Here are the most outstanding articles on AI research. It is a carefully curated list of the most recent developments in data science and AI, presented chronologically with a link to a longer article for more information.

Towards Transparent and Explainable Attention Models

Recent research on the interpretability of attention distributions has developed concepts related to accurate and credible explanations for a model's predictions. Attention distributions can be deemed reliable when a more considerable attention weight corresponds to a more significant influence on the model's prediction. If a reasonable explanation provides a rationale for the model's predictions that is understandable to humans, it can be called valid.

In this study, the researchers initially elucidate why the existing attention mechanisms in LSTM-based encoders cannot offer an accurate or convincing account of the model's predictions. In LSTM-based encoders, the hidden representations at different time steps exhibit high similarity (high conicity). As a result, the attention weights in these cases hold little significance, as even a random rearrangement of the attention weights does not impact the model's predictions.

Through extensive experimentation on many tasks and datasets, the researchers have seen that attention distributions frequently assign significance to insignificant terms like punctuation, failing to provide a credible rationale for the model's predictions. To enhance the reliability and credibility of attention processes, the researchers suggest a revised LSTM cell with a training aim that prioritizes diversity. This objective guarantees that the hidden representations acquired at various time steps are distinct.

The researchers demonstrate that the attention distributions from their study offer more transparency. It is due to three main reasons:

• They provide a more accurate ranking of the hidden states in terms of importance.
• They are more indicative of the words crucial for the model's predictions.
• They exhibit a stronger correlation with gradient-based attribution approaches.

Human assessments demonstrate that the attention distributions acquired by their model provide a credible justification for the model's predictions.

Towards Emotion-aided Multi-modal Dialogue Act Classification

Extensive research has been conducted on Dialogue Act Classification (DAC) to identify communication purposes. However, these studies confine their focus solely to written language. Non-verbal cues such as changes in tone and facial expressions can be used to identify DAs. Therefore, it is advantageous to include multi-modal inputs in the task. Furthermore, the speaker's emotional condition significantly influences the selection of the dialogue act, as emotions frequently impact talks. Therefore, it is necessary to investigate the impact of emotions on the automatic recognition of DAs.

This study investigates the significance of both multi-modality and emotion recognition (ER) in DAC. DAC and ER mutually benefit from multi-task learning. This work makes a significant addition by creating a new EMOTyDA, a multimodal Emotion aware Dialogue Act dataset. EMOTyDA is collected using dialogue datasets that are available as open-source. To showcase the effectiveness of EMOTyDA, the researchers constructed a Deep Neural Network (DNN) that incorporates attention and can handle several modalities and tasks. This DNN enables the simultaneous learning of dialogue acts (DAs) and emotions.

Furthermore, the researchers provide empirical evidence that multimodality and multitasking lead to superior DAC performance compared to unimodal and single-task DAC variations.

Sentiment and Emotion help Sarcasm? A Multi-task Learning Framework for Multi-Modal Sarcasm, Sentiment and Emotion Analysis

This paper posits that sarcasm is intricately linked to mood and emotion. As a result, the researchers propose a multi-task deep learning framework to address these three issues concurrently in a multi-modal conversational setting. The researchers manually labelled the newly available multi-modal MUStARD sarcasm dataset using sentiment and emotion classes, including implicit and explicit categories.

They suggest two attention processes for multi-tasking: Inter-segment Inter-modal Attention (Ie-Attention) and Intra-segment Inter-modal Attention (Ia-Attention). The primary objective of Ie-Attention is to acquire knowledge about the correlation between the various components of the sentence across multiple modalities. On the other hand, Ia-Attention directs its focus to the same part of the text across different modalities. Ultimately, the representations from both attention mechanisms are combined and distributed across the five classes (sarcasm, implicit sentiment, explicit sentiment, implicit emotion, explicit emotion) to perform several tasks simultaneously. 

The experimental findings on the expanded MUStARD dataset demonstrate the effectiveness of their suggested method for identifying sarcasm, surpassing the current cutting-edge methods. The evaluation further indicates that the proposed multi-task architecture achieves superior performance for the primary goal, specifically sarcasm detection, by incorporating two additional tasks: emotion and sentiment analysis.

Image source: Unsplash