A viewpoint extraction and evolution path tracing system and method

Abstract

This invention relates to a system and method for opinion extraction and evolution path tracking, belonging to the field of film review analysis technology. It includes a data input module, a data processing module, a temporal alignment module, a temporal enhancement module, a deep extraction module, an analysis module, and a visualization module. The data input module acquires long film review text data, temporal metadata, and external event data. The text processing module outputs text feature vectors, temporal decay factors, and external event feature vectors. The temporal alignment module outputs aligned temporal context. The temporal enhancement module performs temporal enhancement attention calculation. The deep extraction module extracts deep temporal features. The analysis module outputs sentiment curves and topic drift detection results. The visualization module generates a dynamic reputation evolution map and a visualization analysis report. This application effectively demonstrates the evolution of opinions in long film reviews.

Description

Technical Field

[0001] This invention relates to the field of film review analysis technology, and in particular to a system and method for extracting viewpoints and tracking their evolutionary paths. Background Technology

[0002] Currently, film review analysis typically treats each review as an isolated sample, relying on static textual features or contextual semantic representations for analysis, rarely considering the dynamic changes of film reviews over time. However, film reviews often exhibit a clear trajectory of opinion flow with key milestones such as film release, word-of-mouth development, and awards ceremonies. Audience evaluations of the same elements may systematically shift, making it difficult for analysis results to reflect the continuous evolution of audience opinions over time. Furthermore, compared to short film reviews, long film reviews are longer, contain more densely packed viewpoints, and have more complex layers of argumentation, often resulting in incomplete viewpoints and unclear evolutionary paths of opinions. Summary of the Invention

[0003] To facilitate the representation of the evolution of viewpoints in long film reviews, this application provides a viewpoint extraction and evolution path tracking system and method.

[0004] The viewpoint extraction and evolution path tracing method provided in this application adopts the following technical solution: A method for viewpoint extraction and evolution path tracing includes the following steps: Step S1: Collect long film review text data, time-series metadata, and external event data; Step S2: Preprocess the long film review text data to obtain the text feature vector; preprocess the temporal metadata to obtain the temporal decay factor; preprocess the external event data to obtain the external event feature vector; Step S3: Construct a time-series sequence based on timestamps, and extract the semantic feature vector of the long film review text within each time window to obtain the original time-series semantic sequence; Step S4: Output the fused and aligned temporal context; construct a temporal enhancement vector that fuses internal and external temporal information; Step S5: Use LSTM combined with attention mechanism and temporal convolutional network to perform deep temporal feature extraction; Step S6: Use the DaBERT model to process deep temporal features and output a continuous sentiment score curve. Calculate the evolution trend based on the sentiment score curve and mark the rising, falling and turning points of sentiment. Use the BERTrend model to process deep temporal features and output the topic distribution of each time window. Calculate the degree of change in the topic distribution of adjacent time windows. When the change exceeds a preset threshold, mark it as a topic drift point. Step S7: Integrate the sentiment curve analysis results and topic drift analysis results to output a dynamic word-of-mouth evolution map, and generate a visual analysis report based on the dynamic word-of-mouth evolution map.

[0005] Preferably, step S2 includes the following steps: For long film review text data, perform text cleaning, sentence segmentation, word segmentation, TF-IDF feature extraction, and domain sentiment dictionary weighting sequentially to obtain text feature vectors; For time-series metadata, perform outlier removal, time format standardization, and generate time-series decay factors sequentially to obtain time-series decay factors, which adopt an exponential decay form; For external event data, perform data verification, standardization, and feature engineering to obtain external event feature vectors.

[0006] Preferably, step S4 includes the following steps: Step S41: Use the dynamic time warping algorithm to perform nonlinear alignment of the multi-source time axis, find the optimal curved path, and flexibly align the time semantic sequence. Step S42: Based on the aligned temporal context, use canonical correlation analysis algorithm to analyze the correlation between internal temporal features and external event features; Step S43: Combine the time series sequence after dynamic time warping and alignment with the correlation weights obtained from canonical correlation analysis, and output the fused and aligned time series context; Step S44: Package and fuse the aligned temporal context and the original semantic features to construct a temporal enhancement vector that carries both internal and external temporal information.

[0007] Preferably, step S5 includes the following steps: Step S51: Perform basic attention calculation on the temporal augmentation vector sequence to capture semantic-level relevance; Step S52: Introduce a time decay factor on the basis of the basic attention weights to obtain the actual word weights; Step S53: Input semantic features and temporal features into the temporal gating unit to generate a gating value between 0 and 1, and dynamically balance the proportion of semantic information and temporal information. Step S54: Integrate the basic attention weights, time decay factor and temporal gating mechanism to output a temporal enhanced attention representation; Step S55: Use a long short-term memory network to capture long-term temporal dependencies, and overlay an attention mechanism to dynamically weight the temporal positions, highlighting the key time steps.

[0008] Step S56: Use a temporal convolutional network to process the original temporal sequence. Expand the receptive field through dilated convolution and ensure that information does not leak future time steps through causal convolution, so as to accurately capture local temporal patterns in the text.

[0009] Preferably, the attention calculation formula in step S51 is: ,in, For query vector, The key vector matrix, For value vectors, The dimension of the key vector.

[0010] Preferably, the formula for calculating the actual word weight in step S52 is: ,in, For the words whose weights are to be calculated, This is the timestamp of the word's appearance. For actual word weights, For the weight of the original words, This is the time decay factor.

[0011] The viewpoint extraction and evolution path tracing system provided in this application adopts the following technical solution: A system for extracting viewpoints and tracing evolutionary paths includes a data input module, a data processing module, a temporal alignment module, a temporal enhancement module, a deep extraction module, an analysis module, and a visualization module. The data input module is used to acquire long film review text data, temporal metadata, and external event data. The data processing module is used to output text feature vectors, temporal decay factors, and external event feature vectors. The temporal alignment module is used to output aligned temporal context. The temporal enhancement module is used to perform temporal enhancement attention calculation. The deep extraction module performs deep temporal feature extraction. The analysis module outputs sentiment curves and topic drift detection results. The visualization module is used to generate a dynamic reputation evolution map and a visualization analysis report.

[0012] Preferably, the data processing module includes a text processing module, a time-series processing module, and an event processing module. The text processing module is used to process long film review text data and output text feature vectors. The time-series processing module is used to process time-series metadata and generate time-series decay factors. The event processing module is used to process external event data and output external event feature vectors.

[0013] By adopting the above technical solutions, the entire process of extracting viewpoints and tracking evolution paths, from inputting multi-source heterogeneous data to generating dynamic word-of-mouth evolution maps, is realized. Relying on mechanisms such as temporal semantic extraction, internal and external temporal alignment, temporal enhanced attention, deep temporal modeling, and sentiment and theme analysis, the temporal evolution patterns of viewpoints in long film reviews are accurately captured, providing high-precision, intuitive and visual viewpoint analysis capabilities.

[0014] In summary, this application includes at least one of the following beneficial technical effects: 1. Integrating natural language processing, time series data analysis, and deep learning technologies to solve the problems of insufficient time series modeling capabilities and difficulty in tracing the evolution path of viewpoints in analysis; 2. By introducing dynamic time warping and canonical correlation analysis, we achieve deep integration of internal and external time series for alignment and correlation analysis, which solves the problem of misalignment of multi-source time series data and uncovers key external events that drive the evolution of viewpoints. 3. By designing a temporal enhanced attention mechanism and introducing a time decay factor and a temporal gating mechanism, a dynamic balance between semantic information and temporal information is achieved, enhancing the model's ability to understand the evolution of opinions in long film reviews; 4. A dual-channel architecture of LSTM + attention mechanism and temporal convolutional network is built to take into account both global long-term dependencies and local temporal patterns, providing a more comprehensive temporal representation for sentiment and topic analysis. 5. By integrating DaBERT temporal sentiment analysis with BERTrend temporal topic modeling, a dynamic word-of-mouth evolution map is constructed. Multidimensional visual variables are used to present the sentiment fluctuations, popularity changes, and topic drift of keywords in a three-dimensional way, making the analysis results easier to interpret. Attached Figure Description

[0015] Figure 1 This is a flowchart of a viewpoint extraction and evolution path tracing method in an embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating the timing context after output fusion and alignment in the embodiments of this application.

[0017] Figure 3 This is a flowchart illustrating the construction of a timing enhancement vector that integrates internal and external timing information in the embodiments of this application.

[0018] Figure 4 This is a flowchart illustrating the generation of a visual analysis report in the embodiments of this application.

[0019] Figure 5 This is an architecture diagram of a viewpoint extraction and evolution path tracking system in an embodiment of this application. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1-5 This application will be described in further detail.

[0021] This application discloses a viewpoint extraction and evolution path tracing system. (Refer to...) Figures 1 to 5The system comprises a data input module, a data processing module, a temporal alignment module, a temporal enhancement module, a deep extraction module, an analysis module, and a visualization module. The data processing module includes text processing, temporal processing, and event processing modules. Through the synergistic effect of these modules, it achieves end-to-end opinion extraction and evolutionary path tracking, from multi-source heterogeneous data input to the generation of a dynamic word-of-mouth evolution map. Relying on mechanisms such as temporal semantic extraction, internal and external temporal alignment, temporal enhancement attention, deep temporal modeling, and sentiment and theme analysis, it accurately captures the temporal evolution patterns of opinions in long film reviews, providing high-precision, intuitive, and visualized opinion analysis capabilities.

[0022] The data input module is used to input long film review text data, time-series metadata and external event data into the system, providing raw data support for subsequent opinion extraction and evolution analysis. The data input module supports batch import and real-time streaming access of various data formats.

[0023] The text processing module is used to perform sentence segmentation, word segmentation, text cleaning, TF-IDF feature extraction, and domain sentiment lexicon weighting on long film review text data, outputting text feature vectors. TF-IDF feature extraction sets an upper limit on vocabulary size and a document frequency threshold to filter out keywords with discriminative power; the domain sentiment lexicon weighting incorporates a film review sentiment lexicon to enhance the weight of matched sentiment words, effectively improving the expressive power of sentiment features.

[0024] The time-series processing module removes outliers from the time-series metadata, standardizes the time format of the time-series metadata, and generates a time-series decay factor. The time-series decay factor is calculated based on an exponential decay function, using the interval between the comment's publication time and the current time as input. An adjustable decay coefficient controls the rate at which the influence of historical information decays, providing a reasonable time weight for subsequent time-series modeling.

[0025] The event processing module performs data validation, standardization, and feature engineering on external event data, outputting external event feature vectors. Data validation includes format validity checks and numerical range checks, removing event records with obvious errors.

[0026] The temporal alignment module is used to perform temporal semantic extraction, dynamic time warping alignment, and canonical correlation analysis, and outputs the aligned temporal context.

[0027] The temporal enhancement module is used to construct temporal enhancement vectors, introducing a time decay factor and a temporal gating mechanism for temporal enhancement attention calculation. The construction of the temporal enhancement vector involves concatenating the aligned temporal context with the original semantic features to form an enhanced representation that integrates internal and external temporal information. The time decay factor uses an exponential decay function to dynamically suppress the attention weights of distant words based on the time difference between the word and the current time, allowing the model to focus on recently occurring semantic content. The temporal gating mechanism designs learnable gating units that take semantic and temporal features as input and generate a gating value between 0 and 1 using the sigmoid function, dynamically balancing the proportion of semantic and temporal information in attention calculation.

[0028] The deep extraction module uses LSTM with an attention mechanism, followed by a temporal convolutional network, to extract deep temporal features. LSTM is used to capture long-term temporal dependencies in the text sequence, dynamically weighting key time steps to amplify the impact of emotional turning points. Simultaneously, dilated and causal convolutions are used to exponentially expand the receptive field and accurately capture local temporal patterns in the text, providing a foundation for subsequent sentiment and topic analysis.

[0029] The analysis module is responsible for outputting sentiment curves by DaBERT and topic drift detection results by BERTrend.

[0030] The visualization module integrates sentiment curves and topic drift analysis results to generate a dynamic word-of-mouth evolution map and outputs a visualization analysis report.

[0031] This application discloses a method for viewpoint extraction and evolution path tracing. (Refer to...) Figures 1 to 5 This includes the following steps.

[0032] Step S1: Collect long film review text data, time-series metadata, and external event data. Long film review text, time-series metadata, and external event data are automatically crawled from online film platforms. A data type recognition function automatically distinguishes the three types of input data and directs them to the corresponding processing modules for parallel preprocessing. Long film review text data consists of detailed film review content published by users, uniformly stored as UTF-8 encoded text files; time-series metadata includes the publication time of each film review, stored in CSV format; external event data includes film release dates, film award information, creator updates, social media popularity indexes, etc., stored in JSON format.

[0033] Step S2: Preprocess the long film review text data to obtain the text feature vector; preprocess the time series metadata to obtain the time series decay factor; preprocess the external event data to obtain the external event feature vector.

[0034] The long film review text data was processed sequentially through text cleaning, sentence segmentation, word segmentation, TF-IDF feature extraction, and weighting with a domain sentiment dictionary to obtain text feature vectors. Text cleaning used regular expressions to remove special symbols, emoticons, and meaningless number strings, retaining core Chinese and English vocabulary. Word segmentation used the Jieba Chinese word segmentation tool, supplemented by a custom dictionary for the film and television industry, including words like "montage," "explosive acting," "plot twist," and "special effects" to avoid ambiguous segmentation and improve the accuracy of terminology recognition.

[0035] The time series metadata is processed sequentially, including outlier removal, time format standardization, and generation of time series decay factors. The time series decay factors adopt an exponential decay form, with greater decay occurring as the time difference increases.

[0036] External event feature vectors are obtained by performing data validation, standardization, and feature engineering on external event data.

[0037] Step S3: Construct a time-series sequence based on timestamps, and extract semantic feature vectors from the text within each time window. The time window is divided by day or week, and each time window may contain multiple film reviews. The average of the semantic vectors of all film reviews within the window is taken as the semantic representation of that window, resulting in the original time-series semantic sequence.

[0038] Step S4: Output the fused and aligned temporal context; construct a temporal enhancement vector that fuses internal and external temporal information. This includes the following steps.

[0039] Step S41: Use the dynamic time warping algorithm to nonlinearly align the multi-source timeline, find the optimal curved path, and flexibly align the temporal semantic sequences. For example, different film critics may discuss the same movie at completely different paces. By finding the optimal curved path, the temporal semantic sequences of different lengths and paces are flexibly aligned, smoothing out the time offset caused by individual differences.

[0040] Step S42: Based on the aligned temporal context, canonical correlation analysis (CCAD) is used to analyze the correlation between internal temporal features and external event features. CCAD identifies a projection direction that maximizes the correlation between internal temporal features and external event features, outputting correlation weights to help identify which external events have the greatest impact on the evolution of viewpoints. For example, when an actor wins an award, positive reviews of the actor's acting skills increase significantly; CCAD can quantify the strength of this association.

[0041] Step S43: Combine the time series sequence after dynamic time warping and alignment with the correlation weights obtained from canonical correlation analysis to output the fused and aligned time series context, providing high-quality input for subsequent analysis.

[0042] Step S44: Package and fuse the aligned temporal context and the original semantic features to construct a temporal enhancement vector that simultaneously carries internal and external temporal information. The fusion method uses feature concatenation, so that the representation of each time window simultaneously encompasses textual semantics, temporal alignment information, and external event association information.

[0043] Step S5: Deep temporal feature extraction is performed using LSTM combined with attention mechanism and temporal convolutional network, including the following steps.

[0044] Step S51: Perform basic attention calculation on the temporal augmentation vector sequence to capture semantic relevance. The basic attention method is to obtain the attention weight by taking the dot product of the query and the key, and then summing it with the value in a weighted manner, so that the model focuses its attention on the semantically relevant parts.

[0045] The formula for calculating attention is: ,in, For query vector, The key vector matrix, For value vectors, The dimension of the key vector.

[0046] Step S52: Introduce a time decay factor on top of the basic attention weights to obtain actual word weights, dynamically suppressing the weights of words that are far removed from the current time point. The time decay factor decays exponentially; the larger the time difference, the smaller the decay factor. Recent viewpoints have higher weights, while the weights of historical viewpoints decrease exponentially with the time interval, guiding the model to pay more attention to recent content.

[0047] The actual word weight calculation formula is: ,in, For the words whose weights are to be calculated, This is the timestamp of the word's appearance. For actual word weights, For the weight of the original words, This is the time decay factor.

[0048] Step S53: Input semantic features and temporal features into the temporal gating unit to generate a gating value between 0 and 1, dynamically balancing the proportion of semantic information and temporal information. When a viewpoint undergoes a sudden change, the gating value approaches 0, reducing the influence of historical viewpoints; when a viewpoint evolves stably, the gating value approaches 1, strengthening temporal continuity.

[0049] Step S54: Integrate the basic attention weights, time decay factor, and temporal gating mechanism to output a temporal enhanced attention representation. This approach considers both semantic relevance and temporal importance, enabling a more accurate reflection of the dynamic characteristics of temporal text.

[0050] Step S55: Use a Long Short-Term Memory (LSTM) network to capture long-term temporal dependencies and generate a hidden state sequence. An attention mechanism is superimposed on the output of the LSTM hidden layer to dynamically weight the temporal positions, highlighting key time steps such as emotional turning points and thematic abrupt changes.

[0051] Step S56: Use a temporal convolutional network to process the original temporal sequence. Expand the receptive field through dilated convolution and ensure that information does not leak future time steps through causal convolution. Accurately capture local temporal patterns in the text, such as phrase-level repetition, progression or transition structures.

[0052] Step S6: Process the deep temporal features using the DaBERT model to output a continuous sentiment score curve. Calculate the evolution trend based on the sentiment score curve, marking the periods of sentiment rise, decline, and turning points. Quantify the rate of sentiment change by calculating the rate of change of sentiment scores between adjacent time windows, and mark the moments of significant sentiment change.

[0053] DaBERT incorporates temporal information during the pre-training phase to understand the changing trends of sentiment over time, reflecting the sentiment tendency and fluctuation range of each time window. Sentiment scores range from 0 to 1, with higher scores indicating a more positive sentiment.

[0054] The BERTrend model is used to process deep temporal features, outputting the topic distribution for each time window. The degree of change in topic distribution between adjacent time windows is calculated, and when the change exceeds a preset threshold, it is marked as a topic drift point. Theme drift points reflect the shift in audience focus, such as from discussing "special effects" to discussing "plot depth".

[0055] Among them, BERTrend can dynamically track the rise, differentiation and disappearance of topics, and identify the core themes discussed in film reviews and their changes over time.

[0056] Step S7: Integrate the sentiment trend calculation results with the topic drift detection results to construct a dynamic word-of-mouth evolution map. The map unfolds horizontally along the time axis with keywords as nodes. Node colors distinguish positive, negative, and neutral sentiments, node size reflects keyword frequency, and node connections show topic inheritance or co-occurrence relationships. Drift points are marked with triggering events, such as awards or controversial events. A visual analysis report is generated based on the dynamic word-of-mouth evolution map, supporting interactive user exploration. Users can click on keywords to view detailed sentiment evolution curves, drag the time axis to zoom and observe evolution details at different time granularities, and filter keywords for specific sentiment tendencies or themes for focused analysis. The report supports exporting to PDF and HTML formats for easy sharing and archiving.

[0057] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for viewpoint extraction and evolution path tracing, characterized in that: Includes the following steps: Step S1: Collect long film review text data, time-series metadata, and external event data; Step S2: Preprocess the long film review text data to obtain text feature vectors; time series Metadata preprocessing yields the time-series decay factor; external event data preprocessing yields the external event feature vector. Step S3: Construct a time-series sequence based on timestamps, and extract the semantic feature vector of the long film review text within each time window to obtain the original time-series semantic sequence; Step S4: Output the fused and aligned timing context; Construct a temporal enhancement vector that integrates internal and external temporal information; Step S5: Use LSTM combined with attention mechanism and temporal convolutional network to perform deep temporal feature extraction; Step S6: Use the DaBERT model to process deep temporal features and output a continuous sentiment score curve. Calculate the evolution trend based on the sentiment score curve and mark the rising, falling and turning points of sentiment. Use the BERTrend model to process deep temporal features and output the topic distribution of each time window. Calculate the degree of change in the topic distribution of adjacent time windows. When the change exceeds a preset threshold, mark it as a topic drift point. Step S7: Integrate the sentiment curve analysis results and topic drift analysis results to output a dynamic word-of-mouth evolution map, and generate a visual analysis report based on the dynamic word-of-mouth evolution map.

2. The viewpoint extraction and evolution path tracing method according to claim 1, characterized in that: Step S2 includes the following steps: For long film review text data, perform text cleaning, sentence segmentation, word segmentation, TF-IDF feature extraction, and domain sentiment dictionary weighting sequentially to obtain text feature vectors; For time-series metadata, perform outlier removal, time format standardization, and generate time-series decay factors sequentially to obtain time-series decay factors, which adopt an exponential decay form; For external event data, perform data verification, standardization, and feature engineering to obtain external event feature vectors.

3. The viewpoint extraction and evolution path tracing method according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: Use the dynamic time warping algorithm to perform nonlinear alignment of the multi-source time axis, find the optimal curved path, and flexibly align the time semantic sequence. Step S42: Based on the aligned temporal context, use canonical correlation analysis algorithm to analyze the correlation between internal temporal features and external event features; Step S43: Combine the time series sequence after dynamic time warping and alignment with the correlation weights obtained from canonical correlation analysis, and output the fused and aligned time series context; Step S44: Package and fuse the aligned temporal context and the original semantic features to construct a temporal enhancement vector that carries both internal and external temporal information.

4. The viewpoint extraction and evolution path tracing method according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Perform basic attention calculation on the temporal augmentation vector sequence to capture semantic-level relevance; Step S52: Introduce a time decay factor on the basis of the basic attention weights to obtain the actual word weights; Step S53: Input semantic features and temporal features into the temporal gating unit to generate a gating value between 0 and 1, and dynamically balance the proportion of semantic information and temporal information. Step S54: Integrate the basic attention weights, time decay factor and temporal gating mechanism to output a temporal enhanced attention representation; Step S55: Use a long short-term memory network to capture long-term temporal dependencies, and overlay an attention mechanism to dynamically weight the temporal positions, highlighting the key time steps. Step S56: Use a temporal convolutional network to process the original temporal sequence. Expand the receptive field through dilated convolution and ensure that information does not leak future time steps through causal convolution, so as to accurately capture local temporal patterns in the text.

5. The viewpoint extraction and evolution path tracing method according to claim 4, characterized in that: The attention calculation formula in step S51 is as follows: ,in, For query vector, The key vector matrix, For value vectors, The dimension of the key vector.

6. The viewpoint extraction and evolution path tracing method according to claim 4, characterized in that: The formula for calculating the actual word weight in step S52 is as follows: ,in, For the words whose weights are to be calculated, This is the timestamp of the word's appearance. For actual word weights, For the weight of the original words, This is the time decay factor.

7. A viewpoint extraction and evolution path tracing system, said system being applied to the viewpoint extraction and evolution path tracing method as described in any one of claims 1 to 6, characterized in that: The system includes a data input module, a data processing module, a temporal alignment module, a temporal enhancement module, a deep extraction module, an analysis module, and a visualization module. The data input module is used to acquire long film review text data, temporal metadata, and external event data. The data processing module is used to output text feature vectors, temporal decay factors, and external event feature vectors. The temporal alignment module is used to output aligned temporal context. The temporal enhancement module is used to perform temporal enhancement attention calculation. The deep extraction module performs deep temporal feature extraction. The analysis module outputs sentiment curves and topic drift detection results. The visualization module is used to generate a dynamic word-of-mouth evolution map and a visualization analysis report.

8. The viewpoint extraction and evolution path tracking system according to claim 7, characterized in that: The data processing module includes a text processing module, a time-series processing module, and an event processing module. The text processing module is used to process long film review text data and output text feature vectors. The time-series processing module is used to process time-series metadata and generate time-series decay factors. The event processing module is used to process external event data and output external event feature vectors.

Images