System for visualizing and interpreting a multi-modal model for sentiment analysis

Abstract

The present invention relates to a visual analytics system for helping users better understand and diagnose multimodal models for sentiment analysis. By considering feature importance measured by post-hoc explainability techniques, the system produces explanations of intra- and inter-modality interactions learned by multimodal language models from three levels (i.e., global, subset, and local levels). At the global level, the present system extracts and visually summarizes three types of interactions (i.e., dominance, complementarity, and conflict) and adopts an augmented tree layout. At the subset level, the system summarizes influential and frequent multimodal features using compact templates and enables the exploration of features of interest to users using different symbolic designs. At the local level, the system visualizes individual multimodal instances and their explanations with details.

Description

Technical Field

[0001] This application relates to natural language processing (NLP) and sentiment computing, and more specifically, to a system and method for visualizing and interpreting multimodal models used for sentiment analysis. Background Technology

[0002] Multimodal sentiment analysis is a dynamic topic in natural language processing and sentiment computing, focusing primarily on automatically detecting human attitudes or opinions from multiple communication channels, such as language (i.e., text), speech, and facial expressions. A core challenge is modeling the complex intra- and intermodal interactions, where multimodal features are being fused.

[0003] In early work, features from different modalities were concatenated before being input into the learning model. Conversely, some works employed late-fusion methods, which combined decision values ​​from individual unimodal models using voting schemes or the learning model. However, these methods neglected cross-modal interactions. To address these issues, some works explicitly computed unimodal, bimodal, and trimodal features and fused them with tensor products and dynamic routing. Recently, neural network methods have been widely used to model complex interactions between modalities. For example, researchers have extended LSTM units and gates to learn temporal interaction patterns between multimodal sequences. Attention-based RNNs have been proposed to learn multimodal representations with cyclic translation loss across modalities. We design a multi-view gated memory cell controlled by a neural network to store and predict temporal cross-modal interactions. A transformer attention mechanism is also utilized to learn cross-modal alignment and interactions. While these neural networks significantly improve performance compared to traditional methods, their complex architecture severely impacts model interpretability. In this paper, we construct a visual analytics system to aid in the diagnosis of state-of-the-art black-box models for sentiment analysis tasks.

[0004] Post-hoc explainability techniques refer to separate explanations of a model's understanding after the model training process. They can be broadly categorized into two groups: model-specific approaches and model-agnostic approaches. Model-specific approaches provide explanations for specific types of models, ranging from shallow models (such as tree ensembles and support vector machines) to more complex neural networks. Conversely, model-agnostic approaches are flexible enough to be applied to any machine learning model. According to recent surveys, there are two main types of model-agnostic approaches: simplified explanations and feature-related explanations.

[0005] For techniques that fall under the category of simplified interpretations, researchers often build alternative models (e.g., rule-based learners, decision trees, and linear models) to mimic the behavior of the original model with reduced complexity. One of the most representative techniques is LIME, which constructs a locally linear model based on the neighbors of instances of interest to approximate individual predictions. Feature relevance interpretations quantify feature contributions by calculating relevance or importance scores. A popular example is SHAP, where the mathematical root value is the Shapley value, a method derived from cooperative game theory. SHAP calculates an additional importance score for each feature to describe its impact on a given prediction. It possesses desirable properties (local accuracy, missing values, and consistency) and has been shown to align with human intuition. Other works use local gradients, randomized feature permutations, or influence functions to expose features relevant to the model's predictions.

[0006] However, the methods described above are often used to explain specific instances of a single modality (e.g., sentences, images) and cannot be directly applied to multimodal sentiment analysis. This disclosure aims to fill this gap by enabling multi-level interpretation of intramodal and intermodal interactions learned from global, subset, and local levels.

[0007] As the complexity of both data and machine learning models increases, various visual analytics systems have been proposed to help understand model behavior. Beyond using computational metrics to measure model performance, users need to explore when and why the model makes specific decisions, enabling them to gain the model's trust, identify model limitations, and further refine model design. One of the most common and important interpretation strategies in previous work is revealing the relationship between input data and model predictions. These can be categorized into two groups: instance exploration and feature & subset exploration.

[0008] Instance visualization focuses on illustrating model behavior for individual data samples. ModelTracker is proposed to support performance debugging with a visual summary of binary classification instances. Related techniques extend performance visualization to multi-class scenarios with aligned vertical axis designs, and further employ a class matrix design for instance-level summaries. In addition to visualizing instance distribution, an exploratory debugging prototype is built to allow users to interpret corrections to the model. Furthermore, tools exist that allow users to interactively probe individual models using provided inputs.

[0009] Feature and subset visualization studies explored how to display modal groups of features and instances that influence model decisions. FeatreInsight was developed to support the feature ideation process with visual generalizations that have setting errors. Existing techniques allow for exploring the predictive power of feature candidates across different feature selection algorithms. For specific applications in CV and NLP, features are often visualized as image patches or text segments. Furthermore, researchers have developed interactive tools to facilitate exploration at the group level. Feature attributes are compared to examine differences between different subsets of data. Some works use fairness metrics to partition data into groups for model diagnostics.

[0010] However, these methods do not consider exploring multimodal features and determining the extent to which they influence model decisions. Our system facilitates the multifaceted exploration of multimodal features and generates multi-level visual interpretations of their influence. Summary of the Invention

[0011] This invention is a visual analysis system designed to help model developers and users understand and diagnose how machine learning models utilize multimodal information for sentiment analysis. While recent advancements in techniques to enhance the interpretability of machine learning models have been made, they have typically targeted single-modal scenarios (e.g., images, sentences) and have done little work in interpreting multimodal models. This system generates interpretations of intramodal and intermodal interactions learned by a multimodal language model from three levels: global, subset, and local. At the global level, the system utilizes an extended tree layout to extract and visually summarize three types of interactions (i.e., dominance, complementarity, and conflict). At the subset level, the system summarizes influential and frequent multimodal features using compact templates and is able to explore features of user interest from multiple perspectives using different symbolic designs. At the local level, the system utilizes details to visualize individual multimodal instances and their interpretations. Through expert interviews and two case studies on public opinion datasets, we demonstrate that our system can help users gain a deeper understanding of multimodal models in sentiment analysis.

[0012] On one hand, a visual analysis system is provided for visualizing and interpreting models used in sentiment analysis. This visual analysis system includes: a storage module configured to store the model and multimodal data including processed linguistic features, auditory features, and visual features; an interpretation engine module configured to mimic model behavior of the model using the processed linguistic, auditory, and visual features to generate multi-level interpretations of the model behavior, including global, subset, and local levels; and a visual analysis module configured to display the interpretation results of the multi-level interpretations of the model behavior in multiple views. The global level interpretation presents the importance of each modality among the linguistic, auditory, and visual modalities, as well as the interactions between the linguistic, auditory, and visual modalities. The subset level interpretation summarizes influential and frequent features using templates and can explore features of user interest using different symbols. The local level interpretation visualizes instances and their interpretations.

[0013] In some embodiments, the visual analysis module is further configured to: display the interpretation results of the multi-level interpretation of the model behavior in multiple views using a summary view, a template view, a projection view, and an instance view.

[0014] In some embodiments, the overview view is configured to present the global-level explanation and use a three-layer expanded tree layout to present the importance of each of the language modalities, auditory modalities, and visual modalities, as well as the interactions between the language modalities, the auditory modalities, and the visual modalities.

[0015] In some embodiments, the template view is configured to present an explanation of the subset level and to summarize the frequent and influential templates.

[0016] In some embodiments, the template view includes template type, support, importance, and prediction and error, such that the templates can be sorted according to support, importance, and error.

[0017] In some embodiments, the projected view is configured to: present an explanation of the subset level and connect the template in the template view to the instance, enabling the examination of detailed information about the features in the instance.

[0018] In some embodiments, in the projected view: the visual features are represented using a Chernov face map including yaw, pitch, and roll axis dimensions; the linguistic features are represented by words or phrases; the auditory features are represented by an inner circle and a plurality of sectors surrounding the inner circle, the plurality of sectors respectively indicating the pitch, the amplitude, the guttural sound, and the phase.

[0019] In some embodiments, the instance view is configured to: present the explanation at the local level and provide a local explanation of the importance of each of the language modalities, auditory modalities, and visual modalities and their features by visualizing the features of the instance.

[0020] In some embodiments, the instance view also provides video scenarios for instance-level exploration.

[0021] In some embodiments, the interactions include dominance, complementarity, and conflict. Dominance indicates that one modality dominates the polarity of sentiment prediction; complementarity indicates that two or three modalities influence the model prediction in the same direction; and conflict indicates that modalities influence the model prediction differently.

[0022] In some embodiments, the interpretation engine module is further configured to identify the interactions between the language modalities, auditory modalities, and visual modalities based on processed language features, auditory features, and visual features.

[0023] In some embodiments, the visual analysis system further includes a template construction module configured to: construct a feature set based on the language features, the auditory features, and the visual features; and construct templates for frequently occurring features or features that have a strong impact on model prediction, as frequent and influential templates in the template view.

[0024] In some embodiments, the multi-level interpretation further includes an interpretation of the interactions within each of the language modality, the auditory modality, and the visual modality.

[0025] In some embodiments, the interpretation engine uses the SHAP algorithm to calculate the importance of each of the language modalities, the auditory modalities, and the visual modalities, as well as the interactions between the language modalities, the auditory modalities, and the visual modalities.

[0026] On the other hand, a visual analysis method is provided for visualizing and interpreting models used in sentiment analysis. The visual analysis method includes: storing the model and multimodal data including processed linguistic features, auditory features, and visual features; using the processed linguistic features, auditory features, and visual features to mimic the model's behavior to generate a multi-level explanation of the model's behavior, including global, subset, and local levels; and displaying the explanation results of the multi-level explanations of the model's behavior in multiple views, wherein the global level explanation presents the importance of each modality among the linguistic, auditory, and visual modalities and the interactions between the linguistic, auditory, and visual modalities; the subset level explanation summarizes influential and frequent features using templates and can explore features of user interest using different symbols; and the local level explanation can visualize instances and their explanations. Attached Figure Description

[0027] Figure 1 The system framework is shown to mainly consist of three modules: storage module, interpretation engine, and visual analysis interface.

[0028] Figure 2 The system's visual analysis interface consists of five views. The user panel displays descriptive statistics about the model and dataset. The overview view uses a three-level expanded tree layout to present the importance of each modality and their interactions. The template view and projection view complement each other for subset-level interpretation. Specifically, the template view summarizes the frequent and influential templates for the feature set in the table. The projection view supports multifaceted exploration of instances with features of interest. The instance view provides local interpretation by visualizing the important features and context of each instance.

[0029] Figure 3 The design options are shown in the overview view. A: Expanded Sankey diagram. B: Current expanded tree layout design.

[0030] Figure 4 The symbol designs are shown in the projected view. A: Chernov's face symbol design. A dark ring and bold strokes on the left represent strong facial movements, while one on the right represents very subtle facial movements. B: Auditory symbol design. A large blue sector on the left represents high frequencies, and one on the right represents low frequencies.

[0031] Figure 5 Examples of double negatives are shown. "Not...sin" (in A) and "Not...bad" (in B) are considered indicators of negative emotion in the model. However, these phrases reflect slightly positive feelings.

[0032] Figure 6The co-occurrence pattern of "happiness + sadness" is shown. A: "Happiness + sadness" is a common and important feature template in the table. B: Corresponding illustrations for the original video information and three representative instances of the "happiness + sadness" template.

[0033] Figure 7 The negative effects of speech pitch are shown. A: "Pitch" is the most common auditory template, and it always has a negative effect (as shown by the points in the swarm diagram). B: Selection group of instances with large spacing values ​​(as indicated by the large radius of the blue sector). C: Two high-error cases where the model captures the inflection point of pitch but incorrectly associates pitch with negative effects. Detailed Implementation

[0034] Advantages or improvements of the present invention compared to the prior art

[0035] Existing techniques for sentiment analysis tasks primarily focus on improving the performance of multimodal models. However, these models often operate like black boxes, hindering users from understanding the underlying model mechanisms and fully trusting them when making decisions. This invention includes a visual analytics system to help users diagnose state-of-the-art black-box models used for sentiment analysis tasks.

[0036] Existing post-hoc explanation techniques are typically used to explain specific instances of a pattern (e.g., sentences, images) and cannot be directly applied to multimodal sentiment analysis. This invention fills this gap by enabling multi-level explanations of intramodal and intermodal interactions learned at global, subset, and local levels.

[0037] Existing systems for interpreting machine learning models do not consider exploring multimodal features and determining the extent to which they influence model decisions. The system of this invention facilitates multifaceted exploration of multimodal features and generates multi-level visual explanations of their impact.

[0038] System Overview

[0039] This invention comprises three components ( Figure 1The system comprises a storage module, an interpretation engine module, and a visual analysis module. The storage module stores the user model and multimodal data with processed features. Given the model and data, the interpretation engine uses feature attribute methods (e.g., SHAP) to calculate feature importance. Multilevel interpretations of model behavior are then generated. The visual analysis module supports interactive exploration of the interpretation results using five main views. The user panel is the entry point to the entire interface, displaying descriptive statistics about model performance and the dataset. The summary view, template view, projection view, and instance view then provide multilevel interpretations of model behavior based on language modality, visual modality, and auditory modality. The summary view presents a global summary of the influence of each modality and its interaction with sentiment prediction. The template view and projection view complement each other for subset-level interpretation. Specifically, the template view uses templates to summarize the set of features that frequently and significantly contribute to model predictions. The projection view supports multifaceted exploration of instances with features of interest and their prediction errors. The instance view summarizes instance-level prediction information (e.g., errors) and provides local interpretations of the importance of each modality and its features. Furthermore, auditory and visual features are labeled along the stated words, and feature annotations are provided for the corresponding original video clips for further exploration.

[0040] Multilevel interpretation

[0041] This system generates explanations at both the global and subset levels to facilitate a global understanding of multimodal behavior. These complement the local explanations computed by feature attribute methods (e.g., SHAP).

[0042] Global explanation

[0043] Since intramodal and intermodal interactions are central to multimodal sentiment analysis, they are crucial for users to understand how multimodal models utilize information from different modalities (i.e., verbal (T), auditory (A), and visual (V)). This system characterizes three typical types of interactions between modalities—dominance, complement, and conflict. Dominance indicates that the influence of one modality dominates the polarity of the sentiment prediction (i.e., positive or negative). Complement indicates that two or all three modalities influence the model's prediction in the same direction (i.e., positive or negative). Conversely, conflict indicates that the modalities influence the model's prediction differently from each other. Based on these definitions, we developed a set of rules (Algorithm 1) to identify them. Specifically, the influence of interactions on the model output is based on the importance of each modality (Ii). l I a I vThis is the sum of the importance of all features. We then extract and summarize the interactions (L) that have a strong influence on all predictions. The threshold for our rule is determined by maximizing the distance between interaction types while minimizing the average influence of interactions that are not dominant, complementary, or conflicting (i.e., other):

[0044]

[0045] Among them, L i (i∈{dominance, conflict, complete, others}) is the interaction type output by algorithm (1) for all instances, and dist is the average influence L. i and average influence L j The Euclidean distance between them.

[0046] Algorithm (1): Rules for extracting important relationships of each modality.

[0047] enter:

[0048] {I l ,I a ,I v};Th sig ,Th dom ,Th confl (∈(0,1));

[0049] Output:

[0050] The tag for the interaction type, l;

[0051]

[0052] Feature template

[0053] Feature templates summarize the influence of a set of multimodal features, helping users develop mental models of model decisions. For example, what types of words (e.g., adjectives) are considered important indicators of positive sentiment. To facilitate the exploration of the influence of high-dimensional features, the system organizes low-level features into groups and uses templates to summarize frequent and influential groups (…). Figure 2 C).

[0054] To facilitate the understanding of model behavior, the system constructs several feature sets based on sentence structure in the language modality, emotion-related features in the auditory modality, and facial expressions in the visual modality:

[0055] • Language: Part of Speech (POS) (e.g., noun, adjective, verb);

[0056] • Hearing: pitch, amplitude, glottal / speech quality, and phase;

[0057] • Vision (i.e., the face): various parts of the face (i.e., eyebrows, eyes, nose, lips, and chin), head movements, and facial expressions.

[0058] For language modalities, part-of-speech features provide a compact overview of the structure of language use. Auditory features are grouped according to the state-of-the-art speech processing framework COVAREP. These sets typically involve the emotion and pitch of speech. For facial-related features, grouped according to facial parts, head movements, and facial emotions, they are representative components used in Facial Action Coding Systems (FACS) to describe facial expressions.

[0059] After grouping the low-level features for each modality, the system constructs templates for both frequent feature sets (e.g., "ADJ") and features that have a strong impact on model predictions (e.g., the word "good"). Specifically, we create itemsets for important features and feature sets for all predictions. Then, we build FP-trees in the itemsets to find frequent patterns. For example, if "PRON" and "PART" or the word "NOT" appear frequently, they are recorded in the template. Figure 2 C).

[0060] user interface

[0061] Based on the generated interpretations, the system's user interface facilitates multi-level exploration of model behavior from linguistic, auditory, and visual modal perspectives. All views are tightly integrated with interactions to ensure smooth transitions between different levels of interpretation. They share the same color coding scheme, where a first color (e.g., deep red) signifies strong positive emotions, while a second color (e.g., deep blue) represents strong negative emotions.

[0062] Summary view

[0063] The overview view presents the user panel ( Figure 2 A) provides an overview of the intra-modal and inter-modal interactions learned by the selected model. The effects of each mode and their interactions are visualized in a three-layer extended tree layout. Figure 3 B).

[0064] Visual design: In the parent node, the barcode Figure 1 and lines Figure 2 The distribution of ground truth and model prediction error are shown separately. The vertical height or length of barcode 1 represents the total number of instances, and the color indicates sentiment. Meanwhile, the lines... Figure 2The horizontal position or amplitude represents the absolute error, and the average error is represented by a dashed line.

[0065] The second layer illustrates the importance of individual patterns within the bee swarm plots. They are arranged in descending order of the influence of their respective patterns. For each node in this layer, red points are to the right of coordinate 0.0, and blue points are to the left of coordinate 0.0; the horizontal length of these blue points summarizes the total influence of the pattern. Furthermore, the bee swarm... Figure 3 The points in the graph and their projections (i.e., the barcode below) illustrate the distribution of the pattern's influence across all instances. (Bee colony) Figure 3 The color and horizontal position of the dots in the graph encode the importance value, while the two gray lines of 4 represent the magnitude of the average absolute importance.

[0066] The final layer summarizes information about the four types of interactions, with the most important one shown at the top. For each interaction, the horizontal range of all its graphs indicates the number of instances in that group. To better represent how combinations of modes affect the modes predicted by the model, data instances are grouped closer together if all three modalities of a data instance share similar influence modalities. Specifically, similarity is measured by the furthest distance between the three modalities of each instance. Then, the top line... Figure 5 and barcode Figure 6 The errors and prediction modes similar to the parent node are summarized. Additionally, three barcode charts are appended below to present the distribution of importance for all three modes. The vertical order of the three barcodes shows the total influence of the corresponding mode, which is obtained by adding the blue bars on the left. The color of each bar within the barcode represents its importance value.

[0067] Furthermore, links are drawn between two adjacent layers, from the parent node to its child nodes. The width of the link is proportional to the importance of the child node to the model's predictions.

[0068] Design options: We have considered alternative designs based on Sankey diagrams. Figure 3 A) is used to reveal intramodal and intermodal interactions and their importance to prediction. It consists of three parts: ground truth information on the left, the influence of independent modes in the center, and intermodal interactions on the right. The width of the flow is proportional to the importance of the target node of that flow. A barcode graph for each node further shows the distribution of importance. Additionally, the yellow line 2 of the node shows the error distribution to guide exploration. However, an expert suggested that more detailed information must be shown at each node. For example, which mode dominates the prediction? What is its frequency? Therefore, we added nodes to the graph and further transformed the Sankey graph into a compact tree layout, which led to the current design ( Figure 3 B).

[0069] Template View

[0070] To facilitate the exploration of feature sets and their effects, template views ( Figure 2 C) The table summarizes the frequent and influential templates for multimodal features.

[0071] Visual Design: The template view features four columns of descriptive information: template type, support, importance, and prediction and error. The first column lists the name of the default feature set. If a feature set contains frequent, important features, a green bar will be placed on the right, indicating the number of children of that feature set. Users can click on... Collapse the corresponding rows to obtain details. The second column shows the frequency of the templates. The importance of the templates and the distribution of predictions and errors are visualized in the third and fourth columns. They share the same visual representation as the overview view. Users can sort the templates based on their support, importance, and error. In this way, they can prioritize tasks that diagnose the behavior of complex models.

[0072] Projected view

[0073] To further support subset-level exploration of model behavior, projected views ( Figure 2 D) Connect the multimodal feature templates in the template view to instances. This allows users to examine detailed information about the features on each instance (e.g., feature values, prediction errors). For example, after a user selects the "ADJ" template in the template view, they might be interested in the feature values ​​(i.e., adjectives) associated with large errors or positive predictions. They would then need to further examine the individual instances.

[0074] Visual Design. To summarize the feature set of a set of instances, t-SNE is used to project the high-dimensional features onto a two-dimensional (2D) plane. Therefore, instances with similar features will be placed close together. Assuming that textual, auditory, and visual features are heterogeneous, three different notations are designed to encode the feature set of each instance. Users can switch between views to see the feature distribution for each modality. Furthermore, to aid in diagnosing model behavior (e.g., errors), heatmaps are added as backgrounds to display the distribution of prediction errors or template importance.

[0075] • Language: Since words or phrases already carry semantic meaning, we use them to represent text features. Additionally, we add a circle to each word, and the color of the circle encodes sentiment prediction.

[0076] • Visual: Our symbol design for facial features ( Figure 4A) Inspired by the Chernoff face map, which is popular for displaying facial expressions, the original Chernoff face map cannot reproduce information such as head movement. Therefore, we add three bars around the face to indicate head movement on the yaw, pitch, and roll axes, respectively. The outer ring encodes information about the entire face (e.g., emotion), with darker colors representing larger feature values. Furthermore, facial features (e.g., the nose) and the stroke width of the bars indicate the intensity of the movement. Emotion prediction is revealed by the background color of the face.

[0077] • Hearing: To help understand auditory features, we group them into higher-level categories. For example... Figure 4 As shown in B, each colored sector represents a feature of a category (pitch, phase, amplitude, guttural), where the radius is related to the feature value. The preceding sector 7 summarizes the average of the normalized features, while the smaller following sector 8 displays detailed feature values ​​for each category. Additionally, the inner circle color indicates sentiment prediction.

[0078] Instance View

[0079] Instance View ( Figure 2 E) Provide localized explanations by visualizing the key multimodal features and contexts of each instance (i.e., transcripts and videos).

[0080] Visual Design: The left column presents a visual summary of the impact of each modality on the model's predictions and the prediction errors. Users can sort instances based on different criteria (e.g., error) in the header and prioritize work in instance-level exploration. In each row, the horizontal axis 14 represents the sentiment range, marked with predictions and ground truth. Between the two values, a thick red line 9 represents the error. Below the prediction marker, three colored rectangles represent the aggregated feature importance value for the three modalities. The length and color of each rectangle encode the magnitude and sign of the importance. For example, modalities with a negative impact on predictions are encoded by blue rectangles and placed on the right. Additionally, the feature table below allows users to categorize and retrieve the importance values ​​of features or modalities.

[0081] To facilitate a global understanding of the context of each instance, the right column highlights the important features of the instance. Unlike intuitive text, auditory and visual features are more difficult to identify. Therefore, they are associated with the spoken words, and the most important words are highlighted. Line 10 above the word corresponds to auditory features, while line 11 below represents visual features. The vertical offset of the lines represents feature values, so fluctuations indicate feature changes. Furthermore, the background of the text or feature lines reflects the importance of multimodal features at the word level.

[0082] The instance view also provides video scenarios for instance-level exploration. When a user clicks on a row in the table, the corresponding video clip pops up and plays. To make visual features more intuitive, facial features that are ranked first (based on importance) are highlighted with bounding boxes covering the corresponding parts of the face. Users can also identify facial action units and their specific meanings in detail by hovering over the boxes.

[0083] User Interaction

[0084] The system offers a rich set of interactions to help unify different views and provides detailed information on demand, facilitating multi-level and multi-faceted exploration.

[0085] Scan. Users can scan the barcode in the last layer of the summary view to issue a query for a specific data instance of an interaction type. The template view and instance view will then display the relevant template and partial explanation, respectively.

[0086] Click. Many interactions within the system can be triggered and canceled by clicking. For example, clicking a table row in the template view will filter out irrelevant instances in the projection view and instance view. Users can switch between different modal feature projections by clicking radio buttons in the projection view. When a table row in the instance view is clicked, the corresponding instance in the projection view will be displayed, and its video clip will pop up and play. Additionally, users can click the titles of the template view and projection view to undo previous selections.

[0087] Lasso and semantic zooming. To facilitate scalable exploration, users can use lasso or semantic zooming to focus on a specific instance of interest in the projected view. Details are then displayed in the instance view.

[0088] Search, sort, and filter. To narrow down the exploration space, users can sort and search for instances or features in the template view and instance view tables. By adjusting the sliders in the projected view, users can filter instances based on sentiment predictions for specific modalities and feature importance.

[0089] Detailed example

[0090] This invention has used two multimodal models and tested them on a public opinion dataset (i.e., CMU-MOSEI). The data, experimental setup, and experiments are described below.

[0091] Data preprocessing

[0092] Without loss of generality, our invention was tested on CMU-MOSEI, the largest and most widely used benchmark dataset for multimodal sentiment analysis. It consists of 23,454 monologue film review video clips from 1,000 speakers and 250 topics from YouTube. The sentiment of each video clip was labeled with three annotators using a Likert scale of [-3, 3], where 3 represents strong positive sentiment and -3 represents strong negative sentiment. 0 indicates the review was labeled neutral. In addition to the sentiment labels, each video was associated with information from three communication channels: transcription of linguistic resources (l), visual facial expressions (v), and the speaker's voice as an auditory modality (a).

[0093] The system then transforms the raw data into computational features. For linguistic features, the transcripts are encoded using high-dimensional word vectors. We utilize word embedding techniques (e.g., glove embedding) to represent each word, where each word is transformed into a high-dimensional vector. For the visual modality, most work focuses on facial expressions, which are typically encoded using a Facial Action Coding System (FACS). The FACS utilizes 35 facial action units to encode facial muscle movements. It is used to extract facial features at the frame level. Auditory features are designed using the speech processing framework COVAREP. The extracted features have 74 dimensions, and all of them are related to speech emotion and pitch. To help users gain a quick overview of these basic features, the system further groups them into different classes, as described in section 5.3.2.

[0094] Case 1

[0095] In the first case, Expert 1 used the CMU-MOSEI dataset to study and diagnose a state-of-the-art model for sentiment analysis—the Multimodal Transformer (MULT). Following the setup of the aforementioned work, we trained, validated, and evaluated MULT using the same data splits (training: 16265, validation: 1869, test: 4643).

[0096] During the exploration, Expert 1 observed that the language modality frequently dominated predictions, and the model did not handle negation well in sentiment analysis. He further investigated the dominance of the visual modality, where "happiness" and "sadness" (two facial emotions) often appeared simultaneously. This was attributed to strong facial muscle movements, which were also captured by the model.

[0097] A. Dominance of language modality

[0098] Global Summary: After selecting MULT and the effective set in the user panel, Expert 1 felt interested in how the individual modalities and their interactions contribute to the model's predictions. This was achieved by viewing the second layer of the summary view (…). Figure 2 B), Expert 1 discovered language modalities ( Figure 2 In the B modality (represented by the letter "L"), the influence is greatest among the three modalities because it has the longest left-facing bar and the widest range of points in the swarm graph. Conversely, in the auditory modality located at the bottom ( Figure 2 In section B (represented by the letter "A"), the influence is minimal. Then, Expert 1 examines the final layer, where the dominant group with the widest barcode pattern is shown at the top. In this group, he finds the longest bar attached to the language modality, and the predicted barcode color is very consistent with the language barcode color. Therefore, Expert 1 concludes that the language also plays a dominant role in the dominant relationship. Furthermore, he notices a dense set of blue bars at the end (i.e., the right end) of the language barcode, where the error is relatively large (as shown by the yellow curve 5 above the dashed line). He wants to know what features or combinations thereof lead to the high error. Therefore, we brush the corresponding areas of the blue bars.

[0099] Subset Exploration: Template View Figure 2 C) lists all the frequent and important feature templates used for the instances scanned in the summary view. Sort them in descending order of error, Expert 1 finds "PRON+PART" at the top, with one sub-feature. He then collapses this row and finds 21 instances containing the word "not," which negatively impacts the prediction (blue dots in the beehive plot are shown in the "Importance" column). Next, he clicks on "not" to see details about the feature in the projection view. Zooming in on the word "not," he observes several similar negative words (e.g., "isn't," "wouldn't"). Figure 2 D). They are all located in the red area 12, indicating a large error. Expert 1 speculates that the model does not handle negation well. He will then lasso these words to carefully examine the corresponding instances in the instance view.

[0100] Case Study: To further evaluate how the model handles negation, Expert 1 begins with examples in the table that have large errors. Figure 2E). When examining the examples listed above, Expert 1 observed that negation had a significant negative impact on the overall prediction, and the model could not explain the true sentiment. For example, Expert 1 found that the language modality dominated the negative sentiment prediction, highlighting the case of the word "not" ( Figure 2 E). However, the true sentiment of the sentence is positive, with the opening phrase "Irealy like" demonstrating a positive attitude. However, the model failed to extract keywords and relied on negation (i.e., "not") to predict negative sentiment. Furthermore, Expert 1 noted that when double negations appear in the sentence ( Figure 5 The model tends to treat these negatives separately and view them both as signs of negative emotion. However, in reality, these double negatives reflect mild positive emotion.

[0101] B. Dominance of the visual modality (represented by V)

[0102] Global Summary: Expert 1 refers back to the "Dominant" group in the summary view, where many red bars from the predicted barcode are compared with bars from the visual modality (in... Figure 2 Consistent with (highlighted in red in B). The visual modality dominated the prediction, and the error bar graph above showed a low error rate compared to the previous case A. Driven by this observation, Expert 1 highlighted the red bars to investigate patterns within the visual features.

[0103] Subset exploration: In the template view, "facial emotion" has the greatest support ( Figure 6 A). After expanding the line, Expert 1 discovered that "happiness + sadness" is a frequent and significant combination. This inspired him to explore how a pair of opposite emotions can occur simultaneously. After clicking the template, the corresponding symbol is highlighted in the projection view ( Figure 6 B). The vast majority of them were found outside the red area, which confirmed that instances with "joy + sadness" usually have small prediction errors. He decided to examine these instances.

[0104] Instance Exploration: By browsing instances and their videos in the instance view, "joy" and "sadness" are generally considered important visual features with positive impact. Additionally, Expert 1 found that their co-occurrence might be due to the strong and rich facial expressions present in the videos. These expressions typically involve the movement of related facial motor units during "joy" and "sadness." For example, after Expert 1 clicked on an instance, he noticed the corresponding symbol in the projected view ( Figure 6B) All facial features (i.e., nose, eyes, eyebrows, mouth, and chin) have thick strokes, indicating strong movement. When he watched the original video, the boxes 13 for "joy" and "sadness" always popped up as important visual features. By hovering over box 13 and examining the facial expressions and their interpretations, Expert 1 concluded that extreme facial expressions triggered movement of the motion unit during "joy" and "sadness," and that the model seemed to capture these important visual facial expressions.

[0105] Case 2

[0106] In this case, Expert 2 explored the popular RNN-based model EF-LSTM for multimodal sentiment analysis using the CMU-MOSEI dataset. The dataset setup and feature processing were the same as in Case 1 (see section 5.5.2).

[0107] Through interactive exploration with the system, Expert 2 was surprised to find that EF-LSTM failed to learn the emotions in the text. Furthermore, he noted that among the three modalities, the auditory modality had the greatest impact on emotion prediction, while the pitch of the speech always had a negative effect on emotion prediction.

[0108] A. No meaningful information was learned from the text.

[0109] Global Summary: After selecting the effective set and EF-LSTM, Expert 2 starts with a summary view to gain a global understanding of each modality. Figure 3 B). By comparing the range of points in the three beehive diagrams, he was surprised to find that the auditory modality had the greatest influence, followed by the linguistic modality. Furthermore, the linguistic modality consistently exhibited a positive influence on emotion. These findings were exactly the opposite of intuition. Therefore, Expert 2 decided to first explore text-related interactions by tracing the thickest links from the linguistic modality to the third layer. He noticed that the "complementary" group appeared at the top, and the text played a major role within that group. He then swiped through the entire group to see patterns in the textual features.

[0110] Subset Exploration: Strangely, neither the text template nor the text symbols were seen in the template view or the projected view, respectively. Expert 2 suspected that the model had not learned any important linguistic features (i.e., words) for sentiment analysis. He then referred to the instance view to verify his suspicions.

[0111] Instance Exploration: While exploring instances in the instance view, Expert 2 found that the model failed to recognize words that might be important for sentiment analysis, such as "fantasy" (in row #1) and "excellent" (in row #2). None of them were highlighted in color in the instance details. Furthermore, Expert 2 noted that each word in each sentence across the feature table had an equally low positive importance score (less than 0.1). This explains why language modalities always have a positive impact and further demonstrates that the model fails to capture sentiment in the text.

[0112] B. The negative effects of speech pitch

[0113] Global Summary: Expert 2 focuses on the summary view ( Figure 3 The most influential modality in B) is the auditory modality, which shows a negatively skewed distribution of points. Furthermore, he noted that within the "conflict" group, auditory modality A had a negative effect throughout the time frame (the barcode for auditory modality A is predominantly blue). Therefore, Expert 2 scanned this group to study the negative impact of auditory features.

[0114] Subset Exploration: Expert 2 discovers "pitch" is the most frequent auditory template in the template view ( Figure 7 A). Furthermore, Expert 2 noted that in cases of negatively skewed distribution of points in the third column, spacing always has a negative impact. After clicking on that row, he switched to the projection view to examine the pitch value distribution ( Figure 7 B). He found that the auditory symbols unfolded along a left diagonal line, where the radius of the blue sectors (i.e., the spacing values) generally increased from left to right. He then selected a set of instances with large spacing at the right corner for further examination.

[0115] Exploring Examples: Browse the examples and videos in the Example View ( Figure 7 C), Expert 2 observed that pitch is always the most important auditory feature and is associated with negative effects. Although some important pitch change signals in the video were captured by the model, he believed the model was unreliable because it always treated pitch as an indicator of strong negative emotions, and he found many counterexamples. For example, in two cases ( Figure 7 C) He found that pitch ranked first in the feature table according to its negative importance. He also noticed that the orange line above the word (i.e., the pitch distance value) was highlighted in light blue (i.e., negative). By examining the offset of all the orange lines, he concluded that the light blue highlighted orange line appeared to be the inflection point of the pitch value. He speculated that the model captured an important signal in hearing. He further examined the original video and verified his observations. However, the speaker's pronunciation was enthusiastic, and the pitch should have reflected positive emotion.

[0116] Key features or core ideas of this invention

[0117] A novel visual analytics system provides multi-level and multi-faceted interpretations of intra- and inter-modal interactions learned by multimodal models. At the global level, the system extracts and visually summarizes three types of interactions—dominance, complementarity, and conflict—utilizing an extended tree layout. At the subset level, the system summarizes influential and frequent multimodal features using compact templates, enabling the exploration of features of user interest from multiple perspectives using different symbolic designs. At the local level, the system utilizes details to visualize individual multimodal instances and their interpretations.

[0118] • A method for characterizing and extracting three typical interactions (dominance, complementarity, and conflict) in visual, auditory, and linguistic modalities learned by a model.

[0119] Our invention provides interpretation of any multimodal model used for sentiment analysis, given pre-computed importance scores of input features from facial, auditory, and linguistic modalities.

Claims

1. A visual analysis system for visualizing and interpreting models used in sentiment analysis, comprising: A storage module configured to store the model and multimodal data including processed language features, auditory features, and visual features; An interpretation engine module is configured to: calculate the importance of each of the language modalities, auditory modalities, and visual modalities, as well as the interactions between the language modalities, the auditory modalities, and the visual modalities, to generate multi-level interpretations including global, subset, and local levels; as well as A visual analysis module is configured to display the interpretation results of the multi-level interpretation in multiple views, wherein... The global-level explanation reveals the importance of each of the language, auditory, and visual modalities, as well as the interactions among them. The subset-level interpretation utilizes templates to summarize influential and frequent features, and can use different symbols to explore features of interest to the user. The local level explanation can visualize instances and their explanations.

2. The visual analysis system according to claim 1, wherein, The visual analysis module is further configured to display the interpretation results of the multi-level interpretation using a summary view, a template view, a projection view, and an instance view.

3. The visual analysis system according to claim 2, wherein... The overview view is configured to present the global-level explanation and use a three-layer expanded tree layout to present the importance of each modality among the language modality, auditory modality, and visual modality, as well as the interactions between the language modality, the auditory modality, and the visual modality.

4. The visual analysis system according to claim 2, wherein The template view is configured to present an explanation at the subset level and to summarize the frequent and influential templates.

5. The visual analysis system according to claim 4, wherein The template view includes template type, support, importance, prediction, and error, enabling the templates to be sorted according to support, importance, and error.

6. The visual analysis system according to claim 2, wherein The projected view is configured to present an explanation of the subset level and to connect the template in the template view to the instance, enabling the examination of detailed information about the features in the instance.

7. The visual analysis system of claim 6, wherein in the projected view: The visual features are represented using a Chernov facet diagram that includes yaw, pitch, and roll axis dimensions; The linguistic features are represented by words or phrases; The auditory features are represented by an inner circle and multiple sectors surrounding the inner circle, the multiple sectors indicating pitch, amplitude, guttural sounds, and phase, respectively.

8. The visual analysis system according to claim 2, wherein The instance view is configured to present the explanation at the local level and to provide a local explanation of the importance of each of the language modalities, auditory modalities, and visual modalities and their features by visualizing the features of the instance.

9. The visual analysis system according to claim 8, wherein The instance view also provides video scenarios for instance-level exploration.

10. The visual analysis system according to claim 1, wherein, The interactions include dominance, complementarity, and conflict, among which Dominant representation: One modality dominates the polarity of sentiment prediction. The complementarity means that two or three modalities have the same direction of influence on the model's predictions. The conflict is indicated by the fact that different modes have different effects on model predictions.

11. The visual analysis system according to claim 1, wherein The interpretation engine module is further configured to identify the interactions between the language modality, auditory modality, and visual modality based on the processed language features, auditory features, and visual features.

12. The visual analysis system according to claim 4, further comprising: The template constructor module is configured as follows: A feature set is constructed based on the language features, the auditory features, and the visual features; Templates are constructed for frequently occurring features or features that have a strong impact on model predictions, serving as frequent and influential templates in the template view.

13. The visual analysis system according to claim 1, wherein The multi-level interpretation also includes the interpretation of the interactions within each of the language modality, the auditory modality, and the visual modality.

14. The visual analysis system according to claim 1, wherein The interpretation engine uses the SHAP algorithm to calculate the importance of each modality in the language modality, the auditory modality, and the visual modality, as well as the interactions between the language modality, the auditory modality, and the visual modality.

15. A visual analysis method for visualizing and interpreting models used in sentiment analysis, comprising: Store the model and multimodal data including processed language features, auditory features, and visual features; The importance of each of the language, auditory, and visual modalities, as well as the interactions between the language, auditory, and visual modalities, are calculated to generate multi-level interpretations that include global, subset, and local levels. as well as The interpretation results of the multi-level interpretation are displayed in multiple views, wherein... The global-level explanation presents the importance of each modality among the language, auditory, and visual modalities, as well as the interactions between these modalities. The subset-level interpretation utilizes templates to summarize influential and frequent features, and can use different symbols to explore features of interest to the user. The local level explanation can visualize instances and their explanations.

Images