Text specific information identification and judgment method

Abstract

The invention relates to the field of Internet text content recognition, in particular to a text specific information recognition and judgment method, which comprises the following steps of: (S1) acquiring text data, performing preprocessing and word segmentation, calculating a hash value of each word through an improved hash algorithm, forming a dictionary, calculating short-term and long-term frequencies by utilizing a multi-level time sequence model, and calculating a short-term frequency and a long-term frequency; identifying abnormal high-frequency words as potential new word candidate words, and mapping the abnormal high-frequency words to a high-dimensional vector space to generate a comprehensive semantic vector; and (S2) constructing a context semantic graph model by using the comprehensive semantic vector, analyzing context compatibility of words, constructing a multi-task learning sentiment analysis model, respectively calculating sentiment scores and specific field correlation scores, comprehensively calculating the scores to obtain the total confidence coefficient of each candidate word, identifying potential new words, and obtaining the candidate words. According to the invention, the capability of identifying hidden illegal contents and the reliability of identifying and judging specific text information can be improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet text content recognition, and in particular to a method for identifying and determining specific text information. Background Art

[0002] With the popularization and development of the Internet, social media plays a great role in the dissemination of information in people's daily lives. A large number of new Internet words are also widely spread through these social media. Some new Internet words are summaries of social hot spots, while others are similar substitutes for specific words. Since new Internet words are not fully discovered, it will lead to word segmentation errors, and the quality of word segmentation will directly affect the effect of text analysis.

[0003] There are at least the following technical problems in the existing technology: the existing technology is prone to word segmentation errors, affecting the subsequent analysis results, and is unable to fully capture the frequency differences of words in the short and long term, resulting in inaccurate recognition of abnormally high-frequency words; lacks comprehensive consideration of word context information, paragraph information and full-text information, resulting in incomplete capture of word semantic information, affecting the accuracy of new word recognition; it is unable to dynamically capture the performance of words in different contexts, resulting in poor polysemous word disambiguation effect, and low reliability and accuracy of new word recognition. Summary of the invention

[0004] In order to overcome the deficiencies in the prior art, the present invention provides a method for identifying and determining text specific information.

[0005] To achieve the above object, the present invention discloses a method for identifying and determining specific information in a text, comprising the following steps:

[0006] (S1) obtaining text data and performing preprocessing and word segmentation, calculating the hash value of each word through an improved hash algorithm to form a dictionary, using a multi-level time series model to calculate short-term and long-term frequencies, identifying abnormally high-frequency words as potential new word candidates, and mapping them to a high-dimensional vector space to generate a comprehensive semantic vector;

[0007] (S2) Use the comprehensive semantic vector to build a contextual semantic graph model, analyze the contextual compatibility of words, build a multi-task learning sentiment analysis model, calculate the sentiment score and specific domain relevance score respectively, and obtain the total confidence of each candidate word by comprehensively calculating each score to identify potential new words.

[0008] Preferably, in step (S1), an improved hash algorithm is used to perform word segmentation processing on the text data. The improved hash algorithm improves the accuracy of word segmentation by increasing the weight and position parameters of the words, and groups words or phrases with the same hash value together to form a dictionary. The formula for calculating the hash value by the improved hash algorithm is as follows:

[0009]

[0010] Among them, h(w i ) represents the i-th word w i The hash value of S is the length of the word, c s is the weight of the sth position, p s (w i ) The i-th word w i The hash value of the character at position s, where m is the size of the hash table.

[0011] Preferably, in step (S1), after obtaining the dictionary, the frequency of occurrence of each word is counted; in order to capture the changes of words over time, a multi-level time series model is constructed to identify abnormal high-frequency words through the changes in word frequency in the short term and long term; the short-term frequency F s (w i , t) and the long-term frequency F l (w i , t) is calculated as follows:

[0012]

[0013] Among them, F s (w i , t) and F l (w i , t) represent the word w i The frequency at short and long time t, N s (t) and N l (t) is the total number of texts in the short and long term, δ(w i , w k ) is the indicator function, when w i =w k 1 when ω is 1, otherwise 0. s and ω l is the frequency, φ s and φ l is the phase shift.

[0014] Preferably, in step (S1), the frequency difference between words in the short term and the long term is calculated to identify abnormally high-frequency changes, the frequency difference is standardized, and a threshold is set. If the standardized frequency difference exceeds the threshold, it is considered that the current word has an abnormally high frequency phenomenon within time t, and all words that meet the conditions are output as abnormally high-frequency words, thereby identifying words that appear abnormally high in a specific time period as potential new word candidates.

[0015] Preferably, in step (S1), the context information and paragraph / full text information are mapped to a low-dimensional vector space through a mapping function to obtain a semantic vector representation; the low-dimensional vector is further processed through a multi-layer perceptron to generate a high-dimensional comprehensive semantic vector, as shown below:

[0016]

[0017] in, For the word w i The comprehensive vector representation of α 1j Embedding vector representing the context words The weight in the comprehensive semantic vector, β 1j and γ 1j They are the embedding vectors of the context words in Through the linear transformation coefficient and bias before the tanh activation function, α2 represents the paragraph embedding vector The weights in the comprehensive semantic vector, β2 and γ2 represent the paragraph embedding vector The linear transformation coefficient and bias before passing through the relu activation function, α3 is the full text embedding vector The weights in the comprehensive semantic vector, β3 and γ3 represent the full-text embedding vector Linear transformation coefficients and bias before passing through the relu activation function.

[0018] Preferably, in step (S2), each word in the potential new word cluster and its context information will be used to generate nodes and edges of the semantic graph, build an improved context semantic graph model, and analyze the compatibility of words in multi-layer contexts; the context semantic graph model takes into account the semantic relationship between sentences, paragraphs, and full texts, and can more comprehensively analyze the performance of words in different contexts; a self-supervised learning graph convolutional network is used to embed nodes in the semantic graph and perform polysemous word disambiguation.

[0019] Preferably, in step (S2), a multi-task learning sentiment analysis model is constructed to simultaneously perform sentiment classification and specific domain relevance determination; the multi-task learning model integrates comprehensive vector representation information and can more accurately analyze the sentiment tendency of words and their contexts; the sentiment analysis model is optimized by a cross entropy loss function to calculate the sentiment scores of words and their contexts and the specific domain relevance determination results.

[0020] The sentiment score of a word and its context is calculated using the following formula:

[0021]

[0022] The domain-specific relevance score of a term and its context is calculated using the following formula:

[0023]

[0024] In the above formula, S em (w i ) represents the word w i The sentiment score, S ha (w i ) represents the word w i The domain-specific relevance score, λ p and θ p is the sentiment analysis parameter of the pth sentiment classifier, λ q and θ q is the domain-specific relevance determination parameter of the qth domain-specific relevance classifier, σ is the activation function, and are the weight vectors of sentiment and domain-specific relevance classifiers, respectively, P and Q are the number of classifiers, For the word w i A comprehensive vector representation of .

[0025] Preferably, through the comprehensive confidence formula, the multi-level and multi-stage analysis results can be integrated, the average of the short-term frequency and the long-term frequency is used as the frequency score of the word, and then combined with the context compatibility score, sentiment score, specific field relevance score and the gradient norm with the same order of magnitude of the word, so as to obtain the total confidence of each candidate word; the total confidence of each candidate word is as follows:

[0026]

[0027] In the above formula, C total (w i ) represents the word w i The comprehensive confidence of η1, η2, η3, η4, η5 are weighted coefficients, which are obtained by optimizing the training data. Word w i The gradient norm in vector space, S cp (w i ) represents the word w i The context compatibility score, S em (w i ) represents the word w i The sentiment score, S ha (w i ) represents the word w i Domain-specific relevance scores.

[0028] The present invention has the following technical effects:

[0029] 1. The accuracy of word segmentation is improved by using an improved hash algorithm, thereby enhancing the effect of text preprocessing. The frequency analysis of the segmented words is performed through a multi-level time series model, which can accurately identify abnormally high-frequency words as potential new word candidates. A multi-level text fusion method is introduced to map the context information, paragraph information and full-text information of abnormally high-frequency words into a high-dimensional vector space to generate a comprehensive semantic vector.

[0030] 2. Cluster the semantic vectors in high-dimensional space to identify potential new word clusters; build an improved contextual semantic graph model to analyze the compatibility of words in multi-layer contexts, and use self-supervised learning graph convolutional networks to perform node embedding and polysemy disambiguation; context compatibility analysis can more comprehensively identify potential new words that are incompatible with the context, further improving recognition accuracy.

[0031] 3. By constructing a multi-task learning sentiment analysis model, sentiment classification and specific domain relevance judgment are performed simultaneously. The multi-task learning model integrates the information represented by the comprehensive vector, and can more accurately analyze the sentiment tendency and specific domain relevance of words and their contexts; the results of sentiment analysis and specific domain relevance judgment are effectively applied in the calculation of comprehensive confidence, which improves the ability to recognize content in specific fields.

[0032] 4. Through multi-level weighted calculation, the frequency score, context compatibility score, sentiment score, specific field relevance score and gradient norm of the word are comprehensively considered to obtain the total confidence of each candidate word; setting multi-level confidence thresholds and making graded judgments on the total confidence can accurately identify new words and improve the reliability of the judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a flowchart of a method for identifying and determining text-specific information. DETAILED DESCRIPTION

[0034] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0036] The following specifically describes the specific solution of a method for identifying and determining text-specific information provided by the present invention in conjunction with the accompanying drawings.

[0037] Refer to the attached Figure 1 , which shows a flowchart of a method for identifying and determining text-specific information provided by an embodiment of the present invention. The method includes the following steps:

[0038] S1. Obtain text data and perform preprocessing and word segmentation. Calculate the hash value of each word through an improved hash algorithm to form a dictionary. Use a multi-level time series model to calculate short-term and long-term frequencies, identify abnormally high-frequency words as potential new word candidates, and map them to a high-dimensional vector space to generate a comprehensive semantic vector;

[0039] Obtain text data. For the input text data, use a distributed computing framework for preprocessing, and the preprocessing method can adopt existing technologies; perform word segmentation on the preprocessed text data, and the text data is segmented into multiple parts, and each part is processed in parallel on different computing nodes.

[0040] The steps of word segmentation are as follows:

[0041] Segment the input text into words or phrases and remove stop words (such as meaningless words like "de", "le", etc.); calculate the hash value of each word or phrase, and group the words or phrases with the same hash value together to form a dictionary, generating a dictionary containing all words and their corresponding hash values.

[0042] Specifically, use an improved hash algorithm to perform word segmentation on text data. The improved hash algorithm improves the accuracy of word segmentation by increasing the weight and position parameters of words. The specific formula is as follows:

[0043]

[0044] Among them, h(w i ) represents the hash value of the i-th word w i , S is the word length, c s is the weight of the s-th position, p s (w i ) is the hash value of the character of the i-th word w i at the s-th position, and m is the size of the hash table.

[0045] Furthermore, after obtaining the dictionary, count the occurrence frequency of each word; in order to capture the changes of words over time, construct a multi-level time series model, and identify abnormally high-frequency words through the word frequency changes in the short term and long term. Specifically, the short-term frequency F s (w i , t) and the long-term frequency F l (wi , t) is calculated as follows:

[0046]

[0047] Among them, F s (w i , t) and F l (w i , t) represent the word W i The frequency at short and long time t, N s (t) and N l (t) is the total number of texts in the short and long term, δ(w i , w k ) is the indicator function, when w i =w k 1 when ω is 1, otherwise 0. s and ω l is the frequency, φ s and φ l is the phase shift.

[0048] Calculate the difference in the frequency of words in the short term and the long term to identify abnormally high-frequency changes. Define the frequency difference metric ΔF(w i , t) are as follows:

[0049] ΔF(w i ,t)=|F s (w i , t)-F l (w i ,t)|,

[0050] Normalize the frequency differences for better comparison. You can use z-score normalization:

[0051]

[0052] Among them, z(w i , t) is the normalized frequency difference, μ ΔF and σ ΔF They represent the mean and standard deviation of the frequency differences of all words at time t.

[0053] Set a threshold T z , if the normalized frequency difference z(w i , t) exceeds the threshold T z , it is considered that the current word has an abnormally high frequency phenomenon within time t, and all words that meet the conditions w i The words are output as abnormally high-frequency words, thereby identifying words that appear abnormally frequently in a specific time period as potential new word candidates.

[0054] Furthermore, in order to more comprehensively capture the semantic information of words, the abnormally high-frequency words and their context information and paragraph / full-text information are mapped to a high-dimensional vector space; the context information refers to the abnormally high-frequency words (i.e., the target word w i ) in the sentence where it is located; setting the context window size to c, the context word set {u1, u2, ..., u c} contains w i c words before and after; the paragraph / full text information refers to the target word w i The semantic vector of the paragraph or the whole text; assuming that the vector of the paragraph where the target word is located is d i , where the vector of the full text is f i .

[0055] Specifically, the context information and paragraph / full text information are mapped to a low-dimensional vector space through a mapping function to obtain a semantic vector representation in, is the embedding vector of the jth context word, is the target word w i The embedding vector of the paragraph, is the target word w i The embedding vector of the full text.

[0056] The low-dimensional vector is further processed through a multi-layer perceptron to generate a high-dimensional comprehensive semantic vector. The specific formula is as follows:

[0057]

[0058] in, For the word w i The comprehensive vector representation of α 1j Embedding vector representing context words The weight in the comprehensive semantic vector, β 1j and γ 1j They are the embedding vectors of the context words in Through the linear transformation coefficient and bias before the tanh activation function, α2 represents the paragraph embedding vector The weights in the comprehensive semantic vector, β2 and γ2 represent the paragraph embedding vector The linear transformation coefficient and bias before passing through the relu activation function, α3 is the full text embedding vector The weights in the comprehensive semantic vector, β3 and γ3 represent the full-text embedding vector Linear transformation coefficients and bias before passing through the relu activation function.

[0059] After generating the comprehensive semantic vector, a clustering algorithm is used to cluster the semantic vector in a high-dimensional space to identify potential new word clusters; the multi-view clustering algorithm calculates the similarity between each vector through multiple iterations to form a cluster center, and updates the attribution of each vector according to the change of the cluster center. The specific steps of the clustering algorithm are as follows:

[0060] Initial cluster centers: randomly select a preset number of initial cluster centers;

[0061] Calculate similarity: Use cosine similarity to calculate the similarity between each vector and the cluster center;

[0062] Update clustering: assign each vector to the most similar cluster center based on similarity;

[0063] Update center: recalculate the center vector of each cluster;

[0064] Iteration: Repeatedly calculate similarity, update clusters, and update centers until the clustering results converge.

[0065] S2. Use the comprehensive semantic vector to build a contextual semantic graph model, analyze the contextual compatibility of words, build a multi-task learning sentiment analysis model, calculate the sentiment score and specific domain relevance score respectively, and obtain the total confidence of each candidate word by comprehensively calculating each score to identify potential new words.

[0066] After identifying potential new word clusters, the words in the potential new word clusters are used as nodes to build an improved contextual semantic graph model. Specifically, each word in the potential new word cluster and its context information will be used to generate nodes and edges of the semantic graph, build an improved contextual semantic graph model, and analyze the compatibility of words in multi-layer contexts; the contextual semantic graph model takes into account the semantic relationship between sentences, paragraphs, and full texts, and can more comprehensively analyze the performance of words in different contexts; a self-supervised learning graph convolutional network is used to embed nodes in the semantic graph and perform polysemous word disambiguation. The specific formula is as follows:

[0067]

[0068] Among them, S cp (w i ) represents the word w i The context compatibility score, M(w i ) is the word w i The set of neighbor nodes of For the word w iThe comprehensive vector representation of the neighbor node g of is λ, and λ is the adjustment parameter. Through the above formula, the compatibility score of each word in its context can be calculated, and potential new words that are incompatible with the context can be identified. The specific construction steps of the improved context semantic graph model are as follows:

[0069] Each word and its context in the potential new word cluster are represented as nodes in the context semantic graph, and the edges between nodes represent the semantic relationship between words; each node is initialized to its comprehensive semantic vector; the graph is embedded using a graph convolutional network, and the vector representation of each node is updated; the context compatibility score of each word is calculated based on the updated node vector.

[0070] Furthermore, a multi-task learning sentiment analysis model is constructed to simultaneously perform sentiment classification and specific domain relevance determination; the multi-task learning model integrates comprehensive vector representation information and can more accurately analyze the sentiment tendency of words and their contexts; the sentiment analysis model is optimized through the cross entropy loss function to calculate the sentiment score of words and their contexts and the specific domain relevance determination results. The specific formula is as follows:

[0071]

[0072] Among them, S em (w i ) represents the word w i The sentiment score, S ha (w i ) represents the word w i The domain-specific relevance score, λ p and θ p is the sentiment analysis parameter of the pth sentiment classifier, λ q and θ q is the domain-specific relevance determination parameter of the qth domain-specific relevance classifier, σ is the activation function, and are the weight vectors of sentiment and domain-specific relevance classifiers, respectively, and P and Q are the number of classifiers.

[0073] After calculating the sentiment score and specific domain relevance score of each word, perform multi-level weighted calculation on the analysis results to obtain the total confidence of each candidate word. The specific formula is as follows:

[0074]

[0075] Among them, C total (w i ) represents the word w i The comprehensive confidence of η1, η2, η3, η4, η5 are weighted coefficients, which are obtained by optimizing the training data. Word wi Gradient norm in vector space. Through the comprehensive confidence formula, the multi-level and multi-stage analysis results can be integrated, and the average of short-term frequency and long-term frequency can be used as the frequency score of the word. Then, the context compatibility score, sentiment score, specific domain relevance score and gradient norm of the same order of magnitude of the word are combined to obtain the total confidence of each candidate word.

[0076] Set multi-level confidence thresholds and perform graded judgments on the total confidence; if the comprehensive confidence of a candidate word exceeds the set threshold, it is judged to be a new word; output the recognition judgment result, including the recognized new word and its corresponding confidence score.

[0077] In summary, a method for identifying and determining text-specific information has been completed.

[0078] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0079] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for identifying and determining specific information in a text, characterized in that: The following steps are involved: (S1) obtaining text data and performing preprocessing and word segmentation, calculating the hash value of each word through an improved hash algorithm to form a dictionary, using a multi-level time series model to calculate short-term and long-term frequencies, identifying abnormally high-frequency words as potential new word candidates, and mapping them to a high-dimensional vector space to generate a comprehensive semantic vector; (S2) Use the comprehensive semantic vector to build a contextual semantic graph model, analyze the contextual compatibility of words, build a multi-task learning sentiment analysis model, calculate the sentiment score and specific domain relevance score respectively, and obtain the total confidence of each candidate word by comprehensively calculating each score to identify potential new words.

2. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S1), the improved hash algorithm is used to perform word segmentation on the text data. The improved hash algorithm improves the accuracy of word segmentation by increasing the weight and position parameters of the words, and groups words or phrases with the same hash value together to form a dictionary. The formula for calculating the hash value using the improved hash algorithm is as follows: Among them, h(w i ) represents the i-th word w i The hash value of S is the length of the word, c s is the weight of the sth position, p s (w i ) The i-th word w i The hash value of the character at position s, where m is the size of the hash table.

3. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S1), after obtaining the dictionary, the frequency of occurrence of each word is counted; in order to capture the changes of words over time, a multi-level time series model is constructed to identify abnormal high-frequency words through the changes in word frequency in the short and long term; the short-term frequency F s (w i , t) and the long-term frequency F l (w i , t) is calculated as follows: Among them, F s (w i , t) and F l (w i , t) represent the word w i The frequency at short and long time t, N s (t) and N l (t) is the total number of texts in the short and long term, δ(w i , w k ) is the indicator function, when w i =w k 1 when ω is 1, otherwise 0. s and ω l is the frequency, φ s and φ l is the phase shift.

4. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S1), the frequency difference between words in the short term and the long term is calculated to identify abnormally high-frequency changes, the frequency difference is standardized, and a threshold is set. If the standardized frequency difference exceeds the threshold, it is considered that the current word has an abnormally high frequency phenomenon within time t, and all words that meet the conditions are output as abnormally high-frequency words, thereby identifying words that appear abnormally high in a specific time period as potential new word candidates.

5. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S1), the context information and paragraph / full text information are mapped to a low-dimensional vector space through a mapping function to obtain a semantic vector representation; the low-dimensional vector is further processed through a multi-layer perceptron to generate a high-dimensional comprehensive semantic vector, as shown below: in, For the word w i The comprehensive vector representation of α 1j Embedding vector representing context words The weight in the comprehensive semantic vector, β 1j and γ 1j They are the embedding vectors of the context words in Through the linear transformation coefficient and bias before the tanh activation function, α2 represents the paragraph embedding vector The weights in the comprehensive semantic vector, β2 and γ2 represent the paragraph embedding vector The linear transformation coefficient and bias before passing through the relu activation function, α3 is the full text embedding vector The weights in the comprehensive semantic vector, β3 and γ3 represent the full-text embedding vector Linear transformation coefficients and bias before passing through the relu activation function.

6. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S2), each word in the potential new word cluster and its context information will be used to generate nodes and edges of the semantic graph, build an improved context semantic graph model, and analyze the compatibility of words in multi-layer contexts; the context semantic graph model takes into account the semantic relationship between sentences, paragraphs and full texts, and can more comprehensively analyze the performance of words in different contexts; A graph convolutional network with self-supervised learning is used to embed nodes in semantic graphs for polysemous word disambiguation.

7. The text specific information recognition and determination method according to claim 1, characterized in that: In step (S2), a multi-task learning sentiment analysis model is constructed to simultaneously perform sentiment classification and domain-specific relevance determination; The multi-task learning model integrates comprehensive vector representation information to more accurately analyze the sentiment tendency of words and their contexts. The sentiment analysis model is optimized through the cross entropy loss function to calculate the sentiment scores of words and their contexts and the results of domain-specific relevance determination. The sentiment score of a word and its context is calculated using the following formula: The domain-specific relevance score of a term and its context is calculated using the following formula: In the above formula, S em (w i ) represents the word w i The sentiment score, S ha (w i ) represents the word w i The domain-specific relevance score, λ p and θ p is the sentiment analysis parameter of the pth sentiment classifier, λ q and θ q is the domain-specific relevance determination parameter of the qth domain-specific relevance classifier, σ is the activation function, and are the weight vectors of sentiment and domain-specific relevance classifiers, respectively, P and Q are the number of classifiers, For the word w i A comprehensive vector representation of .

8. The text specific information recognition and determination method according to claim 1, characterized in that: Through the comprehensive confidence formula, we can integrate the multi-level and multi-stage analysis results, take the average of short-term frequency and long-term frequency as the frequency score of the word, and then combine the context compatibility score, sentiment score, specific field relevance score and gradient norm of the same order of magnitude of the word to obtain the total confidence of each candidate word; the total confidence of each candidate word is as follows: In the above formula, C total (w i ) represents the word w i The comprehensive confidence of η1, η2, η3, η4, η5 are weighted coefficients, which are obtained by optimizing the training data. Word w i The gradient norm in vector space, S cp (w i ) represents the word w i The context compatibility score, S em (w i ) represents the word w i The sentiment score, S ha (w i ) represents the word w i Domain-specific relevance scores.

Images