Semanlink - NLP: use cases AND Information retrieval: techniques

> To capture knowledge in a more modular and interpretable way, we augment language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend over documents from a large corpus such as Wikipedia, used during pre-training, fine-tuning and inference. **For the first time, we show how to pre-train such a knowledge retriever in an unsupervised manner**, using masked language modeling as the learning signal and backpropagating through a retrieval step that considers millions of documents

Hum, #TODO: parallel to be drawn with techniques in [KG-augmented Language Models](tag:knowledge_graph_augmented_language_models) which focus "on the problem of capturing declarative knowledge in the learned parameters of a language model."

[Google AI Blog Post](doc:2020/08/google_ai_blog_realm_integrat)

[Summary](https://joeddav.github.io/blog/2020/03/03/REALM.html) for the [Hugging Face awesome-papers reading group](doc:2021/03/huggingface_awesome_papers_pap)

2020-12-12 About

[2012.04584] Distilling Knowledge from Reader to Retriever for Question Answering

Tags:

> a method to train an information retrieval module for downstream tasks, **without using pairs of queries and documents as annotations**.

Uses two models (standard pipeline for open-domain QA):

- the first one retrieves documents from a large source of knowledge (the retriever)
- the second one processes the support documents to solve the task (the reader).

> First the retriever selects support passages in a large knowledge
source. Then these passages are processed by the reader, along with the question, to generate an
answer

Inspired by knowledge distillation: the reader model is the teacher and the retriever is the student.

> More precisely, we use a sequence-to-sequence model as the reader, and use
the attention activations over the input documents as synthetic labels to train the retriever. 
> (**train the retriever by learning to approximate the attention score of the reader**)

Refers to:

- [REALM: Retrieval-Augmented Language Model Pre-Training](doc:2020/12/2002_08909_realm_retrieval_a)
- [Dehghani: Neural Ranking Models with Weak Supervision](doc:?uri=https%3A%2F%2Farxiv.org%2Fabs%2F1704.08803)

2020-12-11 About

Tags:

2020-07-10 About

Ranked Entities in Search Results at Google

Tags:

2020-07-02 About

Information Retrieval for HR (2018)

Tags:

2020-02-14 About

Building a Search Engine with BERT and TensorFlow - Towards Data Science

Tags:

2020-01-12 About

Tags:

2020-01-10 About

[1602.01137] A Dual Embedding Space Model for Document Ranking

Tags:

2019-07-17 About

Paris NLP Meetup #6

Tags:

2018-07-26 About

SIGIR 2018 Tutorial - Knowledge Extraction and Inference from Text: Shallow, Deep, and Everything in Between

Tags:

2018-07-09 About

A Tri-Partite Neural Document Language Model for Semantic Information Retrieval (2018 - ESWC conference)

Tags:

2018-06-08 About

Modèle Neuronal de Recherche d'Information Augmenté par une Ressource Sémantique (2017)

Tags:

2018-02-12 About

Evaluating the Impact of Word Embeddings on Similarity Scoring in Practical Information Retrieval (2017)

Tags:

> Transferring the success of word embeddings to Information Retrieval (IR) task is currently an active research topic. While embedding-based retrieval models could tackle the vocabulary mismatch problem by making use of the embedding’s inherent similarity between distinct words, most of them struggle to compete with the prevalent strong baselines such as TF-IDF and BM25.

Considering a practical ad-hoc IR task composed of two steps, matching and scoring, compares the performance of several techniques that leverage word embeddings in the retrieval models to compute the similarity between the query and the documents (namely word centroid similarity, paragraph vectors, Word Mover’s distance, as well as a novel inverse document frequency (IDF) re-weighted word centroid similarity).

> We confirm that word embeddings can be successfully employed in a practical information retrieval setting. The proposed cosine similarity of IDF re-weighted, aggregated word vectors is competitive to the TF-IDF baseline.

2018-01-28 About

Learning Deep Structured Semantic Models for Web Search using Clickthrough Data - Microsoft Research (2013)

Tags:

2017-12-30 About

DSSM ("Deep Semantic Similarity Model") - Microsoft Research

Tags: