Semanlink - Fine-tuning AND Domain adaptation

> 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

Domain adaptation of word embeddings through the exploitation of in-domain corpora and knowledge bases (PhD Thesis 2021)

Tags:

Thèse de Hicham El Boukkouri, univ Paris-Saclay

[Github](https://github.com/helboukkouri/phd-code)

### Goal

Given a target specialized domain, improve the quality of general-domain
word representations using in-domain corpora and/or knowledge bases

### Contributions

#### a method for specializing general-domain embeddings in a [Low-Resource](tag:nlp_low_resource_scenarios) context.

> - train static representations on the task corpus, 
> - resume the
pre-training of general-domain contextual embeddings on the same task corpus,
> - finally, combine both static and contextual representations into
one final model

#### we tackle the issue of using a general-domain vocabulary in a specialized domain

#### Evaluation of  re-training vs training from scratch on specialized corpora using a specialized vocabulary

training from scratch better, but not that much: re-training from a general model
is still appropriate as it is less expensive and leads to comparable, although
slightly lower, performance

#### Regarding subword-based tokenization systems
> we argue that they are inconvenient in practice -> CharacterBERT, a variant
of BERT that uses ELMo’s character-based system instead of WordPieces. More convenient ti use, superior robustness to misspellings

#### Ways to specialize general-domain representations using knowledge bases

a strong baseline using a simple
method relying on graph embeddings and concatenation, using only is_a relation

> both static and contextual embeddings may effectively be
specialized using this simple approach

#### Knowledge Injection Modules (KIM) that
inject the knowledge representations directly within the BERT-like models' architecture

### Notes

> our experiments focused on a single
setting (i.e. the medical domain and the English language)

> meta-embeddings, an approach that consists in combining
different sets of representations for achieving improved performance

2022-03-23 A propos

NAVER LABS Europe : "@Nils_Reimers of @huggingface on 'Unsupervised domain adaptation for neural search'"

Tags:

2022-03-09 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

Unsupervised Training for Sentence Transformers | Pinecone

Tags:

2021-11-24 A propos

[1908.11860] Adapt or Get Left Behind: Domain Adaptation through BERT Language Model Finetuning for Aspect-Target Sentiment Classification

Tags:

2021-10-21 A propos

[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 A propos

[2004.10964] Don't Stop Pretraining: Adapt Language Models to Domains and Tasks

Tags:

2020-12-01 A propos