PaperlayerView exact revision on arXiv

Source-linked AI summary

BiERU: Bidirectional Emotional Recurrent Unit for Conversational Sentiment Analysis

Wei Li, Wei Shao, Shaoxiong Ji, Erik Cambria

arXiv:2006.00492v3cs.CLcs.LG

TL;DR

Conversational sentiment analysis needs effective contextual encoding because dialogue context can influence an utterance’s sentiment. The paper proposes BiERU, a compact party-ignorant framework combining GNTB for context compositionality with a two-channel feature extractor for emotional features. Across three standard datasets, the model outperforms state-of-the-art methods in most cases while using fewer parameters.

  • Problem

    Conversational sentiment analysis must account for contextual information that may influence an utterance, while existing approaches use complicated structures to model dialogue context and parties.

  • Method

    BiERU uses a party-ignorant bidirectional emotional recurrent unit with GNTB for context compositionality and a two-channel LSTM/CNN feature extractor.

  • Results

    Experiments on three standard datasets show that BiERU outperforms state-of-the-art methods in most cases with fewer parameters.

  • Takeaways & Limitations

    The paper concludes that conversational sentiment models can simplify their structure while improving performance, while supporting conversations with arbitrary turns and speakers.

Abstract

from arXiv · show

Sentiment analysis in conversations has gained increasing attention in recent years for the growing amount of applications it can serve, e.g., sentiment analysis, recommender systems, and human-robot interaction. The main difference between conversational sentiment analysis and single sentence sentiment analysis is the existence of context information which may influence the sentiment of an utterance in a dialogue. How to effectively encode contextual information in dialogues, however, remains a challenge. Existing approaches employ complicated deep learning structures to distinguish different parties in a conversation and then model the context information. In this paper, we propose a fast, compact and parameter-efficient party-ignorant framework named bidirectional emotional recurrent unit for conversational sentiment analysis. In our system, a generalized neural tensor block followed by a two-channel classifier is designed to perform context compositionality and sentiment classification, respectively. Extensive experiments on three standard datasets demonstrate that our model outperforms the state of the art in most cases.

I. INTRODUCTION

Conversational sentiment analysis must encode context because preceding utterances can alter an utterance’s sentiment. BiERU addresses this with a compact, party-ignorant architecture that composes context and extracts emotional features, and experiments report stronger performance with fewer parameters.

  • Contribution: BiERU is a fast, compact, parameter-efficient, party-ignorant framework for conversational sentiment analysis.Its recurrent emotional unit contains a generalized neural tensor block and a two-channel feature extractor.
  • Motivation: Context can enhance, weaken, or reverse an utterance’s raw sentiment, making contextual encoding central to dialogue sentiment analysis.Existing methods use separate steps to obtain context, model its influence, and extract emotional features.
  • Contribution: GNTB performs context compositionality by combining historical context with the current utterance and producing a contextual utterance representation.This formulation combines context acquisition and context integration into one operation before emotional feature extraction.
  • Architecture: TFE uses LSTM and one-dimensional CNN branches to extract emotional features from the contextual utterance vector.The two-channel design is followed by sentiment classification.
  • Results: Experiments on three standard benchmarks indicate that BiERU outperforms state-of-the-art methods with fewer parameters.The paper positions this result as evidence that model simplification and improved performance can coexist.

III. METHOD

The method represents conversational utterances as vectors and processes them with bidirectional ERUs. GNTB incorporates context into each utterance, while TFE extracts emotional features for sentiment classification or regression.

  • Problem Definition: Conversational sentiment analysis predicts sentiment labels or intensities for the utterances in a multiple-turn conversation.The task is generally multi-class classification, though continuous sentiment intensity can make it a regression problem.
  • Utterance Representation: Utterances are embedded and passed through CNN feature extraction using convolution filters of sizes 3, 4, and 5.The resulting activations are max-pooled, transformed with ReLU, and used to form textual utterance representations.
  • BiERU: BiERU applies two ERUs in forward and backward directions, concatenating their outputs for sentiment classification or regression.This bidirectional structure processes the input utterances in both directions.
  • BiERU: GNTB simultaneously encodes context and incorporates it into the current utterance, producing a contextual utterance vector.TFE then receives this vector and extracts emotional features for classification or regression.

1) Generalized Neural Tensor Block:

The generalized neural tensor block (GNTB) performs context compositionality by combining the current utterance with contextual information, producing a contextual utterance vector for emotional feature extraction. Its formulation supports multiple compositionality types and lower space complexity than the neural tensor network.

  • Generalized Neural Tensor Block: GNTB combines the current utterance vector u_t with contextual information to produce contextual utterance vector p_t.It extracts context, incorporates it into the current utterance, and treats the combination as context compositionality.
  • Generalized Neural Tensor Block: Each tensor slice represents a specific type of context compositionality, with k slices forming a k-dimensional compositionality space.The model can represent different compositionality types rather than only one recursive structure.
  • Generalized Neural Tensor Block: GNTB supports structures beyond recursive composition and has space complexity O(kd^2), compared with O(d^3) for NTN.NTN is described as a special case of GNTB when k equals d.
  • Two-channel Feature Extractor: The resulting contextual utterance vector is passed to the two-channel feature extractor, whose LSTM and CNN branches produce complementary emotional features.Both branches receive the same contextual utterance vector and their outputs are combined for downstream prediction.

3) Sentiment Classification & Regression:

The model uses the extracted emotion feature vector for either sentiment classification or sentiment regression. Classification predicts a sentiment-label distribution, while regression predicts a scalar sentiment intensity.

  • Sentiment Classification: A linear network followed by softmax maps emotion feature e_t to the sentiment-label probability distribution S_t.The number of output classes equals the number of sentiment labels.
  • Sentiment Classification: The predicted sentiment label for utterance u_t is selected as the most probable sentiment class.The probability distribution S_t is used for this final class decision.
  • Sentiment Regression: For sentiment regression, a linear network maps emotion feature e_t to the predicted sentiment intensity q_t.The regression output is a scalar rather than a sentiment-label distribution.

4) Training:

Training uses task-specific losses with L2 regularization and Adam optimization. The model focuses on textual modality, while the supplied architecture passages describe forward and backward contextual representations.

  • Training Objectives: Classification training uses cross-entropy loss with L2 regularization to relieve overfitting.The loss formulation is associated with sentiment-label probabilities and expected class labels.
  • Training Objectives: Regression training uses mean square error with L2 regularization to relieve overfitting.The regression objective measures error in predicted sentiment intensity.
  • Context Variants: Fig. 2 contrasts BiERU architectures using global context and local context.The caption distinguishes the two architectural forms by their source of context information.
  • Model Setting: The model uses textual modality only, and the detailed GNTB and TFE structures are presented in Fig. 3.The supplied passage identifies textual modality as the sole modality used in the model.
  • Optimization: The network is trained with stochastic-gradient-descent-based Adam optimization.The trainable parameters are optimized using Adam.

D. Bidirectional Emotion Recurrent Unit Variants

BiERU has global- and local-context variants that use bidirectional recurrent processing to encode conversational context, with experiments conducted on three textual benchmarks.

  • BiERU-gc: BiERU-gc uses recurrent contextual utterance vectors to capture information from all preceding utterances and surrounding dialogue context.The contextual vector at time t recursively contains information from u1 through u_t−1.
  • BiERU-lc: BiERU-lc uses adjacent utterances as context, incorporating information from u_t−1 and u_t+1 for the current utterance.Replacing the recurrent contextual vector with the previous utterance gives local context from neighboring turns.
  • Experimental data: The variants are evaluated on AVEC, IEMOCAP, and MELD using textual conversational sentiment analysis.These datasets include regression and classification settings, but the experiments use only textual modality.
  • Architecture: GNTB and TFE are the central components shown in Figure 3 for contextual encoding and emotion-feature extraction.GNTB processes contextual utterance vectors, while TFE uses LSTM and CNN channels to produce emotion features.

B. Baselines and Settings

The evaluation compares BiERU with established contextual dialogue models under textual-modality settings, using dataset-specific metrics and regularized training.

  • Baselines: The baselines include c-LSTM, CMN, DialogueRNN, DialogueGCN, and AGHMN, spanning recurrent, memory, graph, and party-aware approaches.DialogueRNN explicitly models speaker, context, and emotion states, whereas AGHMN is party-ignorant.
  • Settings: All experiments use CNN-extracted utterance features, with DialogueRNN representations adopted directly for fair comparison.The experiments focus on textual modality.
  • Implementation: Training uses dropout, Adam optimization, manually tuned hyperparameters, batch size 1, rank r = 10, and PyTorch implementation.Table II reports BiERU-lc hyperparameters across the three datasets.

C. Results

BiERU is compared with baseline methods on three standard benchmarks using textual modality, with results averaged over five runs.

  • Evaluation: The experiments compare BiERU with baseline methods on three standard benchmarks using textual modality.The reported results are intended to assess performance against established conversational sentiment-analysis systems.
  • Overall results: The model outperforms all listed baseline methods overall, including DialogueRNN, DialogueGCN, and AGHMN, and markedly exceeds them on some indicators.Results are reported as averages from five experimental runs.
  • Metrics: IEMOCAP uses class-level accuracy and weighted accuracy and F1, AVEC uses regression metrics including Pearson correlation, and MELD uses weighted accuracy.These metrics reflect the differing task formulations across the benchmarks.

1) Comparison with the State of the Art:

BiERU generally improves over state-of-the-art baselines across IEMOCAP, AVEC, and MELD, while local context performs better than global context overall.

  • IEMOCAP: 0.84% and 0.41% are the BiERU-lc gains over DialogueGCN in IEMOCAP weighted average accuracy and F1-score, respectively.BiERU-gc improves those metrics by 0.10% and 0.06%, respectively.
  • AVEC: BiERU outperforms DialogueRNN on AVEC valence, arousal, and expectancy, while matching it on power.BiERU-gc gains 0.04 in Pearson r for arousal, and BiERU-lc gains 0.05; BiERU-lc gains 0.01 for valence and expectancy.
  • Dataset considerations: MELD is harder because conversations are shorter, involve more speakers with fewer turns each, and contain fewer explicit sentiment expressions.These properties make speaker interdependency difficult for party-dependent models such as DialogueRNN.
  • MELD: 60.9% average accuracy is achieved on MELD, exceeding AGHMN by 0.6% and DialogueRNN by 4.8%.The models use GNTB for context compositionality.
  • Variant comparison: BiERU-lc outperforms BiERU-gc overall, including IEMOCAP gains of 0.74% in weighted accuracy and 0.45% in F1-score.The paper attributes this difference partly to redundant information in global context versus neighborhood context.
  • Analysis: The BiERU-lc confusion matrix shows balanced performance across IEMOCAP sentiment classes.The paper uses the heat map as a visualization of class-wise behavior.

E. Visualization

The paper visualizes emotion representations and training behavior to assess feature quality, class separation, convergence, parameter count, and training time.

  • Emotion-feature visualization: PCA visualizations compare BiERU-lc and DialogueRNN emotion representations on IEMOCAP using colors for predicted sentiment labels.The representations are reduced to three dimensions before comparison.
  • Emotion-feature visualization: BiERU-lc extracts happy emotion features more effectively and separates happy–excited and frustrated–angry more clearly than DialogueRNN.Neutral overlaps more with other emotions in both models because it is described as an intermediate emotion.
  • Efficiency analysis: BiERU-lc and BiERU-gc show comparable convergence speed with the compared models in the training-curve analysis.The curves plot training and testing loss using the same loss function for all compared models.
  • Efficiency analysis: BiERU-gc uses fewer trainable parameters and less training time than DialogueRNN and DialogueGCN, while BiERU-lc reduces parameters further through low-rank approximation.For IEMOCAP, BiERU has about 0.5M parameters versus DialogueRNN’s around 1M; for MELD, BiERU-lc has 0.6M versus DialogueRNN’s 2.9M.

G. Ablation Study

The ablation study evaluates GNTB and TFE separately and jointly, showing that their combination provides the effective BiERU configuration while the paper also summarizes broader benchmark performance.

  • Ablation Study: Sole GNTB or TFE performs poorly on accuracy and F1-score because each module omits either emotional features or contextual information.GNTB outputs mainly contain context information, whereas TFE outputs lack context information.
  • Ablation Study: Combining GNTB and TFE in BiERU increases accuracy and F1-score dramatically, indicating that the modules couple effectively.The ablation uses the IEMOCAP dataset and reports weighted-average accuracy and F1-score.
  • Conclusion: BiERU combines context compositionality with emotion-feature extraction while reducing parameters and supporting different structures.The conclusion states that simplifying the model structure and improving performance simultaneously is feasible.
  • Conclusion: BiERU outperforms current state-of-the-art models on three standard datasets in most cases and can model conversations with arbitrary turns and speakers.The paper identifies broader emotion categorization as future work for distinguishing similar emotions.
Loading 2006.00492v3…