Retrieval on English AI Terms Dictionary

Hybrid Search

Sat, 18 Jul 2026 10:01:39 +0000

Definition

Hybrid Search integrates two distinct retrieval methods: dense vector search, which captures semantic meaning and context, and sparse vector (keyword) search, which matches exact terms. By leveraging the strengths of both approaches, it mitigates the limitations of relying on a single method, such as missing synonyms in keyword search or lacking precision in pure semantic search. This approach is widely used in modern enterprise search engines and RAG applications to deliver highly relevant results across diverse query types.

Dataset:Ms Marco

Sat, 18 Jul 2026 09:53:44 +0000

Definition

MS MARCO (Microsoft Machine Reading Comprehension) is a widely used dataset in natural language processing, particularly for information retrieval and question answering. It consists of anonymized search queries from Bing and corresponding relevant passages from web documents. Researchers use it to train models to rank documents based on relevance to a query or to extract direct answers, serving as a foundational benchmark for modern dense retrieval and passage ranking models.

Dataset:Natural Questions

Sat, 18 Jul 2026 09:53:44 +0000

Definition

Natural Questions (NQ) is a benchmark dataset introduced by Google to advance research in open-domain question answering. It maps real, anonymized search queries from Google to long-form answers found within Wikipedia articles. The dataset includes both ‘short answers’ (specific spans of text) and ’long answers’ (paragraphs containing the short answer). It is essential for training models that can retrieve and synthesize information from vast knowledge bases to answer complex, real-world questions.

Dataset:Gooaq

Sat, 18 Jul 2026 09:53:29 +0000

Definition

GooAQ is a dataset compiled from the Google Answers service, featuring a massive collection of user-submitted questions along with detailed, paid responses. It serves as a valuable resource for training models in open-domain question answering and information retrieval. The diversity of topics and the structured nature of the Q&A pairs allow researchers to develop systems capable of understanding complex user intents and retrieving relevant factual information from vast corpora.

Dataset:Embedding Data/Paq Pairs

Sat, 18 Jul 2026 09:53:15 +0000

Definition

The PAQ (Pseudo-Answer Quality) dataset contains millions of automatically generated question-answer pairs extracted from Wikipedia. It is specifically engineered to train dense retrievers by providing negative samples and positive matches for learning embedding spaces where relevant passages are clustered closely together. This approach significantly improves the efficiency and accuracy of open-domain question answering systems.

Summary

A large-scale dataset of question-answer pairs derived from Wikipedia, designed for dense passage retrieval training.

Question Answering

Sat, 18 Jul 2026 09:42:48 +0000

Definition

Question Answering (QA) involves retrieving or generating accurate responses to user queries from a given context or knowledge base. It ranges from closed-domain QA, which relies on specific documents, to open-domain QA, which uses vast amounts of external data. Modern QA systems leverage transformer architectures to understand semantic intent and extract relevant information, powering virtual assistants, search engines, and customer support bots.

Summary

An NLP task where a system automatically provides precise answers to questions posed in natural language.

Matching

Sat, 18 Jul 2026 09:34:02 +0000

Definition

Matching is a critical technique in machine learning used to establish relationships between disparate data entities. In computer vision, feature matching identifies corresponding points across images. In recommendation systems, it pairs users with relevant items based on similarity metrics. Algorithmically, it can range from simple nearest-neighbor searches to complex bipartite graph matching problems. Effective matching relies heavily on robust embedding spaces and distance metrics to ensure that semantically or structurally similar items are correctly paired, enhancing retrieval accuracy and personalization.