Semanlink - Arxiv Doc ; Representation learning AND Unsupervised machine learning

[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 About

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

Tags:

2022-04-01 About

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

Tags:

2022-03-31 About

[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 About

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

Tags:

2022-02-25 About

[2112.09118] Towards Unsupervised Dense Information Retrieval with Contrastive Learning

Tags:

2021-12-21 About

[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 About

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

Tags:

2021-09-06 About

[2104.06979] TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning

Tags:

> The most
successful previous approaches like InferSent (Conneau
et al., 2017), Universial Sentence Encoder
(USE) (Cer et al., 2018) and SBERT (Reimers and
Gurevych, 2019) heavily relied on labeled data to
train sentence embedding models.
>
> TSDAE can
achieve up to 93.1% of the performance of indomain
supervised approaches. Further, we
show that TSDAE is **a strong domain adaptation
and pre-training method for sentence
embeddings**, significantly outperforming other
approaches like Masked Language Model.

> During training, TSDAE
encodes corrupted sentences into fixed-sized
vectors and requires the decoder to reconstruct the
original sentences from this sentence embedding.

- <https://www.sbert.net/examples/unsupervised_learning/TSDAE/README.html>
- [github](https://github.com/UKPLab/sentence-transformers/tree/master/examples/unsupervised_learning/TSDAE)
- [UKPLab/sentence-transformers: Sentence Embeddings with BERT & XLNet](doc:2020/07/ukplab_sentence_transformers_s)
- [twitter](https://twitter.com/KexinWang2049/status/1433361957579538432):

> **TSDAE can learn domain-specific sentence embeddings with unlabeled sentences**
>
> Most importantly, instead of STS (Semantic Textual Similarity), **we suggest evaluating unsupervised sentence embeddings on the domain-specific tasks&datasets, which is the real use case for them**. Actually, STS scores do not correlate with performance on specific tasks.

2021-09-01 About

[2107.00676] A Primer on Pretrained Multilingual Language Models

Tags:

2021-07-13 About

[2103.12953] Supporting Clustering with Contrastive Learning

Tags:

2021-05-20 About

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

Tags:

2021-04-19 About

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

Tags:

2021-04-12 About

[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 About

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

Tags:

2020-11-26 About

[1911.06136] KEPLER: A Unified Model for Knowledge Embedding and Pre-trained Language Representation

Tags:

2020-11-03 About

[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 About

[2010.00402] From Trees to Continuous Embeddings and Back: Hyperbolic Hierarchical Clustering

Tags:

2020-10-03 About

[1909.04164] Knowledge Enhanced Contextual Word Representations

Tags:

2020-05-13 About

[1908.10084] Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Tags:

2019-08-28 About

[1905.07129] ERNIE: Enhanced Language Representation with Informative Entities

Tags:

2019-08-05 About

[1906.02715] Visualizing and Measuring the Geometry of BERT

Tags:

2019-06-07 About

[1902.11269] Efficient Contextual Representation Learning Without Softmax Layer

Tags:

**how to accelerate contextual representation learning**.

> Contextual representation models are difficult to train due to the large parameter sizes and high computational complexity

> We find that the softmax layer (the output layer) causes significant inefficiency due to the large vocabulary size.
Therefore, we redesign the learning objectiv.
> Specifically, the proposed approach bypasses the softmax layer by performing language modeling with dimension reduction, and allows the models to leverage pre-trained word embeddings.
Our framework reduces the time spent on the output layer to a negligible level, eliminates almost all the trainable parameters of the softmax layer and performs language modeling without truncating the vocabulary.
When applied to ELMo, our method achieves a 4 times speedup and eliminates 80% trainable parameters while achieving competitive performance on downstream tasks.

**decouples learning contexts and words**

> Instead of using
a softmax layer to predict the distribution of the
missing word, we utilize and extend the SEMFIT
layer (Kumar and Tsvetkov, 2018) to **predict the
embedding of the missing word**.

2019-03-02 About

[1511.06335] Unsupervised Deep Embedding for Clustering Analysis

Tags:

2019-02-19 About

[1811.05370] Unsupervised Transfer Learning for Spoken Language Understanding in Intelligent Agents

Tags:

2018-11-20 About

[1601.03764] Linear Algebraic Structure of Word Senses, with Applications to Polysemy

Tags:

> Here it is shown that multiple word senses reside
in linear superposition within the word
embedding and simple sparse coding can recover
vectors that approximately capture the
senses

> Each extracted word sense is accompanied by one of about  2000 “discourse atoms” that gives a succinct description of which other words co-occur with that word sense.

> The success of the approach is mathematically explained using a variant of
the random walk on discourses model

("random walk": a generative model for language). Under the assumptions of this model,  there
exists a linear relationship between the vector of a
word w and the vectors of the words in its contexts (It is not the average of the words in w's context, but in a given corpus the matrix of the linear relationship does not depend on w. It can be estimated, and so we can compute the embedding of a word from the contexts it belongs to)

[Related blog post](/doc/?uri=https%3A%2F%2Fwww.offconvex.org%2F2016%2F07%2F10%2Fembeddingspolysemy%2F)

2018-08-28 About

[1807.03748] Representation Learning with Contrastive Predictive Coding

Tags:

2018-07-21 About

[1806.06259] Evaluation of sentence embeddings in downstream and linguistic probing tasks

Tags:

2018-06-19 About

[1710.04087] Word Translation Without Parallel Data

Tags:

2017-10-14 About