machines.

3. The system according to claim 1, wherein: The product categories include apparel, consumer electronics, home appliances, computers and office, sports and outdoors, health and beauty, toys, and shoes.

4. The system according to claim 3, wherein: The product category is clothing, the industry attributes include size, color and material, and the description includes function, style, feel, potential audience and scenario.

5. The system according to claim 1, wherein: The computer executable code is configured to determine the tag similarity using Fisher's exact test.

6. The system according to claim 1, wherein: The computer executable code is configured to determine a tag distance between the two tags by averaging the tag distances between the two tags in the product description, wherein the two tags co-occur.

7. The system according to claim 1, wherein: The computer executable code is configured to determine that tag causality exists between the two tags when at least one of the terms "results from", "arise from", "because" and "due to" exists between the two tags in any of the product descriptions, wherein the two tags co-occur.

8. The system according to claim 1, wherein: The computer executable code is configured to generate the candidate tag sequence using N-gram.

9. The system according to claim 1, wherein: The computer executable code is configured to filter the candidate tag sequences in the following manner: removing candidate tag sequences containing two of the introductory tags; and At least a tag sequence whose tag distance is less than a predefined tag distance threshold, a tag sequence whose tag similarity is greater than a predefined tag similarity threshold, or a tag sequence having the tag causality is retained.

10. The system according to claim 1, wherein: The computer executable code is further configured to generate a new product description based on the content plan.

11. A method for obtaining content planning performed by a computing device, comprising: Providing, by a computing device, a label for each clause of each sentence of each product description of each product category, wherein the label comprises a plurality of introductory labels and a plurality of descriptive labels, the plurality of introductory labels corresponding to industry attributes of the product category, and the plurality of descriptive labels corresponding to a plurality of descriptions of the product category; Determining, by the computing device, tag exclusivity between any two of the tags, wherein when the two tags are both the introductory tags, the two tags are defined as having the tag exclusivity; Determining, by the computing device, a tag similarity between any two tags in the product description based on a common occurrence rate of the two tags; determining, by the computing device, a tag distance between any two tags in the product description based on the number of the clauses between the two tags, wherein the two tags appear together; Determining, by the computing device, tag causality for characterizing the causal relationship between any two tags in the product description based on the causal relationship between the two tags, wherein the two tags appear together; generating, by the computing device, a candidate tag sequence from each of the product descriptions using the tags in the each of the product descriptions; and The computing device filters the candidate tag sequence using the tag exclusivity, the tag similarity, the tag distance, and the overlapping relationship between the tag causality to obtain the content plan.

12. The method according to claim 11, wherein: The product categories include clothing, refrigerators, air conditioners, televisions and washing machines. When the product category is clothing, the industry attributes include size, color and material, and the description includes function, style, feeling, potential audience and scenario.

13. The method according to claim 11, in, Determining the label similarity using Fisher's exact test; wherein the tag distance between the two tags is determined by averaging the tag distance between the two tags in the product description, wherein the two tags appear together; and Wherein, when at least one of the terms "resulting in", "originated from", "because" and "due to" exists between the two tags in any of the product descriptions, it is determined that the tag causality exists between the two tags, wherein the two tags appear together.

14. The method according to claim 11, wherein: The candidate tag sequence is generated using N-gram.

15. The method according to claim 11, wherein filtering the candidate tag sequences comprises: removing candidate tag sequences containing two of the introductory tags; as well as At least a tag sequence whose tag distance is less than a predefined tag distance threshold, a tag sequence whose tag similarity is greater than a predefined tag similarity threshold, or a tag sequence having the tag causality is retained.

16. The method according to claim 11, further comprising: A new product description is generated based on the content plan.

17. A non-transitory computer readable medium for storing computer executable code, wherein: When the computer executable code executes the method for obtaining content planning at a processor of a computing device, the computer executable code is configured to: Providing a label for each clause of each sentence of each product description of each product category, wherein the label includes a plurality of introductory labels and a plurality of descriptive labels, the plurality of introductory labels corresponding to industry attributes of the product category, and the plurality of descriptive labels corresponding to a plurality of descriptions of the product category; Determining tag exclusivity between any two of the tags, wherein when the two tags are both the introductory tags, the two tags are defined as having the tag exclusivity; Determine the tag similarity between any two tags based on the co-occurrence rate of the two tags in the product description; determining a tag distance between any two tags in the product description based on the number of the clauses between the two tags, wherein the two tags co-occur; Based on a causal relationship between any two tags in the product description, determining tag causality for characterizing the causal relationship between the two tags, wherein the two tags appear together; generating a candidate tag sequence from each of the product descriptions using the tags in each of the product descriptions; and The candidate tag sequence is filtered using the tag exclusivity, the tag similarity, the tag distance, and the overlapping relationship between the tag causality to obtain the content plan.

18. The non-transitory computer readable medium of claim 17, wherein: The computer executable code is configured to: Determining the label similarity using Fisher's exact test; The tag distance between the two tags is determined by averaging the tag distance between the two tags in the product description, where the two tags appear together; and when at least one of the terms "resulting in", "originated from", "because" and "due to" exists between the two tags in any of the product descriptions, it is determined that tag causality exists between the two tags, where the two tags appear together.

Images