Scholarlink

<p>Information Retrieval (IR) in low-resource languages remains a largely underexplored domain due to the scarcity of annotated corpora, standardized tools, and semantic resources. Assamese, is an Indo-Aryan language spoken by over 15 million people, which lacks a robust digital infrastructure for retrieval tasks. This paper presents the first end-to-end IR pipeline, Assamese Information Retrieval using Classification and Embeddings (AIRCE). We introduce AssamIR-28k, a manually curated and genre-labeled news article corpus, and develop a hybrid retrieval system that integrates a genre-aware IndicBERT-based query classifier with a TF-IDF weighted FastText embedding ranker. AIRCE leverages both classification and semantic embedding techniques to deliver efficient and accurate retrieval. AIRCE reduces the search space and improves retrieval quality by addressing morphological variation and lexical mismatch prevalent in Assamese. Compared against six strong lexical baselines which include BM25, PL2, and TF-IDF, AIRCE achieves statistically significant gains across five standard IR metrics. Notably, it improves nDCG@10 by +4.1 percentage points and accuracy@10 by +6.3 points over the best baseline, with all improvements verified via one-sided paired t-tests (p &lt; 0.05). We further provide a qualitative analysis highlighting the AIRCE’s strengths in handling synonymy, inflection, and compounding. This work establishes a reproducible benchmark for Assamese IR and paves the way for scalable, context-aware retrieval in other low-resource languages.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Assamese information retrieval using classification and embedding techniques

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Manash Pratim Lahkar,
  • Utpal Sharma,
  • Tribikram Pradhan

摘要

Information Retrieval (IR) in low-resource languages remains a largely underexplored domain due to the scarcity of annotated corpora, standardized tools, and semantic resources. Assamese, is an Indo-Aryan language spoken by over 15 million people, which lacks a robust digital infrastructure for retrieval tasks. This paper presents the first end-to-end IR pipeline, Assamese Information Retrieval using Classification and Embeddings (AIRCE). We introduce AssamIR-28k, a manually curated and genre-labeled news article corpus, and develop a hybrid retrieval system that integrates a genre-aware IndicBERT-based query classifier with a TF-IDF weighted FastText embedding ranker. AIRCE leverages both classification and semantic embedding techniques to deliver efficient and accurate retrieval. AIRCE reduces the search space and improves retrieval quality by addressing morphological variation and lexical mismatch prevalent in Assamese. Compared against six strong lexical baselines which include BM25, PL2, and TF-IDF, AIRCE achieves statistically significant gains across five standard IR metrics. Notably, it improves nDCG@10 by +4.1 percentage points and accuracy@10 by +6.3 points over the best baseline, with all improvements verified via one-sided paired t-tests (p < 0.05). We further provide a qualitative analysis highlighting the AIRCE’s strengths in handling synonymy, inflection, and compounding. This work establishes a reproducible benchmark for Assamese IR and paves the way for scalable, context-aware retrieval in other low-resource languages.