PaperlayerView exact revision on arXiv

Source-linked AI summary

Multimodal Routing: Improving Local and Global Interpretability of Multimodal Language Analysis

Yao-Hung Hubert Tsai, Martin Q. Ma, Muqiao Yang, Ruslan Salakhutdinov, Louis-Philippe Morency

arXiv:2004.14198v2cs.CL

TL;DR

Strong-performing multimodal models for sentiment and emotion tasks often have limited interpretability. Multimodal Routing dynamically weights features against prediction concepts, and experiments report competitive performance alongside local and global interpretations of feature importance.

  • Problem

    Strong-performing multimodal models for human-centric tasks are often black boxes with limited interpretability, while global and sample-specific feature importance remain useful to understand.

    Source: §1 · p. 1

  • Method

    Multimodal Routing encodes unimodal, bimodal, and trimodal features, then routes them to prediction concepts using input-specific weights that support local interpretation.

    Source: §3 · p. 3, §3.1 · p. 3

  • Results

    Routing-weight analyses show feature importance varies across samples; globally, acoustic features are crucial for negative sentiment or emotions, and acoustic-visual interactions for angry emotion.

    Source: §6 · p. 8

  • Takeaways & Limitations

    Local and global interpretation of feature importance strengthens understanding of multimodal language learning.

    Source: §1 · p. 2

    The formulation is scoped to audio, visual, and text inputs and multiclass or multilabel prediction tasks.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

The human language can be expressed through multiple sources of information known as modalities, including tones of voice, facial gestures, and spoken language. Recent multimodal learning with strong performances on human-centric tasks such as sentiment analysis and emotion recognition are often black-box, with very limited interpretability. In this paper we propose Multimodal Routing, which dynamically adjusts weights between input modalities and output representations differently for each input sample. Multimodal routing can identify relative importance of both individual modalities and cross-modality features. Moreover, the weight assignment by routing allows us to interpret modality-prediction relationships not only globally (i.e. general trends over the whole dataset), but also locally for each single input sample, meanwhile keeping competitive performance compared to state-of-the-art methods.

1 Introduction

Multimodal Routing aims to make multimodal language models interpretable at both dataset-wide and individual-sample levels by dynamically weighting modality features in relation to output labels. The paper evaluates the approach on sentiment analysis and emotion recognition and examines local and global feature importance.

  • The model’s local weights can associate different modality features with different emotion concepts across input samples.The examples show vision and visual-text features linked to sadness in one sample, and visual-text and text features linked to happiness in another.Source: Figure 1 · p. 1
  • Interpretability can reveal crucial explanatory features and support insight into multimodal learning, model design, and dataset debugging.Global interpretation captures average feature-importance trends, while local interpretation reflects feature importance for individual samples.Source: §1 · p. 1
  • Multimodal Routing dynamically adjusts weights between modality features and output labels for each sample, providing local interpretation during training and inference.It addresses unimodal, bimodal, and trimodal explanatory features.Source: §1 · p. 2
  • The experiments cover sentiment analysis and emotion recognition on IEMOCAP and CMU-MOSEI, with local qualitative and global statistical analyses of adjusted weights.The paper also compares the model with state-of-the-art methods on these tasks.Source: §1 · p. 2

2 Related Work

Prior interpretability approaches do not jointly explain single-modality and cross-modality contributions to predictions. Multimodal Routing instead routes unimodal, bimodal, and trimodal features to decision concepts.

Related Work

  • Post-hoc methods such as LIME (Ribeiro et al., 2016), SHAP (Lundberg and Lee, 2017), and L2X (Chen et al., 2018) explain unimodal, but not bimodal or trimodal, features.Deep neural models of multimodal language were also described as uninterpretable (Wang et al., 2019; Tsai et al., 2019a).Source: §2 · p. 2
  • Graph-MFN does not separate unimodal from multimodal contributions, while MFM (Tsai et al., 2019b) analyzes only trimodal interactions.Neither explains how single modalities and modality interactions jointly contribute to the final prediction.Source: §2 · p. 2
  • Inspired by Capsule Networks (Sabour et al., 2017), Multimodal Routing routes unimodal, bimodal, and trimodal explanatory features to concepts of the model’s decision.Capsule Networks route between layers of capsules; this method routes between multimodal features and decision concepts.Source: §2 · pp. 2–3

3 Method

Multimodal Routing encodes unimodal, bimodal, and trimodal features, routes them into label-associated concepts, and predicts from those concepts. Its routing weights support local, sample-specific and global, dataset-level interpretation.

3 Method

  • The final prediction pipeline has three stages: encoding multimodal inputs, routing features into label-associated concepts, and predicting from the inferred concepts.Source: §3 · p. 3

3.1 Multimodal Routing

  • Encoding represents unimodal, bimodal, and trimodal interactions as explanatory features, while iterative routing assigns features to similar concepts and updates concepts using activation- and routing-weighted features.The encoding functions use Multimodal Transformer (MulT).Source: §3.1 · pp. 3–4

3.2 Interpretability

  • The framework interprets routing coefficients as feature-to-concept assignments, providing a basis for explaining modality contributions locally and globally.Source: §3.2 · p. 4

3.2.1 Local Interpretation

  • For a sample, a feature’s contribution to a concept’s logit depends on its activation, routing agreement, and projected-feature alignment; sample-specific activations and routing weights enable local interpretation.Source: §3.2.1 · p. 4

3.2.2 Global Interpretation

  • Global interpretation analyzes dataset-average routing coefficients using 95% confidence intervals, with the bounded coefficients’ mean treated as approximately normally distributed when the sample is sufficiently large.Source: §3.2.2 · pp. 4–5

4 Experiments

Experiments on CMU-MOSEI and IEMOCAP compare Multimodal Routing with interpretable ablations and established methods. The reported results show improvements over interpretable baselines and competitive performance against state-of-the-art methods across the evaluated tasks.

4.1 Datasets

  • CMU-MOSEI has 23,454 clips for sentiment prediction and emotion recognition, while IEMOCAP has 10K clips for multilabel emotion analysis; experiments use word-aligned text, visual, and acoustic features.CMU-MOSEI sentiment metrics include Acc7, Acc2, and F1; its emotion task uses six-class accuracy and F1. IEMOCAP reports four-class accuracy and F1 for happy, sad, angry, and neutral.Source: §4.1 · p. 5

4.2 Ablation Study and Baseline Models

  • The interpretable baselines are GAM, which sums unimodal, bimodal, and trimodal features without routing coefficients, and Multimodal Routing∗, which uses one routing iteration.Comparisons also include EF-LSTM, LF-LSTM, RAVEN, and Multimodal Transformer as non-interpretable methods.Source: §4.2 · p. 5

4.3 Results and Discussions

  • On CMU-MOSEI sentiment, Multimodal Routing improves over GAM and noniterative Multimodal Routing∗, outperforms EF-LSTM, LF-LSTM, and RAVEN, and performs competitively against MulT.The authors attribute the improvement over interpretable baselines to iterative routing dynamically associating features with prediction concepts.Source: §4.3 · pp. 5–6, Table 1 · p. 6
  • On IEMOCAP emotion recognition, Multimodal Routing performs close to the state of the art, with a drop on “happy” and an advantage on “angry.”The results are reported in Table 1 alongside CMU-MOSEI sentiment.Source: Table 1 · p. 6, §4.3 · p. 6
  • On CMU-MOSEI emotion recognition, Multimodal Routing performs competitively on most metrics; skewed labels lead all models to the same accuracy for “surprise.”RAVEN and MulT are omitted because they did not report CMU-MOSEI emotion results.Source: Table 2 · p. 6, §4.3 · p. 6

5 Interpretation Analysis

Multimodal Routing supports local interpretation through sample-specific feature contributions and global interpretation through averaged routing coefficients compared with uniform weighting. The analyses show how modality and interaction importance varies across inputs and prediction labels.

5.1 Local Interpretation Analysis

  • For a neutral sentiment example, the text-vision interaction receives high contribution because neither text nor vision alone explains the prediction.The speaker’s transcript includes “don’t,” while the frown does not establish neutral sentiment; together, the modalities support the neutral prediction.Source: §5.1 · p. 7
  • In a happy-emotion example, the trimodal interaction contributes significantly, consistent with text, a dramatic smile, and happy tone all informing the prediction.Source: §5.1 · p. 7
  • The six examples show different feature-importance patterns across individual samples, even for the same sentiment, making local routing useful for debugging and interpretation.Averaging samples instead yields a general trend for global interpretation.Source: §5.1 · p. 7

5.2 Global Interpretation Analysis

  • Global sentiment analysis associates neutral predictions with language, extremely negative predictions with acoustic features, and extremely positive predictions with text-acoustic interaction.The analysis compares confidence intervals for averaged routing coefficients with uniform weighting; features are distinguishable when intervals fall outside 1/J.Source: Table 3 · p. 8, §5.2 · p. 8
  • For emotion, global routing associates fear with text, disgust with acoustic features, surprise with text-acoustic interaction, and angry with acoustic-visual interaction; happy and sad are less distinct.For happy and sad, either trimodal interaction has the most significant connection or routing is not significantly different among modalities.Source: Table 3 · p. 8, §5.2 · p. 8

6 Conclusion

Multimodal Routing links explanatory features to predictions for individual inputs and supports global interpretation through routing coefficients. The authors report varied local feature importance and dataset-level modality patterns, with little performance loss compared to state-of-the-art methods.

  • For each input, Multimodal Routing dynamically associates unimodal, bimodal, or trimodal explanatory features with predictions when those features explain them well.Source: §6 · p. 8
  • Routing coefficients reveal substantial sample-to-sample variation in explanatory-feature importance.Source: §6 · p. 8
  • Globally, acoustic features are crucial for predicting negative sentiment or emotions, while acoustic-visual interactions are crucial for predicting angry emotion.The authors state these observations align with prior psychological research.Source: §6 · p. 8
  • The authors report achieving local and global interpretation without much loss of performance compared to state-of-the-art methods.Source: §6 · p. 8

Encoding pi from input

This section describes the model’s feature and pi encoding, training configuration, and CMU-MOSEI task formulations. Its global analyses identify modality and interaction patterns associated with sentiment and emotion predictions.

Encoding pi from input

Training Details and Hyper-parameters

  • Training uses Adam (Kingma and Ba, 2014), batch size 32, dropout 0.5, and learning rates of 1e-4 for CMU-MOSEI sentiment and IEMOCAP and 1e-5 for CMU-MOSEI emotion.Routing uses two iterations between features and 64-dimensional concepts; all experiments share the same hyperparameter configuration.Source: Training Details and Hyper-parameters · p. 10

Global Interpretation Result

  • Acoustic modality supports extremely negative sentiment, text-acoustic interaction supports extremely positive sentiment, and neutral or less-extreme predictions draw on many modalities and interactions.Most modality/interaction activations exceed 0.8, except pvl.Source: Table 4 · p. 11, Global Interpretation Result · p. 11
  • CMU-MOSEI emotion links audio-visual interactions to angry, text to fear, audio to disgust, and text-audio interactions to surprise.Most modality/interaction activations exceed 0.8, except pvl.Source: Table 5 · p. 11, Global Interpretation Result · p. 11

Remarks on CMU-MOSEI Sentiment

  • For CMU-MOSEI sentiment, the model predicts integer labels as a classification task, so mean average error and correlation metrics are not reported.Source: Remarks on CMU-MOSEI Sentiment · p. 11

Remarks on CMU-MOSEI Emotion

  • For CMU-MOSEI emotion, scores in [0, 3] become presence labels: each emotion with a score greater than zero is marked present, yielding a multiclass label.Source: Remarks on CMU-MOSEI Emotion · p. 11
Loading 2004.14198v2…