Multilingual encoder-based language models are widely used for code-mixed analysis, yet their internal representations of code-mixed inputs—and their relationship to the constituent languages—remain poorly understood. Using Hindi–English as a case study, we construct a unified trilingual corpus of parallel English, Hindi (Devanagari), and Romanized code-mixed sentences. We then probe cross-lingual representation alignment in standard multilingual encoders and their code-mix-adapted variants using CKA, token-level saliency, and entropy-based uncertainty analysis. We find that while standard models align English and Hindi well, code-mixed inputs remain loosely connected to either language — and that continued pre-training on code-mixed data improves English–code-mixed alignment at the cost of English–Hindi alignment. Interpretability analyses further reveal a clear asymmetry: models process code-mixed text through an English-dominant semantic subspace, while native-script Hindi provides complementary signals that reduce representational uncertainty. Motivated by these findings, we introduce a trilingual post-training alignment objective that brings code-mixed representations closer to both constituent languages simultaneously, yielding more balanced cross-lingual alignment and downstream gains on sentiment analysis and hate speech detection — showing that grounding code-mixed representations in their constituent languages meaningfully helps cross-lingual understanding. Code is available at: https://github.com/debajyotimaz/tri_align_EMNLP_2026.

Large language models often exhibit a substantial gap between their performance in English and in lower-resourced languages on equivalent knowledge queries—a cross-lingual consistency issue that remains underexplored for Indian languages and their code-mixed counterparts. To study this gap, we introduce IndicKLAR, an Indic extension of the KLAR-CLC benchmark covering 18 of the 22 scheduled Indian languages. For 11 widely used language pairs, we additionally provide code-mixed variants. Both monolingual and code-mixed inputs verified by native speakers. This three-way alignment enables us to examine how knowledge recall consistency varies across English, code-mixed, and native Indian language inputs. Across nine open-weight models, we find that the accuracy gap between native-language and English inputs can reach ∼0.50, while code-mixed inputs substantially reduce this gap, bringing performance within ∼0.05 of English without any model-level intervention. Motivated by this finding, we evaluate several prompting strategies that differ in how explicitly language conversion is exposed: a two-stage translate-then-answer setup, a one-stage joint translation-and-answer prompt, and Translate-in-Thought (TinT)—a single-step strategy in which the model internally converts the input and outputs only the final answer. Across the native → code-mixed → English performance trajectory, we observe a consistent flip point—the transition from incorrect to correct prediction—between the native and code-mixed settings. Notably, this pattern holds both when the code-mixed representation is explicitly provided as input or when the model is prompted to convert internally using TinT.

In this paper, we reported our experiments with various strategies to improve code-mixed humour and sarcasm detection. Particularly, we tried three approaches: (i) native sample mixing, (ii) multi-task learning (MTL), and (iii) prompting and instruction finetuning very large multilingual language models (VMLMs). In native sample mixing, we added monolingual task samples to code-mixed training sets. In MTL learning, we relied on native and code-mixed samples of a semantically related task (hate detection in our case). Finally, in our third approach, we evaluated the efficacy of VMLMs via few-shot context prompting and instruction finetuning. Some interesting findings we got are (i) adding native samples improved humor (raising the F1-score up to 6.76%) and sarcasm (raising the F1-score up to 8.64%) detection, (ii) training MLMs in an MTL framework boosted performance for both humour (raising the F1-score up to 10.67%) and sarcasm (increment up to 12.35% in F1-score) detection, and (iii) prompting and instruction finetuning VMLMs couldn’t outperform the other approaches. Finally, our ablation studies and error analysis discovered the cases where our model is yet to improve. We provided our code for reproducibility.

This paper describes the system architectures and the models submitted by our team “IISERB Brains” to SemEval 2022 Task 6 competition. We contested for all three sub-tasks floated for the English dataset. On the leader-board, we got 19th rank out of 43 teams for sub-task A, 8th rank out of 22 teams for sub-task B, and 13th rank out of 16 teams for sub-task C. Apart from the submitted results and models, we also report the other models and results that we obtained through our experiments after organizers published the gold labels of their evaluation data. All of our code and links to additional resources are present in GitHub for reproducibility.

There is a huge difference between a scientific journal reporting ‘wine consumption might be correlated to cancer’, and a media outlet publishing ‘wine causes cancer’ citing the journal’s results. The above example is a typical case of a scientific statement being exaggerated as an outcome of the rising problem of media manipulation. Given a pair of statements (say one from the source journal article and the other from the news article covering the results published in the journal), is it possible to ascertain with some confidence whether one is an exaggerated version of the other? This paper presents a surprisingly simple yet rational three-step approach that performs best for this task. We solve the task by breaking it into three sub-tasks as follows – (a) given a statement from a scientific paper or press release, we first extract relation phrases (e.g., ‘causes’ versus ‘might be correlated to’) connecting the dependent (e.g., ‘cancer’) and the independent (‘wine’) variable, (b) classify the strength of the relationship phrase extracted and (c) compare the strengths of the relation phrases extracted from the statements to identify whether one statement contains an exaggerated version of the other, and to what extent. Through rigorous experiments, we demonstrate that our simple approach by far outperforms baseline models that compare state-of-the-art embedding of the statement pairs through a binary classifier or recast the problem as a textual entailment task, which appears to be a very natural choice in this settings.

In this paper, we demonstrate how code-switching patterns can be utilised to improve various downstream NLP applications. In particular, we encode various switching features to improve humour, sarcasm and hate speech detection tasks. We believe that this simple linguistic observation can also be potentially helpful in improving other similar NLP applications.

This paper describes our approach to solve Semeval task 3: EmoContext; where, given a textual dialogue i.e. a user utterance along with two turns of context, we have to classify the emotion associated with the utterance as one of the following emotion classes: Happy, Sad, Angry or Others. To solve this problem, we experiment with different deep learning models ranging from simple bidirectional LSTM (Long and short term memory) model to comparatively complex attention model. We also experiment with word embedding conceptnet along with word embedding generated from bi-directional LSTM taking input characters. We fine-tune different parameters and hyper-parameters associated with each of our models and report the value of evaluating measure i.e. micro precision along with class wise precision, recall and F1-score of each system. We report the bidirectional LSTM model, along with the input word embedding as the concatenation of word embedding generated from bidirectional LSTM for word characters and conceptnet embedding, as the best performing model with a highest micro-F1 score of 0.7261. We also report class wise precision, recall, and f1-score of best performing model along with other models that we have experimented with.

In this paper we propose a deep learning framework for sarcasm target detection in predefined sarcastic texts. Identification of sarcasm targets can help in many core natural language processing tasks such as aspect based sentiment analysis, opinion mining etc. To begin with, we perform an empirical study of the socio-linguistic features and identify those that are statistically significant in indicating sarcasm targets (p-values in the range(0.05,0.001)). Finally, we present a deep-learning framework augmented with socio-linguistic features to detect sarcasm targets in sarcastic book-snippets and tweets. We achieve a huge improvement in the performance in terms of exact match and dice scores compared to the current state-of-the-art baseline.

n this paper, we present a set of computational methods to identify the likeliness of a word being borrowed, based on the signals from social media. In terms of Spearman’s correlation values, our methods perform more than two times better (∼ 0.62) in predicting the borrowing likeliness compared to the best performing baseline (∼ 0.26) reported in literature. Based on this likeliness estimate we asked annotators to re-annotate the language tags of foreign words in predominantly native contexts. In 88% of cases the annotators felt that the foreign language tag should be replaced by native language tag, thus indicating a huge scope for improvement of automatic language identification systems.