Semanlink - Discuté avec Raphaël ; Arxiv Doc AND Entity embeddings

[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

[2010.03496] 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

[2001.09522] TaxoExpan: Self-supervised Taxonomy Expansion with Position-Enhanced Graph Neural Network

Tags:

how to add a set of new concepts to an existing taxonomy.

[Tweet](https://twitter.com/mickeyjs6/status/1253772146142216194?s=20) [GitHub](https://github.com/mickeystroller/TaxoExpan)

> we study the taxonomy expansion task: given an
existing taxonomy and a set of new emerging concepts, we aim
to automatically expand the taxonomy to incorporate these new
concepts (without changing the existing relations in the given taxonomy).

> To the best of our knowledge, this is the first study on **how to
expand an existing directed acyclic graph (as we model a taxonomy
as a DAG) using self-supervised learning**.

Self-supervised framework, the existing taxonomy being used as training data: it learns a model to predict whether a query concept is the direct hyponym of an anchor concept.

> 2 techniques:
>
> 1. a **position-enhanced graph neural network that encodes the local structure of an anchor concept** in the existing taxonomy,
> 2. a noise-robust training objective that enables the learned model to be insensitive to the label noise in the self-supervision data.

Regarding 1: uses [GNN](/tag/graph_neural_networks.html) to model the "ego network" of concepts (potential “siblings”
and “grand parents” of the query concept).

> Regular
GNNs fail to distinguish nodes with different relative positions to
the query (i.e., some nodes are grand parents of the query while
the others are siblings of the query). To address this limitation, we
present a simple but effective enhancement to inject such position
information into GNNs using position embedding. We show that
such embedding can be easily integrated with existing GNN architectures
(e.g., [GCN](/tag/graph_convolutional_networks) and GAT) and significantly boosts the
prediction performance

Regarding point 2: uses InfoNCE loss, cf. [Contrastive Predictive Coding](/doc/?uri=https%3A%2F%2Farxiv.org%2Fabs%2F1807.03748)

> Instead of predicting
whether each individual ⟨query concept, anchor concept⟩ pair
is positive or not, we first group all pairs sharing the same query
concept into a single training instance and learn a model to select
the positive pair among other negative ones from the group.

(Hum, ça me rappelle quelque chose)

> assume each concept (in existing taxonomy + set of new concepts) has an initial embedding
vector learned from some text associated with this concept.

To keep things tractable, only attempts to find a single parent node of each new concept.

2020-04-25 About