Relevance in e-commerce product search is critical to ensuring that results accurately reflect customer intent. While large language models (LLMs) have recently advanced natural language processing capabilities, their high inference latency and significant infrastructure demands make them less suitable for real-time e-commerce applications. Consequently, transformer-based encoder models are widely adopted for relevance classification tasks. These models typically evaluate the relevance of a product to a given query by encoding the query and product title as input features. As e-commerce stores expand into new marketplaces, the need for language- and region-specific relevance models grows, often resulting in the sequential development and maintenance of separate models per marketplace. To address this challenge, we introduce a multilingual continual learning (CL) framework that mitigates catastrophic forgetting. Our proposed method, FR-LoRA (Fisher Regularized LoRA), integrates Elastic Weight Consolidation (EWC) with marketplace-specific LoRA modules, where each LoRA is regularized using the Fisher information matrix. FR-LoRA retains the same inference-time footprint as the base model, ensuring zero additional latency while enabling frequent, scalable updates. Empirically, our approach achieves a ~3% ROC-AUC improvement over single-marketplace baselines and outperforms several recent CL baselines on both proprietary and public datasets.

Entity Recognition (ER) is a common natural language processing task encountered in a number of real-world applications. For common domains and named entities such as places and organisations, there exists sufficient high quality annotated data and foundational models such as T5 and GPT-3.5 also provide highly accurate predictions. However, for niche domains such as e-commerce and medicine with specialized entity types, there is a paucity of labeled data since manual labeling of tokens is often time-consuming and expensive, which makes entity recognition challenging for such domains. Recent works such as NEEDLE [48] propose hybrid solutions to efficiently combine a small amount of strongly labeled (human-annotated) with a large amount of weakly labeled (distant supervision) data to yield superior performance relative to supervised training. The extensive noise in the weakly labeled data, however, remains a challenge. In this paper, we propose WeSDoM (Weak Supervision with Domain Models), which leverages pretrained encoder models from the same domain but different tasks to create domain ontologies that can enable the creation of less noisy weakly labeled data. Experiments on internal e-commerce and public biomedical NER datasets demonstrate that WeSDoM outperforms existing SOTA baselines by a significant margin. We achieve new SOTA F1 scores on two popular Biomedical NER datasets, BC5CDR-chem 94.27, BC5CDR-disease 91.23.

Automatic extraction of attribute preferences from search queries is a critical problem in providing accurate product recommendations to customer. The task becomes even more challenging in cold-start settings where we do not have any supervised/labelled data available to train ML models. In this work, we implement a novel dataset generation pipeline (LLM-API) that leverages Large Language Models (LLMs), search logs and proprietary product information data from an ecommerce website to create a high quality dataset. Our proposed pipeline of LLM-API is robust as it can generalize to any product category with minimal changes in the LLM prompts. For the problem of converting product search queries to API calls we propose a multi-task schema generator model which we train on our generated dataset. Experiments on an internal test set reveals that our proposed model achieves an improvement of ≈9.6% and ≈5% in Exact Match and Micro-F1 respectively, over competitive baselines. Benchmarking our approach on public test set of search queries further reveals a gain of ≈8.6% and ≈10.5% in Exact Match and Micro-F1. We further demonstrate that our approach outperforms a state-of-the-art LLM (Claude) applied on our task using few-shot prompting and CoT reasoning, while at the same time, achieves improvement in inference latency.

Query reformulation (QR) is a widely used technique in web and product search. In QR, we map a poorly formed or low coverage user query to a few semantically similar queries that are rich in product coverage, thereby enabling effective targeted searches with less cognitive load on the user. Recent QR approaches based on generative language models are superior to informational retrieval-based methods but exhibit key limitations: (i) generated reformulations often have low lexical diversity and fail to retrieve a large set of relevant products of a wider variety, (ii) the training objective of generative models does not incorporate a our goal of improving product coverage. In this paper, we propose RLQR (Reinforcement Learning for Query Reformulations), for generating high quality diverse reformulations which aim to maximize the product coverage (number of distinct relevant products returned). We evaluate our approach against supervised generative models and strong RL-based methods. Our experiments demonstrate a 28.6% increase in product coverage compared to a standard generative model, outperforming SOTA benchmarks by a significant margin. We also conduct our experiments on an external Amazon shopping dataset and demonstrate increased product coverage over SOTA algorithms.

Pre-trained language models like BERT have reported state-of-the-art performance on several Natural Language Processing (NLP) tasks, but high computational demands hinder its widespread adoption for large scale NLP tasks. In this work, we propose a novel routing based early exit model called BE3R (BERT based Early-Exit using Expert Routing), where we learn to dynamically exit in the earlier layers without needing to traverse through the entire model. Unlike the exiting early-exit methods, our approach can be extended to a batch inference setting. We consider the specific application of search relevance filtering in Amazon India marketplace services (a large e-commerce website). Our experimental results show that BE3R improves the batch inference throughput by 46.5% over the BERT-Base model and 35.89% over the DistilBERT-Base model on large dataset with 50 Million samples without any trade-off on the performance metric. We conduct thorough experimentation using various architectural choices, loss functions and perform qualitative analysis. We perform experiments on public GLUE Benchmark and demonstrate comparable performance to corresponding baseline models with 23% average throughput improvement across tasks in batch inference setting.

Numerous problems of practical significance such as clickthrough rate (CTR) prediction, forecasting, tagging and so on, involve complex interaction of various user, item and context features. Manual feature engineering has been used in the past to model these combinatorial features but it requires domain expertise and becomes prohibitively expensive as the number of features increases. Feedforward neural networks alleviate the need for manual feature engineering to a large extent and have shown impressive performance across multiple domains due to their ability to learn arbitrary functions. Despite multiple layers of non-linear projections, neural networks are limited in their ability to efficiently model functions with higher order interaction terms. In recent years, Factorization Machines and its variants have been proposed to explicitly capture higher order combinatorial interactions. However not all feature interactions are equally important, and in sparse data settings, without a suitable suppression mechanism, this might result into noisy terms during inference and hurt model generalization. In this work we present Dual Attentional Higher Order Factorization Machine (DA-HoFM), a unified attentional higher order factorization machine which leverages a compositional architecture to compute higher order terms with complexity linear in terms of maximum interaction degree. Equipped with sparse dual attention mechanism, DA-HoFM summarizes interaction terms at each layer, and is able to efficiently select important higher order terms. We empirically demonstrate effectiveness of our proposed models on the task of CTR prediction, where our model exhibits superior performance compared to the recent state-of-the-art models, outperforming them by up to 6.7% on the logloss metric.

Lack of calibrated product sizing in popular categories such as apparel and shoes leads to customers purchasing incorrect sizes, which in turn results in high return rates due to fit issues. We address the problem of product size recommendations based on customer purchase and return data. We propose a novel approach based on Bayesian logit and probit regression models with ordinal categories Small, Fit, Largeto model size fits as a function of the difference between latent sizes of customers and products. We propose posterior computation based on mean-field variational inference, leveraging the Polya-Gamma augmentation for the logit prior, that results in simple updates, enabling our technique to efficiently handle large datasets. Our Bayesian approach effectively deals with issues arising from noise and sparsity in the data providing robust recommendations. Offline experiments with real-life shoe datasets show that our model outperforms the state-of-the-art in 5 of 6 datasets. and leads to an improvement of 17-26% in AUC over baselines when predicting size fit outcomes.

We propose a novel latent factor model for recommending product size fits {Small, Fit, Large} to customers. Latent factors for customers and products in our model correspond to their physical true size, and are learnt from past product purchase and returns data. The outcome for a customer, product pair is predicted based on the difference between customer and product true sizes, and efficient algorithms are proposed for computing customer and product true size values that minimize two loss function variants. In experiments with Amazon shoe datasets, we show that our latent factor models incorporating personas, and leveraging return codes show a 17-21% AUC improvement compared to baselines. In an online A/B test, our algorithms show an improvement of 0.49% in percentage of Fit transactions over control.