Innovative Requirement Analysis Method and System Based on Span Association Prediction

Abstract

The present invention proposes an innovative demand analysis method and system based on span correlation prediction, which relates to natural language processing technology. This method uses Bi-LSTM and Bi-GRU to encode the comment corpus, combines context-related boundary representations with an attention mechanism on spans to update the clause vectors; integrates the prediction results of sentiment and reason and relative position information through a multi-dimensional information interaction mechanism, screens and predicts sentiment-reason pairs, corresponds them to Maslow's hierarchy of needs, and extracts the essence of the needs; inputs the audience needs and customer business needs into the X model and outputs product innovation functions. The present invention enhances the information interaction between clauses, improves the extraction accuracy of sentiment-reason pairs by more accurately analyzing the relationship between sentiment and reason, matches them to Maslow's hierarchy of needs, and combines with the X model framework to provide a strategic orientation for product function innovation.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and particularly relates to an innovative demand analysis method and system based on span association prediction. Background Art

[0002] In the current digital age, with the explosive growth of social media and online interaction platforms, user feedback data has become a key resource for enterprises to understand market demands, innovate product functions, and enhance user experiences.

[0003] Sentiment Analysis is a research direction in the fields of Natural Language Processing (NLP), Text Mining, and Computational Linguistics. Sentiment Analysis is also known as opinion mining, sentiment orientation analysis, opinion extraction, sentiment mining, etc. It is a process of analyzing, processing, summarizing, and reasoning on subjective texts with emotional colors.

[0004] Traditional sentiment analysis techniques can only obtain the sentiment polarity (positive, neutral, negative) of texts. However, more in-depth information (such as the reasons causing emotions, the objects expressing emotions, etc.) is still worthy of further exploration. In contrast, pure sentiment analysis may only tell users that they are dissatisfied but does not explain why, resulting in ambiguity in demand recognition. The extraction of emotion - reason pairs is beneficial for product teams to directly focus on the core of the problem instead of blindly groping in extensive emotional feedback, thus accelerating the problem-solving process.

[0005] The emotion - reason pair extraction task is a fine-grained task in text sentiment analysis, aiming to extract all emotions and their root causes in a document. Currently, the emotion - reason pair extraction task is solved in a step-by-step manner, that is, first completing two subtasks of emotion extraction and reason extraction, and then completing the pairing task of emotion - reason pairs. However, this fails to handle the potential relationship between the two subtasks and the emotion - reason pair extraction task well. At the same time, the grammatical information contained in the document itself is ignored. To solve these problems, some emotion - reason pair extraction methods have been proposed in the prior art to extract emotion - reason pairs more accurately. Such as:

[0006] Chinese Patent Document CN116578671A discloses an emotion-cause pair extraction method and device. According to the word hidden state weights and the hidden states of each word included in the ECPE text for extracting emotion-cause pairs, the first clause feature and the first clause feature vector corresponding to the ECPE text are determined; according to the first clause feature, the graph attention network and the semantic dependency adjacency matrix, the clause adjacency matrix of the graph neural network is obtained; according to the first clause feature vector and the clause adjacency matrix of the graph neural network, the second clause feature and the second clause feature vector corresponding to the graph neural network are obtained; according to the emotion-cause pair feature, the graph attention network and the matching possibility matrix between clause pairs, the clause pair feature and the clause pair feature vector corresponding to the graph neural network are obtained; a classifier is trained to predict the clause pair feature vector corresponding to the graph neural network. This method mainly improves the accuracy of the emotion-cause pair prediction probability based on the loss function of emotion-cause extraction.

[0007] Chinese Patent Document CN114065769A discloses a training method for an emotion-cause pair extraction model. The method includes: inputting a document sample into a first encoding network to encode words and clauses, obtaining an emotion clause representation and a cause clause representation; predicting the two clause representations of each clause to obtain two clause prediction results, an emotion output and a cause output; inputting the emotion output and the cause output into a graph attention network for updating; based on a pairing network, obtaining corresponding emotion representations and cause representations according to the two updated outputs, and pairing the emotion representations and the cause representations to obtain emotion-cause pairs; according to a prediction network, obtaining emotion-cause pair prediction results; calculating a loss value according to the prediction results and updating the model. This method uses a graph attention network to extract the mutual relationship between each clause, enriches the information contained in the emotion output and the cause output of each clause, and improves the accuracy.

[0008] Chinese Patent Document CN116910255A provides an emotion-cause pair extraction method. The method includes: calculating the correlation degree, information fusion, and the contribution degree of the word vector to the clause vector between each clause vector and the corresponding word vector to obtain the target clause vector corresponding to each clause; establishing an adjacency matrix of the mutual relationship between clauses according to the target clause vector corresponding to each clause, and extracting the semantic information of each clause from the adjacency matrix; calculating the prediction labels of all candidate emotion-cause pairs according to the semantic information of each clause, and screening out all emotion-cause pairs in the text to be extracted according to the prediction labels. This method mainly enhances the ability to capture clause semantics and improves the pairing accuracy of emotion clauses and cause clauses by capturing the interaction between clauses and words and the contribution degree of words to the clause vector representation.

[0009] Although the above-mentioned methods for extracting emotion-cause pairs can all improve the accuracy of the extracted emotion-cause pairs to a certain extent, the semantic representation of clauses is relatively limited, the information interaction is not deep enough, and the multi-dimensional interaction between emotion cause prediction and context is not fully considered, so the accuracy still needs to be improved. At the same time, the above methods do not further analyze in combination with the deep needs of customers and audiences, and cannot provide a good strategic guidance for product function innovation. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide an innovative demand analysis method and system based on span correlation prediction. Aiming at the problems existing in the prior art, the method provided by the present invention introduces span correlation prediction, more accurately understands the meaning and relationship between emotion and cause through the information of the context, enhances the information interaction between clauses, precisely pairs emotion and cause, and matches these pairs to Maslow's hierarchy of needs, combined with the X model framework, to provide a strategic guidance for product function innovation, aiming to deeply explore and meet the deep needs of different users, and drive the continuous improvement of product experience and market competitiveness.

[0011] To solve the above technical problems, the present invention adopts the following technical solutions:

[0012] In the first aspect, the present invention provides an innovative demand analysis method based on span correlation prediction, including the following steps:

[0013] S1. Comment preprocessing: Obtain product comments and train to obtain corresponding word vectors, and use Bi-LSTM and Bi-GRU to perform two-level encoding on the clauses in the comment corpus, and list the obtained clause vectors one by one;

[0014] S2. Span representation: Combine the boundary representation related to the context and the attention mechanism on the span, and apply it to the learning of emotion and cause to obtain the updated representation vector of the clause;

[0015] S3. Span correlation pairing; Using the clause representation vectors obtained in step S1, with each clause as the pivot, pair the clauses with different spans in the comment corpus one by one, and integrate the span information into the updated representation vector of the clause in step S2;

[0016] S4. Emotion-cause pair prediction: Integrate the prediction results of emotion and cause and the relative position information, and then use the multi-dimensional information interaction mechanism to screen and predict the emotion-cause pairs;

[0017] S5. Demand hierarchy matching: Corresponding the predicted emotion-cause pairs to Maslow's hierarchy of needs to extract the essence of the needs;

[0018] S6. Product Innovation: Input the audience needs based on Maslow's hierarchy of needs classification and the customer business needs into the X model, and output the product innovation functions.

[0019] Further, step S1 specifically includes:

[0020] S101. Pre-training Model: Obtain product reviews on a large-scale Chinese microblog corpus and use the Word2vec algorithm for training to obtain corresponding word vectors.

[0021] S102. Two-level Encoding: Use Bi-LSTM and Bi-GRU to perform two-level encoding on the clauses in the review corpus, list the obtained clause vectors one by one, take each clause as a pivot, pay attention to its context information from multiple aspects, and enhance the subordinate clauses with targeted information to strengthen the causal relationship between clauses.

[0022] Further, the specific method of two-level encoding in step S102 is as follows:

[0023] (1) Word-level Encoding: For each word in a clause, use the word vector obtained from the Word2vec model as the initial input feature w n ;

[0024] (2) Clause-level Encoding: After completing the word-level encoding, use w n as the input of Bi-LSTM and Bi-GRU, model the context relationship between words, and obtain the clause representation vector S n .

[0025] Further, step S2 specifically includes:

[0026] Step S201. Given the review corpus D = {S1, S2, S3,..., S n}}, perform the following operations to update the clause information through the clause representation vector S n :

[0027]

[0028] Among them, represents a 3-layer feedforward neural network, which performs secondary processing on the clause vector S n to obtain a new feature vector a n , represents the attention weight of a n within the span i of each clause, softmax represents the normalization process, span start (i) and span end(i) represent the start position and end position of span i respectively, exp() represents the exponential operation, ∑ represents the summation operation, and ⊙ represents the product of the corresponding elements of two matrices. represents the clause representation vector after information interaction;

[0029] Step S202: Within the span w, the clause representation vector S output by the Bi-GRU n is fused with the clause representation vector after information interaction obtained in step S201 to finally obtain the clause updated representation vector X n :

[0030]

[0031] where add[] represents vector concatenation, concatenating the clause representation vector S n and the clause representation vector after information interaction to obtain the concatenated clause updated representation vector X n .

[0032] Furthermore, step S3 specifically includes:

[0033] Step 301: Take each clause in the comment corpus as a pivot, pair it with the surrounding clauses within a certain preset range, and fuse information such as relative positions to obtain the vector representation of the final candidate pair:

[0034]

[0035] where W e and W c are the sentiment weight matrix and the reason weight matrix respectively, b e and b c are the sentiment bias term and the reason bias term respectively, is the i-th sentiment clause vector, is the i-th reason clause vector, E i is the sentiment prediction of the i-th clause, and C i is the reason prediction of the i-th clause;

[0036] Here, according to general language habits, the reasons corresponding to a certain sentiment are generally within a certain range before and after the sentiment clause, and even appear in the same clause as the sentiment. Therefore, in this step, each clause in the comment corpus is taken as a pivot, paired with the surrounding clauses within a certain range, and information such as relative positions is fused to obtain the vector representation of the final candidate pair.

[0037] Then, through the clause updated representation vector X obtained by the span representation in step S202 n, pair the clauses within the span w to obtain sentiment - reason candidate pairs, denoted as:

[0038] Pair i,j = [X i , X j , i ∈ [1, n], j ∈ [i - w, i + w + 1]

[0039] Among them, Pair i,j represents the candidate sentiment - reason pair composed of the sentiment clause at position i and the reason clause at position j, and w represents the span;

[0040] Step 302: To fuse the obtained information and provide multi - aspect information support for the prediction of the sentiment - reason pair Pair i,j . Fuse the sentiment and reason predictions with the candidate sentiment - reason pair representation, and at the same time add the relative position information loc i,j to improve the prediction accuracy. The specific formula is as follows:

[0041] P i,j = [Pair i,j , E i , C j , loc i,j

[0042] Among them, i represents the index of the clause at the sentiment position in the sequence, representing the specific position where the sentiment occurs, that is, the position of the sentiment clause; j represents the index of the clause at the reason position in the sequence, corresponding to the sentiment i position, that is, the position of the reason clause; Pair i,j represents the candidate sentiment - reason pair, E i is the sentiment prediction of the i - th clause, and C j is the reason prediction of the j - th clause;

[0043] Furthermore, the specific process of step S4 is as follows:

[0044] Obtain the representation vectors of all candidate sentiment - reason pairs through span - associated pairing Then use a feed - forward neural network combined with softmax for prediction:

[0045]

[0046] Among them, represents the prediction result of the sentiment - reason pair, FFN pair represents a 3 - layer feed - forward neural network, W pair is the candidate sentiment - reason pair weight matrix, and b pair is the candidate sentiment - reason pair bias term;

[0047] ​During the training process, multi-task learning is used to jointly train the model, and the cross-entropy loss function is used to calculate each task, namely the sentiment extraction task, the reason extraction task, and the sentiment-reason pair extraction task:

[0048]

[0049] and are the true prediction results of sentiment, reason, and sentiment-reason pair respectively, and L emo , L cau and L pair are the loss functions of sentiment, reason, and sentiment-reason pair respectively;

[0050] The loss function of the final model is expressed as the sum of the three:

[0051] L = L pair + L emo + L cau + λ||θ|| 2

[0052] where: λ represents the weight of the L2 regularization term, and θ is the L2 regularization term of all parameters in the model.

[0053] The L2 regularization term, also known as weight decay, is a commonly used technique for reducing the generalization error of the model.

[0054] Furthermore, the Maslow's hierarchy of needs in step S5 is presented in the form of a pyramid, which successively includes, from high to low levels: physiological needs, safety needs, social needs, esteem needs, and self-actualization needs;

[0055] List the relevant keywords of Maslow's hierarchy of needs, such as:

[0056] Physiological needs include food, water, sleep, health, rest, comfort, etc.;

[0057] Safety needs include safety, protection, stability, work, income, health insurance, etc.;

[0058] Social needs include friendship, love, family, relationship, belonging, acceptance, social interaction, team, etc.;

[0059] Esteem needs include self-esteem, respect, recognition, status, achievement, prestige, praise, etc.;

[0060] Self-actualization needs include growth, potential, creativity, self-actualization, goals, dreams, personal development, etc.

[0061] Use a thesaurus (such as WordNet) to expand each keyword set for the above keywords to obtain a keyword library, so that the obtained keyword library covers more variants and synonyms. Then, match the sentiment-reason pairs output in step S4 with the obtained keyword library to determine the Maslow's hierarchy of needs corresponding to the sentiment-reason pairs and extract the essence of the needs.

[0062] Furthermore, the analysis process in step S6 in combination with the X model is specifically as follows:

[0063] The upper two ends of the X model are respectively represented as customer business needs and audience needs, the intersection point is represented as the heterogeneous user goal, the lower left end is represented as product innovation functions, and the right end is represented as Maslow's needs (human nature). The two major analysis steps of the X model are as follows:

[0064] (1) Bridge the customer business requirement items and the audience requirement items extracted based on Maslow's hierarchy of needs to obtain the heterogeneous user goal, and then combine the analysis of human nature needs. This step aims to deeply explore the common vision and value pursuit behind this heterogeneous user group and touch the most fundamental level of needs (human nature needs). For those trivial or pseudo-needs that are superficial and contribute little to the core value of the product, they are marked with a dotted line, indicating that although these needs exist, they do not need to be delved into too deeply in the current analysis stage.

[0065] (2) Analyze from Maslow's human nature needs to the heterogeneous user goal analysis and then to the product innovation function analysis, or directly from the heterogeneous user goal analysis to the product innovation function analysis. This step ensures that the product design not only meets the fundamental demands of human nature but also effectively solves the actual problems faced by heterogeneous users. This path reflects the process from the macroscopic concept to the specific implementation, ensuring that every design of the product function responds to the deep human nature needs and achieving the precise matching of the heterogeneous user goal and the product function.

[0066] The X model originates from the user-centered design concept and is a transformation process from problems to solutions, emphasizing a systematic methodology for deeply understanding user needs and transforming them into specific product functions.

[0067] On the second aspect, the present invention also provides an innovative requirement analysis method using the above-mentioned span correlation prediction, which specifically includes:

[0068] A comment preprocessing module, including a pre-training sub-model and a two-level encoding sub-module, where

[0069] The pre-training sub-model is used to obtain product comments and perform training to obtain corresponding word vectors;

[0070] The two-level encoding sub-module is used to perform two-level encoding on the clauses in the comment corpus using Bi-LSTM and Bi-GRU, and list the obtained clause representation vectors one by one;

[0071] Span representation module: used to combine context-related boundary representations with attention mechanisms on spans, apply them to the learning of sentiment and reasons, and obtain clause update representation vectors;

[0072] Span association pairing module: used to pair clauses with different spans in the comment corpus with each clause as a pivot based on the obtained clause representation vectors, and integrate span information into the clause update representation vectors;

[0073] Sentiment-reason pair prediction module: used to integrate prediction results of sentiment and reasons, relative positions and other information, and then use a multi-dimensional information interaction mechanism to screen and predict sentiment-reason pairs;

[0074] Hierarchy of needs matching module: used to correspond the predicted sentiment-reason pairs to Maslow's hierarchy of needs and extract the essence of needs;

[0075] Product innovation module: inputs the audience needs classified based on Maslow's hierarchy of needs and the customer business needs into the X model and outputs product innovation functions.

[0076] Explanations of some professional terms in the present invention are as follows:

[0077] Bi-LSTM: The full name is Bidirectional Long Short-Term Memory Network. The principle of Bi-LSTM is to add a reverse LSTM network to the LSTM network, enabling the model to consider both past and future information simultaneously. In the forward LSTM, the input sequence is processed from left to right, while in the reverse LSTM, the input sequence is processed from right to left. The outputs of the two LSTM networks are concatenated to form the final output of Bi-LSTM.

[0078] Bi-GRU: The full name is Bidirectional Gated Recurrent Unit. It is a network structure that combines the characteristics of bidirectional recurrent neural networks (BRNN) and gated recurrent units (GRU). Bi-GRU consists of an input layer, a forward calculation layer, a backward calculation layer, and an output layer.

[0079] Word2vec algorithm: It is an algorithm for generating word vectors. It learns the vector representation of each word by training a neural network model. The Word2vec algorithm includes two main models: CBOW (Continuous Bag of Words) and Skip-gram, which predict the context of words in different ways. The CBOW model predicts the target word based on the words in the context, while the Skip-gram model predicts the words in the context based on the target word.

[0080] WordNet: It groups entries according to the meaning of the words. Each group of entries with the same meaning is called a synset (synonym set). WordNet provides a short and concise definition for each synset and records the semantic relationships between different synsets. Nouns, verbs, adjectives, and adverbs are each organized into an independent synonym network. Each synonym set represents a basic semantic concept, and these sets are connected to each other through various relationships.

[0081] The present invention has the following beneficial effects:

[0082] The present invention provides an innovative demand analysis method and system based on span association prediction. Compared with the prior art:

[0083] 1. The extraction of emotion - cause pairs in the present invention can not only capture the emotional reactions of the audience, but also deeply explore the specific reasons that cause these emotions. This in - depth analysis helps to more accurately understand the internal motivations and actual problems encountered by the audience when using products or services, thus providing a more targeted basis for product innovation. By analyzing the reasons behind the emotions, the specific details and priorities of the audience's needs can be revealed.

[0084] 2. The method provided by the present invention can accurately locate the emotional expressions and their direct or indirect causes in a wide range of text environments through span association prediction, and delve into the specific situations and factors that trigger these emotions, providing a more detailed and specific data basis for demand analysis.

[0085] 3. The demand analysis method provided by the present invention combines Maslow's hierarchy of needs analysis, which can place the identified emotions and needs within the framework of human basic needs, helping decision - makers understand which needs are basic survival needs, which involve security, social interaction, respect, or self - actualization, so as to prioritize the solution of the most urgent needs and maximize the product value.

[0086] 4. In the innovation demand analysis method provided by the present invention, the X model is combined for analysis. By bridging the business requirement items of customers and the requirement items of the audience, the X model can conduct comprehensive analysis, taking into account both the commercial goals of customers and the actual needs and experiences of the audience. This bridging of dual demand sources, based on in-depth mining and comprehensive innovation of multi-dimensional information, makes the analysis results more comprehensive and accurate. By clarifying the association between the goals of different types of users and product functions, it is ensured that the product design can respond to the deep human needs of users. At the same time, the model encourages being both in-depth and appropriate when understanding the demand goals, avoiding over-design, and maintaining the agility and pertinence of product iteration. Description of the Drawings

[0087] Figure 1 It is a schematic diagram of the overall process of the innovation demand analysis method based on span correlation prediction provided in Embodiment 1 of the present invention.

[0088] Figure 2 It is a structural framework diagram of the X model provided in the embodiments of the present invention.

[0089] Figure 3 It is a structural framework diagram of the innovation demand analysis system based on span correlation prediction provided in Embodiment 2 of the present invention. Detailed Embodiments

[0090] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0091] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0092] Embodiment 1:

[0093] As Figure 1 shown, this embodiment provides an innovation demand analysis method based on span correlation prediction, including the following steps:

[0094] S1. Comment preprocessing: Obtain the product comment corpus and train it to obtain the corresponding word vectors, and use Bi-LSTM and Bi-GRU to perform two-level encoding on the clauses in the comment corpus, and list the obtained clause representation vectors one by one. Specifically, it includes:

[0095] S101. Pre-training model: Obtain product comments on a large-scale Chinese microblog corpus and use the Word2vec algorithm for training to obtain the corresponding word vectors;

[0096] S102. Two - level Encoding: Use Bi - LSTM and Bi - GRU to perform two - level encoding on the clauses in the review corpus. List the obtained clause representation vectors one by one. Taking each clause as a pivot, pay attention to its context information from multiple aspects, and enhance the subordinate clauses with targeted information to strengthen the causal relationship between clauses. The specific method of two - level encoding in step S102 is as follows:

[0097] (1) Word - level Encoding: For each word in a clause, use the word vector obtained from the Word2vec model as the initial input feature w of the word n ;

[0098] (2) Clause - level Encoding: After completing the word - level encoding, use w n as the input of Bi - LSTM and Bi - GRU, model the context relationship between words, and obtain the clause representation vector S of each clause n .

[0099] S2. Span Representation: Combine the boundary representation related to the context and the attention mechanism on the span, and apply it to the learning of sentiment and reasons to update the clause representation vector. Specifically, it includes:

[0100] Step S201. Given the review corpus D = {S1, S2, S3, …, S n}}, update the clause information through the clause representation vector S n by the following operations:

[0101]

[0102]

[0103] where, represents a three - layer feed - forward neural network, which performs secondary processing on the clause representation vector S n to obtain a new feature vector a n , represents the attention weight of a n of each clause within the span i, softmax represents the normalization process, span start (i) and span end (i) represent the start position and end position of the span i respectively, exp() represents the exponential operation, ∑ represents the summation operation, ⊙ represents the element - by - element product of two matrices, represents the clause representation vector after information interaction;

[0104] Step S202. Within the span w, combine the clause representation vector S n output by Bi - GRU with the clause representation after information interaction obtained in step S201 Perform fusion to finally obtain the clause update representation vector X n :

[0105]

[0106] Among them, add[] represents vector concatenation. Concatenate the clause representation vector S n and the clause representation vector after information interaction to obtain the concatenated clause update representation vector X n .

[0107] S3. Span association pairing; Using the clause representation vectors obtained in step S1, with each clause as the pivot, pair the clauses with different spans in the comment corpus one by one, and integrate the span information into the clause update representation vector X n obtained in step S2. Specifically, it includes:

[0108] Step 301. Take each clause in the comment corpus as the pivot, pair it with the surrounding clauses within a certain preset range, and integrate information such as relative positions to obtain the vector representation of the final candidate pair:

[0109]

[0110] Among them, W e and W c are the sentiment weight matrix and the reason weight matrix respectively, b e and b c are the sentiment bias term and the reason bias term respectively, is the i-th sentiment clause vector, is the i-th reason clause vector, E i is the sentiment prediction of the i-th clause, C i is the reason prediction of the i-th clause;

[0111] Since according to general language habits, the reason corresponding to a certain sentiment is generally within a certain range before and after the sentiment clause, and even appears in the same clause as the sentiment. Therefore, in this step, each clause in the comment corpus is used as the pivot, paired with the surrounding clauses within a certain range, and information such as relative positions is integrated to obtain the vector representation of the final candidate pair.

[0112] Then, using the clause update representation vector X n obtained through step S202 span representation, pair the clauses within the span w range to obtain the sentiment-reason candidate pair, expressed as:

[0113] Pair i,j =[X i ,X j , i ∈ [1, n], j ∈ [i - w, i + w + 1]

[0114] Among them, Pair i,j represents a candidate sentiment - reason pair composed of a sentiment clause at position i and a reason clause at position j, and w represents the span;

[0115] Step 302: In order to fuse the acquired information and provide multi - aspect information support for the prediction of the sentiment - reason pair Pair i,j Fuse the sentiment and reason predictions with the candidate sentiment - reason pair representation, and at the same time add the relative position information loc i,j to improve the prediction accuracy. The specific formula is as follows:

[0116] P i,j =[Psir i,j ,E i ,C j ,loc i,j

[0117] Among them, i represents the index of the clause at the sentiment position in the sequence, representing the specific position where the sentiment occurs, j represents the index of the clause at the reason position in the sequence, corresponding to the sentiment position i, and Pair i,j represents the candidate sentiment - reason pair, E i is the sentiment prediction of the i - th clause, and C j is the reason prediction of the j - th clause;

[0118] S4: Sentiment - reason pair prediction: Integrate the prediction results of sentiment and reason, as well as information such as relative position, and then use a multi - dimensional information interaction mechanism to screen and predict sentiment - reason pairs.

[0119] The specific process is as follows:

[0120] Obtain the representation vectors of all candidate sentiment - reason pairs through span - associated pairing Then use a feed - forward neural network combined with softmax for prediction:

[0121]

[0122] Among them, represents the prediction result of the sentiment - reason pair, FNN pair represents a 3 - layer feed - forward neural network, W pair is the weight matrix of the candidate sentiment - reason pair, and b pair is the bias term of the candidate sentiment - reason pair;

[0123] During the training process, use multi - task learning to jointly train the model, and use the cross - entropy loss function to calculate each task, namely the sentiment extraction task, the reason extraction task, and the sentiment - reason pair extraction task:

[0124] ​

[0125] and are the true prediction results of emotion, reason, and emotion-reason pairs respectively, and L emo , L cau and L pair are the loss functions of emotion, reason, and emotion-reason pairs respectively;

[0126] The loss function of the final model is finally expressed as the sum of the three:

[0127] L = L pair + l emo + L cau + λ||θ|| 2

[0128] where: λ represents the weight of the L2 regularization term, θ is the L2 regularization term of all parameters in the model, and the L2 normalization is set to 1×10 -5 .

[0129] Using Stochastic Gradient Descent (SDG) and Adam innovation with shuffled batches, the batch size and learning rate are set to 32 and 0.008 respectively, and all weight matrices and biases are randomly initialized to a uniform distribution U(0.01, 0.01).

[0130] S5. Hierarchy of needs matching: Corresponding the predicted emotion-reason pairs with Maslow's hierarchy of needs to extract the essence of needs. Maslow's hierarchy of needs in step S5 is presented in a pyramid form and includes, in order from highest to lowest level: physiological needs, safety needs, social needs, esteem needs, and self-actualization needs;

[0131] List the relevant keywords of Maslow's hierarchy of needs, such as:

[0132] Physiological needs include food, water, sleep, health, rest, comfort, etc.;

[0133] Safety needs include safety, protection, stability, work, income, health insurance, etc.;

[0134] Social needs include friendship, love, family, relationship, belonging, acceptance, social interaction, team, etc.;

[0135] Esteem needs include self-esteem, respect, recognition, status, achievement, prestige, praise, etc.;

[0136] Self-actualization needs include growth, potential, creativity, self-actualization, goals, dreams, personal development, etc.

[0137] Use a thesaurus (such as WordNet) to expand each keyword set for the above keywords to obtain a keyword library, so that the obtained keyword library covers more variants and synonyms. Then, match the sentiment - reason pairs output in step S4 with the obtained keyword library to determine the Maslow's hierarchy of needs corresponding to the sentiment - reason pairs and extract the essence of the needs.

[0138] S6. Product innovation: Input the audience needs classified based on Maslow's hierarchy of needs and the customer business needs into the X model, and output the product innovation functions. The specific analysis process in combination with the X model is as follows:

[0139] As Figure 2 shown, the two upper ends of the X model in this embodiment are respectively represented as customer business needs and audience needs, the intersection point is represented as the heterogeneous user goal, the lower left end is represented as the product innovation function, and the right end is represented as Maslow's needs (human nature). The two major analysis steps of the X model are as follows:

[0140] (1) Bridge the customer business need items and the audience need items extracted based on Maslow's hierarchy of needs to obtain the heterogeneous user goal, and then combine the analysis of human nature needs. This step aims to deeply explore the common vision and value pursuit behind this heterogeneous user group, touching the most fundamental level of needs (human nature needs). For those trivial needs or pseudo - needs that are superficial and contribute little to the core value of the product, they are marked with a dotted line, indicating that although these needs exist, they do not need to be delved into too deeply in the current analysis stage.

[0141] (2) Analyze from Maslow's human nature needs to the heterogeneous user goal and then to the product innovation function, or directly analyze from the heterogeneous user goal to the product innovation function. This step ensures that the product design not only meets the fundamental demands of human nature but also effectively solves the actual problems faced by users. This path reflects the process from the macro concept to the specific implementation, ensuring that each design of the product function responds to the deep - seated human nature needs and achieving an accurate match between the heterogeneous user goal and the product function.

[0142] Embodiment 2

[0143] As Figure 3 shown, this embodiment provides an innovation demand analysis system 20 based on span - correlation prediction, which adopts the innovation demand analysis method based on span - correlation prediction described in Embodiment 1, and specifically includes:

[0144] A comment pre - processing module 21, including a pre - trained sub - model 211 and a two - level encoding sub - module 212, where,

[0145] The pre - trained sub - model 211 is used to obtain product comments and perform training to obtain corresponding word vectors;

[0146] A two-level encoding submodule 212 is used to perform two-level encoding on the clauses in the review corpus using Bi-LSTM and Bi-GRU, and list the obtained clause representation vectors one by one;

[0147] The span representation module 22 is used to combine the context-related boundary representation with the attention mechanism on the span, and apply it to the learning of emotions and reasons to obtain the clause update representation vector;

[0148] The span association pairing module 23 is used to pair the obtained clause representation vectors with clauses of different spans in the review corpus one by one, taking each clause as a pivot, and integrating the span information into the clause update representation vector;

[0149] The emotion-cause pair prediction model 24 is used to integrate the prediction results of emotion and cause as well as information such as relative position, and then use the multi-dimensional information interaction mechanism to screen and predict the emotion-cause pair;

[0150] The needs hierarchy matching module 25 is used to match the predicted emotion-reason pairs with Maslow's needs hierarchy to extract the essence of the needs;

[0151] The product innovation module 26 inputs the audience needs and customer business needs based on Maslow's hierarchy of needs into the X-model and outputs product innovation functions.

[0152] Example 3: Application Example

[0153] In order to verify the feasibility and effectiveness of the method proposed in the present invention, this embodiment is described by taking the museum's Asian civilization exhibition hall project as an example.

[0154] The following audience comment corpus is available:

[0155] "My friend and I came here for a trip. We were looking forward to visiting the local museum. We were interested in the exhibits that others commented online as worth seeing. However, there were too many people taking photos of the beautiful exhibits. It took a long time to make a special trip to take a look, and we felt tired and annoyed after queuing for a long time"; "It's very comfortable to bring the children here for a stroll when I have time. It would be even better if there were free refreshments. The content of the exhibition hall is very rich, which allows children to learn more about history. However, there are too many words and the child doesn't like to read them. She is not interested in listening to my explanations. She has indeed come here so many times, but she is no longer interested and doesn't want to come anymore. There is nothing new."

[0156] The method provided in Example 1 is adopted to obtain emotion-reason candidate pairs through span association pairing, for example: "Looking forward to - traveling, checking in to visit the museum, recommending exhibits, recommending exhibits online, too many people taking photos of the beautiful exhibits, waiting for a long time, queuing for a long time", and then predicting the candidate pairs to obtain the following emotion-reason pairs: "Looking forward to - traveling, checking in to visit the museum", "Interested - recommending exhibits online", "Tired and annoyed - too many people taking photos of the beautiful exhibits, waiting for a long time, queuing for a long time", "Quite comfortable - bring the children here for a stroll", "Even better - free refreshments", "Quite rich - the content of the exhibition hall, let the children learn more about history", "Not interested in listening - too many words", "Not interested and unwilling to come - I have been here so many times, there is nothing new".

[0157] For the above audience comment corpus, the predicted emotion-reason pairs are matched with Maslow's hierarchy of needs through the WordNet keyword library and displayed in the form of a pyramid, including the following in descending order:

[0158] Physiological needs L1: Physical fatigue caused by long waiting and queuing reflects the basic physiological needs for rest and a comfortable environment; the desire for free refreshments involves basic dietary needs.

[0159] Safety Need L2: Not directly expressed in the comments, the museum filled the potential safety management gap by ensuring that the venue had good lighting and unobstructed emergency exits, meeting the audience's potential needs for a safe environment.

[0160] Social Need L3: Visiting a museum with friends reflects the need for a sense of belonging and social interaction; the desire to take photos also reflects the social need to share experiences in social circles to gain recognition and a sense of belonging.

[0161] Respect Need L4: The pursuit of visiting highly rated exhibits reflects the need for respect for knowledge, aesthetics and cultural values.

[0162] Self-actualization needs L5: Pursuing the depth and breadth of knowledge to meet the needs of personal growth and potential realization; tourists’ desire for novelty and innovation; parents hope that their children can learn through visits, which reflects the need for their children’s achievement and self-improvement.

[0163] Bridging the requirements of heterogeneous users, i.e., the requirements put forward by the customer: The customer hopes to build a mobile platform with some introductions about this exhibition hall on the home page. When visitors enter the exhibition hall, they can scan the code to enter the platform. While visitors are viewing the exhibits in the exhibition hall, they can simultaneously view more detailed information about the exhibits on their mobile phones. The customer also hopes to use the multimedia projectors in the exhibition hall to play some historical materials related to the exhibits, plus some special effects, to provide visitors with a better visual experience. Finally, if possible, use some methods to attract customers to the museum's cultural and creative products to achieve the purpose of promotion. However, when asked whether the form of display must be the client interface of the mobile platform, the customer's requirement became "as long as it can meet the requirements of the museum, the form and content can be changed."

[0164] Ignoring the means and processes, analyze the original requirements, and extract the customer's business requirement items: detailed information introduction of the exhibition hall D1; providing visitors with a better exhibition experience D2; promoting cultural and creative products D3.

[0165] Based on the audience needs classified by Maslow's hierarchy of needs, bridge the customer's business needs, further explore the goals of heterogeneous users, and put forward innovative product function suggestions by combining human needs.

[0166] Step 1, bridge the audience need items based on Maslow's hierarchy of needs and the customer's business need items to obtain the goals of heterogeneous users, and then analyze in combination with human needs:

[0167] 1. Bridge the audience's esteem needs L4 and self-actualization needs L5 with the customer's business need items D1 and D2. While meeting the basic acquisition of historical information and enjoyment during the exhibition process, pursue the depth and connotation of cultural experience, absorb the rich historical knowledge of the exhibition, and achieve the improvement of personal cultural qualities and emotional resonance.

[0168] 2. Bridge the audience's physiological needs L1, safety needs L2, and self-actualization needs L5 with the customer's business need item D2. Visitors pursue to absorb rich historical knowledge in a safe and comfortable environment, especially a learning space friendly to children, to cultivate the learning interests and continuous learning abilities of the next generation, which reflects the pursuit of a safe and comfortable environment and the emphasis on the growth and education of future generations.

[0169] 3. Bridge the audience's social needs L3 with the customer's business need item D3. Exhibits popular among the public often can arouse the emotional resonance of visitors. Designing relevant cultural and creative products can allow visitors to take away souvenirs with emotional connections. As an extension of the exhibits, cultural and creative products enable visitors to have a physical carrier of the museum experience, meet the collection desire and display needs, and expand the influence and popularity of the museum, increase income, and spread cultural value and historical connotation more widely to achieve the purpose of education and cultural inheritance.

[0170] 4. Bridge the audience's social needs L3 and the customer's business needs D2. Some visitors come to the museum because of its popularity and exquisite exhibits, and are eager to witness and share unique and impressive exhibits and visiting experiences. Sharing on social networks not only records personal footprints, but also satisfies the audience's desire to show their personal cultural tastes and experiences on social platforms, in order to achieve social recognition, respect and self-worth, and also virtually promotes the museum's word-of-mouth communication and brings wider exposure;

[0171] Step 2: Starting from the heterogeneous user goals and combining human needs, ensure that every design of product function is a response to deep-seated human needs, and put forward specific product function innovation suggestions:

[0172] 1. Through heterogeneous user target 1, we launched the online guide platform function, placed the exhibition hall introduction on the homepage, and visitors scanned the code to enter the platform; browse the exhibition hall guide page to quickly understand the exhibit layout and key exhibit information; check the brief graphic introduction and short video of the relevant exhibits online, and choose the desired viewing route.

[0173] 2. Through heterogeneous user target 2, launch customized family guided tour routes and supporting parent-child exhibition manuals to encourage family participation, improve the cultural literacy of family members, and promote children's long-term learning.

[0174] 3. Heterogeneous user goal 1 bridges goal 2, launching an interesting interactive knowledge area. An entertaining knowledge interactive area is added at the exit of the exhibition hall. A large screen is set up at the exit for multimedia projection. Animation teaching and knowledge competitions are used to awaken children's enthusiasm for learning. All visitors are encouraged to actively participate through gamified interaction, which activates historical knowledge in an interesting way and stimulates their desire to learn.

[0175] Further bridge heterogeneous user goal 3, through game interactive quizzes to get rewards and cultural and creative products to bridge, visitors can win points through games, redeem cultural and creative products or coupons, which not only enhances knowledge understanding, but also increases the fun of participation, and can also meet the customer needs of promoting cultural and creative products. Game quizzes serve as knowledge consolidation and testing, encourage repeated visits to seek correct answers, effectively extend the stay time, achieve the perfect integration of education and entertainment, and enhance the attraction of the exhibition and the effectiveness of education.

[0176] 4. Through heterogeneous user target 3, launch the design and sales of cultural and creative products, cooperate with designers and artists to design beautiful and practical cultural and creative products; set up cultural and creative product areas, develop offline and online dual channels, and facilitate tourists to purchase through multiple channels; combine popular exhibits for marketing promotion, such as launching exhibit theme series products to increase tourists' desire to buy.

[0177] 5. Bridge Target 3 through Heterogeneous User Target 2, and create a rest area and a photo-taking area beside the cultural and creative store. The rest area provides free drinking water and paid food related to cultural and creative peripherals to meet the rest needs of tourists, and at the same time promotes the consumption of cultural and creative products; add a photo-taking area, design unique backgrounds related to the exhibits in the exhibition hall, facilitate tourists to take photos, punch in and share, and provide photo printing services, etc.

[0178] 6. Bridge Target 4 through Heterogeneous User Target 3. Based on the user-generated content (UGC) incentive mechanism, set up activities such as "Best Sharing", encourage tourists to upload photos and experiences of their visits, and excellent works can obtain coupons for cultural and creative products or customized souvenirs, enhancing the sense of participation and belonging of tourists. At the same time, through the word-of-mouth communication of tourists, improve the popularity and attractiveness of the museum, and attract more tourists to come.

[0179] In summary, the method provided by the present invention can extract emotion - reason pairs more accurately. The emotion - reason pairs extracted by the present invention can not only capture the emotional reactions of the audience, but also deeply explore the specific reasons that cause these emotions. Further in-depth analysis in combination with the X model helps to more accurately understand the internal motivation and actual problems encountered by the audience when using products or services, and bridge the business needs of customers, thereby providing a more targeted basis for product innovation.

[0180] The above are only some preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An innovation demand analysis method based on span correlation prediction, characterized in that: The specific steps include: S1. Review preprocessing: Obtain product reviews and train them to obtain corresponding word vectors, and use Bi-LSTM and Bi-GRU to perform two-level encoding on the clauses in the review corpus, and list the obtained clause representation vectors one by one; S2, span representation: Combining the context-related boundary representation with the attention mechanism on the span, it is applied to the learning of emotions and reasons to obtain the clause update representation vector; S3, span association pairing: Using the clause representation vector obtained in step S1, take each clause as the pivot, pair the clauses of different spans in the review corpus one by one, and integrate the span information into the clause update representation vector in step S2; S4, emotion-cause pair prediction: Integrate the prediction results of emotion and cause as well as the relative position information, and then use the multi-dimensional information interaction mechanism to screen and predict emotion-cause pairs; S5, matching the demand hierarchy: matching the predicted emotion-reason pair with Maslow's hierarchy of needs, and extracting the essence of the needs; specifically, Maslow's hierarchy of needs is presented in the form of a pyramid, and includes, in descending order, physiological needs, safety needs, social needs, respect needs, and self-actualization needs; listing the relevant keywords of each demand hierarchy, and using the synonym library to expand each keyword set to obtain a keyword library, matching the emotion-reason pair output in step S4 with the obtained keyword library, determining the Maslow's hierarchy of needs corresponding to the emotion-reason pair, and extracting the essence of the needs; S6. Product innovation: Input the audience needs and customer business needs based on Maslow's hierarchy of needs into the X model and output product innovation functions; the specific analysis process is as follows: The upper two ends of the X-model represent customer business needs and audience needs, respectively. The intersection represents heterogeneous user goals. The lower left end represents product innovation functions, and the right end represents Maslow's needs. The two major analysis steps of the X-model are as follows: (1) Bridge the customer business demand items and the audience demand items extracted based on Maslow's hierarchy of needs to obtain heterogeneous user goals, and then combine them with human needs analysis; (2) From Maslow's human needs analysis to heterogeneous user goal analysis and then to product innovation function analysis or from heterogeneous user goal analysis directly to product innovation function analysis; By bridging the customer's business needs and the audience's needs, we can clarify the relationship between heterogeneous user goals and product functions, and ensure that product design can respond to users' deep human needs.

2. The innovation demand analysis method based on span association prediction according to claim 1 is characterized in that: Step S1 specifically includes: S101, pre-trained model: obtain product reviews on a large-scale Chinese Weibo corpus and use the Word2vec algorithm to train to obtain the corresponding word vectors; S102. Two-level encoding: Use Bi-LSTM and Bi-GRU to perform two-level encoding on the clauses in the review corpus, list the obtained clause vectors one by one, take each clause as the fulcrum, pay attention to its contextual information in multiple aspects, and enhance the clauses with targeted information to strengthen the causal relationship between clauses.

3. The innovation demand analysis method based on span association prediction according to claim 2 is characterized in that: The specific method of the two-level encoding in step S102 is: (1) Word-level encoding: For each word in a sentence, the word vector obtained from the Word2vec model is used as the initial input feature w of the word. n ; (2) Clause-level encoding: After completing word-level encoding, w n As the input of Bi-LSTM and Bi-GRU, the contextual relationship between words is modeled to obtain the clause representation vector S of each clause n .

4. The innovation demand analysis method based on span association prediction according to claim 3 is characterized in that: Step S2 specifically includes: Step S201: Given a review corpus D = {S1, S2, S3, ..., S n }, and represent the vector S through the clause n To update the clause information, do the following: in, represents a 3-layer feedforward neural network, and the clause represents the vector S n Perform secondary processing to obtain a new feature vector a n , A represents each clause n The attention weight within span i, softmax represents normalization, span start (i) and span end (i) represents the starting position and the ending position of span i, exp() represents exponential operation, ∑ represents summation operation, ⊙ represents the product of the elements at corresponding positions of two matrices, The clause representation vector after information interaction; Step S202: Within the span w, the clause representation vector S output by Bi-GRU is converted into n The clause representation vector after interacting with the information obtained in step S201 Fusion is performed, and finally the clause update representation vector X is obtained n : Among them, add[] represents vector concatenation, and the clause represents vector S n The clause representation vector after information interaction Concatenate and obtain the concatenated clause update representation vector X n .

5. The innovation demand analysis method based on span association prediction according to claim 4 is characterized in that: Step S3 specifically includes: Step 301: Take each clause in the review corpus as a pivot, pair it with surrounding clauses within a certain preset range, and fuse the relative position information to obtain the vector representation of the final candidate pair: Among them, W e and W c are the sentiment weight matrix and the cause weight matrix, respectively, e and b c are the emotional bias term and the cause bias term, is the i-th sentiment clause vector, is the i-th reason clause vector, E i is the sentiment prediction of the ith clause, C i Predict the cause of the i-th clause; The clause obtained by span representation in step S202 updates the representation vector X n , pair the clauses in the span w to obtain the emotion-reason candidate pairs, expressed as: Pair i,j =[X i ,X j ],i∈[1,n],j∈[i-w,i+w+1] Among them, Pair i,j represents the candidate emotion-reason pair consisting of the emotion clause at position i and the reason clause at position j, and w represents the span; Step 302: Fusion the emotion and cause predictions with the candidate emotion-cause pair representations, while adding relative position information loc i,j To improve the prediction accuracy, the specific formula is as follows: P i,j =[Pair i,j ,E i ,C j ,loc i,j ] Among them, i represents the index of the clause at the emotion position in the sequence, indicating the specific position where the emotion occurs; j represents the index of the clause at the cause position in the sequence, corresponding to the position of emotion i. i,j represents the candidate emotion-reason pair, E i is the sentiment prediction of the ith clause, C j Predict the cause of the j-th clause.

6. The innovation demand analysis method based on span association prediction according to claim 1 is characterized in that: The specific process of step S4 is as follows: Obtain the representation vectors of all candidate emotion-reason pairs through span association pairing Then use a feedforward neural network combined with softmax for prediction: in, Represents the prediction result of sentiment-reason pair, FNN pair represents a 3-layer feedforward neural network, W pair is the weight matrix of candidate emotion-reason pairs, b pair is the candidate emotion-reason pair bias term; The cross entropy loss function is used to calculate each task, namely the sentiment extraction task, the cause extraction task, and the sentiment-cause pair extraction task: and are the true prediction results of emotion, reason and emotion-reason pair, L emo , L cau and L pair They are the loss functions for sentiment, reason, and sentiment-reason pair respectively; The loss function of the final model is expressed as the sum of the three: L=L pair +L emo +L cau +λ||θ|| 2 Where: λ represents the weight of the L2 regularization term, and θ is the L2 regularization term of all parameters in the model.

7. An innovation demand analysis system based on span correlation prediction, characterized in that: The innovation demand analysis method based on span association prediction according to any one of claims 1 to 6 is adopted, specifically comprising: The comment preprocessing module includes a pre-training sub-model and a two-level encoding sub-module, where: The pre-trained sub-model is used to obtain product reviews and train them to obtain corresponding word vectors; The two-level encoding submodule is used to perform two-level encoding on the clauses in the review corpus using Bi-LSTM and Bi-GRU, and list the obtained clause representation vectors one by one; The span representation module is used to combine the context-related boundary representation with the attention mechanism on the span, and apply it to the learning of emotions and reasons to obtain the clause update representation vector; The span association pairing module is used to pair the obtained clause representation vectors with clauses of different spans in the review corpus one by one, taking each clause as the pivot, and integrating the span information into the clause update representation vector; The emotion-cause pair prediction module is used to integrate the prediction results of emotion and cause as well as the relative position information, and then use the multi-dimensional information interaction mechanism to screen and predict the emotion-cause pairs; The needs hierarchy matching module is used to match the predicted emotion-reason pairs with Maslow's needs hierarchy and extract the essence of needs; The product innovation module inputs the audience needs and customer business needs based on Maslow's hierarchy of needs into the X model and outputs product innovation functions.

Images