Semanlink - Good AND Deep Learning

Zeyuan Allen-Zhu sur X : " surprisingly, when pre-training good data (e.g., Wiki) together with "junks" (e.g., Common Crawl), LLM's capacity on good data may decrease by 20x times!"

Tags:

2024-04-10 A propos

[2305.12517] Retrieving Texts based on Abstract Descriptions

Tags:

2023-06-15 A propos

Google "We Have No Moat, And Neither Does OpenAI"

Tags:

> low-cost public involvement was enabled
by a vastly cheaper mechanism for fine tuning called low
rank adaptation ()[LoRA](tag:lora)

> **Part of what makes LoRA so effective is that ... it’s stackable.**
>
> By contrast, training giant models from scratch not only
throws away the pretraining, but also any iterative
improvements that have been made on top.

> LoRA updates are very cheap to produce (~$100) for the
most popular model sizes.

> Many of these projects are saving time by training on
small, highly curated datasets...
> These
datasets are built using synthetic methods (e.g. filtering
the best responses from an existing model) and
scavenging from other projects

> Directly Competing With Open Source
Is a Losing Proposition

> Paradoxically, the one clear winner in all of this is Meta.
Because the leaked model was theirs ([LLaMA](tag:llama)), they have
effectively garnered an entire planet's worth of free labor.
Since most open source innovation is happening on top of
their architecture, there is nothing stopping them from
directly incorporating it into their products.

2023-05-04 A propos

[2106.13474] Adapt-and-Distill: Developing Small, Fast and Effective Pretrained Language Models for Domains

Tags:

2021-10-21 A propos

[2110.06176] Mention Memory: incorporating textual knowledge into Transformers through entity mention attention

Tags:

2021-10-13 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

[1909.01259] Neural Attentive Bag-of-Entities Model for Text Classification

Tags:

A model that performs **text classification using entities in a knowledge base**.

> Entities provide unambiguous and relevant semantic signals that are beneficial for capturing semantics in texts. We combine **simple high-recall entity detection based on a dictionary** (word->list of entities), to detect entities in a document, with a novel neural **attention mechanism that enables the model to focus on a small number of unambiguous and relevant entities**.

2 steps:

1. Entity detection
2. Classification using the detected entities (+text) as inputs

Regarding entity linking, a local model which uses cosine
similarity between the embedding of the target
entity and the word-based representation of
the document to capture the relevance of an entity
given a document.

Embeddings from the KB: computed using [#Wikipedia2Vec](tag:wikipedia2vec) (similar words and entities
close to one another in a unified vector space)

Model using attention, with 2 features :

- cosine similarity between the
embedding of the entity and the word based
representation of the document
- the probability that the entity
name refers to the entity in KB.

Somewhat [related](doc:2020/01/investigating_entity_knowledge_)

### Conclusion:

>a neural
network model that performs text classification using
entities in Wikipedia. We combined simple
dictionary-based entity detection with a neural attention
mechanism to enable the model to focus
on a small number of unambiguous and relevant
entities in a document.

2020-09-02 A propos

[1909.04164] Knowledge Enhanced Contextual Word Representations

Tags:

2020-05-13 A propos

Semantic Text Matching for Long-Form Documents (2019)

Tags:

2020-01-23 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

The knowledge graph as the default data model for learning on heterogeneous knowledge (2017)

Tags:

2019-08-22 A propos

Transformers from scratch | Peter Bloem

Tags:

2019-08-21 A propos

Romain Vial (Hyperlex) at Paris NLP meetup, slides

Tags:

2019-01-24 A propos

Learning Deep Architectures for AI By Yoshua Bengio (2009)

Tags:

2018-11-06 A propos

[1801.06146] Universal Language Model Fine-tuning for Text Classification

Tags:

2018-01-19 A propos

[1709.08568] The Consciousness Prior

Tags:

2017-09-29 A propos

Attention and Memory in Deep Learning and NLP – WildML

Tags:

cf. visual attention

In standard [#seq2seq](/tag/sequence_to_sequence_learning) NMT, the decoder is supposed to generate a translation solely based on the last hidden state of the encoder - which therefore must capture everything from the source sentence (it must be a sentence embedding). Not good. Hence the attention mechanism.

> we allow the decoder to “attend” to different parts of the source sentence at each step of the output generation. Importantly, we let the model learn what to attend to based on the input sentence and what it has produced so far

> each decoder output word now depends on a weighted combination of all the input states, not just the last state.

Possible to interpret what the model is doing by looking at the Attention weight matrix

Cost: We need to calculate an attention value for each combination of input and output word (-> attention is a bit of a misnomer: we look at everything in details before deciding what to focus on)

> attention mechanism is simply giving the network access to its internal memory, which is the hidden state of the encoder

> Unlike typical memory, the memory access mechanism here is soft, which means that the network retrieves a weighted combination of all memory locations, not a value from a single discrete location

2016-01-03 A propos

Deep Learning, NLP, and Representations - colah's blog

Tags:

2015-10-16 A propos