Multi-head self-attention-based Transformers have shown promise in different learning tasks. Albeit these models exhibit significant improvement in understanding short-term and long-term contexts from sequences, encoders of Transformers and their variants fail to preserve layer-wise contextual information. Transformers usually project tokens onto sparse manifolds and fail to preserve mathematical equivalence among the token representations. In this work, we propose TransJect, an encoder model that guarantees a theoretical bound for layer-wise distance preservation between a pair of tokens. We propose a simple alternative to dot-product attention to ensure Lipschitz continuity. This allows TransJect to learn injective mappings to transform token representations to different manifolds with similar topology and preserve Euclidean distance between every pair of tokens in subsequent layers. Evaluations across multiple benchmark short- and long-sequence classification tasks show maximum improvements of 6.8% and 5.9%, respectively, over the variants of Transformers. Additionally, TransJect displays 79% better performance than Transformer on the language modeling task. We further highlight the shortcomings of multi-head self-attention from the statistical physics viewpoint. Although multi-head self-attention was incepted to learn different abstraction levels within the networks, our empirical analyses suggest that different attention heads learn randomly and unorderly. In contrast, TransJect adapts a mixture of experts for regularization; these experts are more orderly and balanced and learn different sparse representations from the input sequences. TransJect exhibits very low entropy and can be efficiently scaled to larger depths.

Understanding emotions during conversation is a fundamental aspect of human communication, driving NLP research for Emotion Recognition in Conversation (ERC). While considerable research has focused on discerning emotions of individual speakers in monolingual dialogues, understanding the emotional dynamics in code-mixed conversations has received relatively less attention. This motivates our undertaking of ERC for code-mixed conversations in this study. Recognizing that emotional intelligence encompasses a comprehension of worldly knowledge, we propose an innovative approach that integrates commonsense information with dialogue context to facilitate a deeper understanding of emotions. To achieve this, we devise an efficient pipeline that extracts relevant commonsense from existing knowledge graphs based on the code-mixed input. Subsequently, we develop an advanced fusion technique that seamlessly combines the acquired commonsense information with the dialogue representation obtained from a dedicated dialogue understanding module. Our comprehensive experimentation showcases the substantial performance improvement obtained through the systematic incorporation of commonsense in ERC. Both quantitative assessments and qualitative analyses further corroborate the validity of our hypothesis, reaffirming the pivotal role of commonsense integration in enhancing ERC.

Sentiment analysis with deep learning in resource-constrained languages is a challenging task. In this paper, we introduce a novel approach for sentiment analysis in resource-constrained scenarios using character embedding and cross-lingual sentiment analysis with transliteration. We use this method to introduce the novel task of inducing sentiment polarity of words and sentences and aspect term sentiment analysis in the no-resource scenario. We formulate this task by taking a metalingual approach whereby we transliterate data from closely related languages and transform it into a meta language. We also demonstrated the efficacy of using character-level embedding for sentence representation. We experimented with 4 Indian languages – Bengali, Hindi, Tamil, and Telugu, and obtained encouraging results. We also presented new state-of-the-art results on the Hindi sentiment analysis dataset leveraging our metalingual character embeddings.

Code-mixed text infused with low resource language has always been a challenge for natural language understanding models. A significant problem while understanding such texts is the correlation between the syntactic and semantic arrangement of words. The phonemes of each character in a word dictates the spelling representation of a term in low resource language. However, there is no universal protocol or alphabet mapping for code-mixing. In this paper, we highlight the impact of spelling variations in code-mixed data for training natural language understanding models. We emphasize the impact of using phonetics to neutralize this variation in spelling across different usage of a word with the same semantics. The proposed approach is a computationally inexpensive technique and improves the performances of state-of-the-art models for three dialog system tasks viz. intent classification, slot-filling, and response generation. We propose a data pipeline for normalizing spelling variations irrespective of language.