Semanlink - Arxiv Doc AND NLP@Google

> **approach for training dense retrieval models that does not require any labeled training data**. Dense retrieval is a central challenge for open-domain tasks, such as Open QA, where state-of-the-art methods typically require large supervised datasets with custom hard-negative mining and denoising of positive examples.
>
> ART, in contrast, only requires access to unpaired inputs and outputs (e.g. questions and potential answer documents).
>
> It uses a new document-retrieval autoencoding scheme, where
> 1. an input question is used to retrieve a set of evidence documents, and
> 2. the documents are then used to compute the probability of reconstructing the original question.
>
> Training for retrieval based on question reconstruction enables effective unsupervised learning of both document and question encoders, which can be later incorporated into complete Open QA systems without any further finetuning.

[Tweet](doc:2022/07/devendra_singh_sachan_sur_twitt)

> Given an
input question, ART first retrieves a small set
of possible evidences documents. It then recon
structs
the original question by attending to these
documents
>
> The
key idea in ART is to consider the retrieved documents
as a noisy representation of the original
question and question reconstruction probability
as a way of denoising that provides soft-labels for
how likely each document is to have been the correct
result

Refers to [[IZACARD 2012.04584] Distilling Knowledge from Reader to Retriever for Question Answering](doc:2020/12/2012_04584_distilling_knowled)

2022-07-06 A propos

[2205.15952] Knowledge Graph -- Deep Learning: A Case Study in Question Answering in Aviation Safety Domain

Tags:

2022-06-11 A propos

[2205.08184] SKILL: Structured Knowledge Infusion for Large Language Models

Tags:

2022-05-18 A propos

[2205.05131] Unifying Language Learning Paradigms

Tags:

2022-05-12 A propos

[2205.04260] EASE: Entity-Aware Contrastive Learning of Sentence Embedding

Tags:

2022-05-11 A propos

[2203.08913] Memorizing Transformers

Tags:

2022-05-07 A propos

[1909.00426] Global Entity Disambiguation with BERT

Tags:

2022-04-18 A propos

[2110.08151] mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models

Tags:

[Ikuya Yamada sur Twitter : "Is entity representation effective to improve multilingual language models?..."](doc:2022/04/ikuya_yamada_sur_twitter_is_)

> Recent studies have shown that multilingual pretrained language models can be effectively improved with cross-lingual alignment information from Wikipedia entities. However, **existing methods only exploit entity information in pretraining and do not explicitly use entities in downstream tasks**. In this study, we explore the **effectiveness of leveraging entity representations for downstream cross-lingual tasks**.
>
> the key insight is that incorporating entity representations into the input allows us to extract more language-agnostic features.

[Github](https://github.com/studio-ousia/luke)

> Entity representations are known to enhance
language models in mono-lingual settings
(Zhang et al., 2019: [ERNIE](tag:ernie.html); Peters et al., 2019:  [[1909.04164] Knowledge Enhanced Contextual Word Representations](doc:2020/05/1909_04164_knowledge_enhanced); Wang et al.,
2021 [[1911.06136] KEPLER: A Unified Model for Knowledge Embedding and Pre-trained Language Representation](doc:2020/11/1911_06136_kepler_a_unified_); Xiong et al., 2020; Yamada et al., 2020: [[2010.01057] LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](doc:2020/11/2010_01057_luke_deep_context))
presumably by introducing real-world knowledge.
We show that using entity representations facilitates
cross-lingual transfer by providing languageindependent
features.
>
> Multilingual extension of LUKE. The model is trained with the multilingual
masked language modeling (MLM) task as well
as the masked entity prediction (MEP) task with
Wikipedia entity embeddings

> We investigate two ways of using the entity representations
in cross-lingual transfer tasks:
> 1. perform
entity linking for the input text, and append
the detected entity tokens to the input sequence.
The entity tokens are expected to provide language independent
features to the model
> 2. use the entity
[MASK] token from the MEP task as a languageindependent
feature extractor.

2022-04-17 A propos

[2008.11228] A simple method for domain adaptation of sentence embeddings

Tags:

2022-04-01 A propos

[2004.05119] Beyond Fine-tuning: Few-Sample Sentence Embedding Transfer

Tags:

2022-03-31 A propos

[2203.14655] Few-Shot Learning with Siamese Networks and Label Tuning

Tags:

> the problem of building text classifiers with little or no training data.
>
> In recent years, an approach based on neural textual entailment models has been found to give strong results on a diverse range of tasks.

(cf. #[NLI](tag:nli), using the input text as the premise and the text representing the label as the hypothesis)

> In this work, we show that **with proper pre-training, Siamese Networks that embed texts and labels** offer a competitive alternative.
>
> We introduce **label tuning: fine-tuning the label embeddings only**. While giving lower performance than model fine-tuning (which updates all params of the model), this approach has the architectural advantage that a single encoder can be shared by many different tasks (we only fine-tune the label embeddings)
> The drop in quality can
be compensated by using a variant of **[Knowledge distillation](tag:knowledge_distillation)**

[Github](https://tinyurl.com/label-tuning), [Tweet](doc:2022/03/thomas_muller_sur_twitter_pa)

2022-03-30 A propos

[2006.05987] Revisiting Few-sample BERT Fine-tuning

Tags:

2022-03-21 A propos

[2109.06304] Phrase-BERT: Improved Phrase Embeddings from BERT with an Application to Corpus Exploration

Tags:

2022-02-25 A propos

[1906.00300] Latent Retrieval for Weakly Supervised Open Domain Question Answering

Tags:

2022-01-11 A propos

[2112.07577] GPL: Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval

Tags:

An unsupervised domain adaptation technique for dense retrieval models

1. synthetic queries
are generated for each passage from the target corpus (using an existing pre-trained [T5](tag:text_to_text_transfer_transformer)
encoder-decoder)
2. the generated queries are used for mining negative
passages (retrieving the most similar
paragraphs using an existing dense retrieval
model == hard negatives!)
3. the query-passage pairs are labeled by a cross-encoder and used to train the domain-adapted
dense retriever (using method described in [Hofstätter et al.,
2020](doc:2021/12/2010_02666_improving_efficien))

[Nils Reimers sur Twitter](doc:2021/12/nils_reimers_sur_twitter_do_), [GitHub](https://github.com/UKPLab/gpl),  by the author of [TSDAE](doc:2021/09/2104_06979_tsdae_using_trans)

Claims to improve "Doc2Query" [Document Expansion by Query Prediction](doc:2022/01/1904_08375_document_expansion): ([src](https://twitter.com/KexinWang2049/status/1471435779415150598))

> - GPL: Uses doc2query to construct synthetic data and does knowledge distillation (i.e. training) on that data.
> - Doc2query: Generates queries to extend the documents and use BM25 on top of them w/o training.

2021-12-15 A propos

[2106.04647] Compacter: Efficient Low-Rank Hypercomplex Adapter Layers

Tags:

2021-09-29 A propos

[2010.12566] DICT-MLM: Improved Multilingual Pre-Training using Bilingual Dictionaries

Tags:

2021-09-06 A propos

[2102.07043] Reasoning Over Virtual Knowledge Bases With Open Predicate Relations

Tags:

> a method for constructing **a virtual KB (VKB) trained entirely from text**

Open Predicate Query Language (OPQL): constructing a virtual knowledge base (VKB) that supports KB reasoning & open-domain QA, tackling the incompleteness of knowledge bases by constructing a virtual KB only from text

> OPQL constructs
a VKB by **encoding and indexing a set of
relation mentions** in a way that naturally enables
reasoning and can be trained without any structured
supervision.

> can be used
as an **external memory integrated into a language
model**

cf. this earlier paper [[2002.10640] Differentiable Reasoning over a Virtual Knowledge Base](doc:2020/07/2002_10640_differentiable_rea). But does not require an initial structured KB for distant
supervision.

> The key idea in constructing the OPQL VKB is to use a
dual-encoder pre-training process, similar to 
[[1906.03158] Matching the Blanks: Distributional Similarity for Relation Learning](doc:2021/05/1906_03158_matching_the_blank)

Related work section refers to [[1909.04164] Knowledge Enhanced Contextual Word Representations](doc:2020/05/1909_04164_knowledge_enhanced). Also refers to [[2007.00849] Facts as Experts: Adaptable and Interpretable Neural Memory over Symbolic Knowledge](doc:2020/07/2007_00849_facts_as_experts_) (some authors in common)

2021-06-20 A propos

[1906.03158] Matching the Blanks: Distributional Similarity for Relation Learning

Tags:

2021-05-13 A propos

[1909.10506] Learning Dense Representations for Entity Retrieval

Tags:

> We show that it is feasible to perform **entity
linking by training a dual encoder (two-tower)
model that encodes mentions and entities in
the same dense vector space**, where candidate
entities are retrieved by approximate nearest
neighbor search. Unlike prior work, **this setup
does not rely on an alias table followed by a
re-ranker, and is thus the first fully learned entity
retrieval model**.

Contributions:

> -  a dual encoder architecture for
learning entity and mention encodings suitable for
retrieval. A key feature of the architecture is that it
employs a modular **hierarchy of sub-encoders that
capture different aspects of mentions and entities**
> - a simple, fully unsupervised **hard negative
mining** strategy that produces massive gains
in retrieval performance, compared to using only
random negatives
> - high
quality candidate entities very efficiently using approximate nearest neighbor search
> - outperforms discrete retrieval
baselines like an alias table or BM25

> strong retrieval
performance across all 5.7 million Wikipedia entities in
around 3ms per mention

> since we are using a two-tower or dual
encoder architecture, **our model cannot use any kind of attention over
both mentions and entities at once**, nor feature-wise
comparisons as done by Francis-Landau et al. (2016).
This is a fairly severe constraint – for example, **we cannot
directly compare the mention span to the entity title**
– but it permits retrieval with nearest neighbor search
for the entire context against a single, all encompassing
representation for each entity

2021-05-01 A propos

[2011.05864] On the Sentence Embeddings from Pre-trained Language Models

Tags:

2021-04-19 A propos

[2007.12603] IR-BERT: Leveraging BERT for Semantic Search in Background Linking for News Articles

Tags:

2021-04-12 A propos

[2007.15779] Domain-Specific Language Model Pretraining for Biomedical Natural Language Processing

Tags:

2021-04-11 A propos

[1902.00751] Parameter-Efficient Transfer Learning for NLP

Tags:

2021-04-11 A propos

[1901.04085] Passage Re-ranking with BERT

Tags:

2021-03-26 A propos

[1911.03681] E-BERT: Efficient-Yet-Effective Entity Embeddings for BERT

Tags:

> way of **injecting factual knowledge about entities into the pretrained BERT model**.

(Feeding entity vectors
into BERT as if they
were wordpiece vectors without additional encoder
pretraining)

>
> **We align [Wikipedia2Vec](tag:wikipedia2vec) entity vectors (Yamada et al., 2016) with BERT's native wordpiece vector space and use the aligned entity vectors as if they were wordpiece vectors**. The resulting entity-enhanced version of BERT (called E-BERT) is similar in spirit to [ERNIE](tag:ernie) (Zhang et al., 2019) and [KnowBert](tag:knowbert) (Peters et al., 2019), but it **requires no expensive further pretraining of the BERT encoder**.
>
> Our vector space alignment strategy is inspired by
cross-lingual word vector alignment

Related work on Entity-enhanced BERT:

> ([ERNIE](doc:2019/08/_1905_07129_ernie_enhanced_la) and [Knowbert](doc:2020/05/1909_04164_knowledge_enhanced)) are based on the design principle
that BERT be adapted to entity vectors. They introduce
new encoder layers to feed pretrained entity
vectors into the Transformer, and they require additional
pretraining to integrate the new parameters.
In contrast, E-BERT’s design principle is that entity
vectors be adapted to BERT.
>
> Two other knowledge-enhanced MLMs are [KEPLER](doc:2020/11/1911_06136_kepler_a_unified_)
(Wang et al., 2019c) and K-Adapter (Wang
et al., 2020)... Their factual knowledge
does not stem from entity vectors – instead, they
are trained in a multi-task setting on relation classification
and knowledge base completion.

Not to be cofounded with [[2009.02835] E-BERT: A Phrase and Product Knowledge Enhanced Language Model for E-commerce](doc:2020/12/2009_02835_e_bert_a_phrase_a)

2021-01-12 A propos

[2009.02835] E-BERT: A Phrase and Product Knowledge Enhanced Language Model for E-commerce

Tags:

2020-12-14 A propos

[2002.08909] REALM: Retrieval-Augmented Language Model Pre-Training

Tags:

**Augment language model pre-training with a retriever module**, which
is trained using the masked language modeling objective.

> 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 A propos

[2010.01057] LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention

Tags:

2020-11-26 A propos

[2010.03496] Inductive Entity Representations from Text via Link Prediction

Tags:

BLP "BERT for Link Prediction". Central idea: **training an entity encoder with a
link prediction objective** (using the textual descriptions of entities when computing entity representations - hence not failing with entities unknown in training)

> a method for **learning representations
of entities**, that uses a **pre-trained Transformer** based
architecture as an entity encoder, and
**link prediction training on a knowledge graph
with textual entity descriptions**.

> using entity descriptions,
an entity encoder is trained for link prediction in
a knowledge graph. The encoder can then be used
without fine-tuning to obtain features for entity classification
and information retrieval

Cites [Xie et al](doc:2020/10/representation_learning_of_know) and [Kepler](doc:2020/11/1911_06136_kepler_a_unified_). They claim that their
objective targeted exclusively for link prediction (and not an objective that combines language modeling
and link prediction as Kepler)
performs better than Kepler's more complex one.

2020-11-03 A propos

[2004.07202] Entities as Experts: Sparse Memory Access with Entity Supervision

Tags:

2020-07-11 A propos

[2007.00849] Facts as Experts: Adaptable and Interpretable Neural Memory over Symbolic Knowledge

Tags:

> a neural language model that includes **an explicit interface between symbolically interpretable factual information and subsymbolic neural knowledge.**... **The model can be updated without re-training by manipulating its symbolic representations**. In particular this model allows us to add new facts and overwrite existing ones.

> a **neural language model which learns to access information
in a symbolic knowledge graph.**

> This
model builds on the recently-proposed [Entities as
Experts](doc:2020/07/2004_07202_entities_as_expert) (EaE) language model (Févry et al., 2020),
which extends the same transformer (Vaswani
et al., 2017) architecture of BERT (Devlin et al., 2019) with an additional external memory for entities.
>
> After training EaE, the embedding associated
with an entity will (ideally) capture information
about the textual context in which that
entity appears, and by inference, the entity’s semantic
properties
>
> we include an additional
memory called a fact memory, which encodes
triples from a symbolic KB.
>
> This combination results in a
neural language model which learns to access information
in a the symbolic knowledge graph.

TODO:

- read again IBM's [Span Selection Pre-training for Question Answering](doc:2019/09/_1909_04120_span_selection_pre) ("an effort to avoid encoding general knowledge in the transformer network itself")
- compare with [[1907.05242] Large Memory Layers with Product Keys](doc:2019/07/_1907_05242_large_memory_layer)
- how does it relate with [[2002.08909] REALM: Retrieval-Augmented Language Model Pre-Training](doc:2020/12/2002_08909_realm_retrieval_a)?

2020-07-09 A propos

[1909.04164] Knowledge Enhanced Contextual Word Representations

Tags:

2020-05-13 A propos

[1911.03814] Scalable Zero-shot Entity Linking with Dense Entity Retrieval

Tags:

2020-05-02 A propos

[1909.03193] KG-BERT: BERT for Knowledge Graph Completion

Tags:

2020-03-22 A propos

[1909.07606] K-BERT: Enabling Language Representation with Knowledge Graph

Tags:

2020-03-08 A propos

[2002.12327] A Primer in BERTology: What we know about how BERT works

Tags:

2020-02-28 A propos

[2002.11402] Detecting Potential Topics In News Using BERT, CRF and Wikipedia

Tags:

2020-02-27 A propos

[2002.02925] BERT-of-Theseus: Compressing BERT by Progressive Module Replacing

Tags:

2020-02-10 A propos

[2003.05473] Investigating Entity Knowledge in BERT with Simple Neural End-To-End Entity Linking (CoNNL 2019)

Tags:

Training BERT-base-uncased on English Wikipedia and then fine-tuned and evaluating it
on an entity linking (EL) benchmark (EL implemented as a token classification over the entity vocabulary)

> BERT+Entity is a straightforward extension on top
of BERT, i.e. we initialize BERT with the publicly
available weights from the BERT-base-uncased
model and add an output classification layer on
top of the architecture. Given a contextualized token,
the classifier computes the probability of an
entity link for each entry in the entity vocabulary.

Can BERT’s architecture learn all entity
linking steps jointly? To answer:

> an extreme
simplification of the **entity linking setup that
works surprisingly well**: simply cast it as **a
per token classification over the entire entity
vocabulary** (over 700K classes in our case).

> the model
is the first that performs entity linking without any
pipeline or any heuristics, compared to all prior
approaches. We found that with our approach we
can learn additional entity knowledge in BERT that
helps in entity linking. **However, we also found
that almost none of the downstream tasks really
required entity knowledge**.

### Related work

- > [Durrett and Klein (2014)](/doc/2020/01/a_joint_model_for_entity_analys) were the first to propose
jointly modelling Mention detection, Candidate generation and Entity disambiguation in a graphical
model and could show that each of those steps are
interdependent and benefit from a joint objective

This paper uses neural techniques instead of CRF.

- > [Yamada](/showprop.do?pptyuri=http%3A%2F%2Fwww.semanlink.net%2F2001%2F00%2Fsemanlink-schema%23arxiv_author&pptyval=Ikuya%2BYamada) (2016, 2017) was the first to
investigate neural text representations and entity
linking, but their approach is limited to ED.

cf. [#Wikipedia2Vec](tag:wikipedia2vec). Compare with [newer work by Yamada](doc:2020/09/1909_01259_neural_attentive_b)

2020-01-09 A propos

[1902.10909] BERT for Joint Intent Classification and Slot Filling

Tags: