Add pretrained relation extraction models

[alanakbik]

This PR adds a few relation extraction models, trained over a modified version of TACRED. Two models are added: relations and relations-fast.

To use these models, you also need an entity tagger. The tagger identifies entities, then the relation extractor possible entities. For instance use this code:

from flair.data import Sentence
from flair.models import RelationExtractor, SequenceTagger

# 1. make example sentence
sentence = Sentence("George was born in Washington")

# 2. load entity tagger and predict entities
tagger = SequenceTagger.load('ner-fast')
tagger.predict(sentence)

# check which entities have been found in the sentence
entities = sentence.get_labels('ner')
for entity in entities:
    print(entity)

# 3. load relation extractor
extractor: RelationExtractor = RelationExtractor.load('relations-fast')

# predict relations
extractor.predict(sentence)

# check which relations have been found
relations = sentence.get_labels('relation')
for relation in relations:
    print(relation)

[alejandrojcastaneira]

Thank you very much for adding this, can you provide some insights or a short description of the architecture of the relations extraction models?

[alanakbik]

Hello @alejandrojcastaneira it we used a very simple architecture here: the RelationExtractor goes through all pairwise combinations of entities found in a sentence. For each entity pair, it takes the embeddings of both entities and concatenates them. This concatenated entity pair embedding is then classified.

[SohaK]

@alanakbik does the embedding include some sort of positional embedding? I am asking since I have a use-case that for example in a sentence there are two persons and two birth_places. Can this Relation Extraction model be trained to distinguish between these two birth_place relations based on context of the sentence?

[djstrong]

Yes if you use contextual embeddings.

[SohaK]

@alanakbik @djstrong in my question I meant positional context not semantic context. Is there any paper or blog that shed some light to the FLAIR Relation Extraction?

[djstrong]

Semantic context is usually positional. Flair RE is extracting embeddings of the spans, concats them and pass through fully connected neural network.

[SohaK]

@djstrong is the 'embedding' the dynamic embedding of the span in the current sentence, or the pretrained embedding?

[djstrong]

You can choose embeddings (dynamic or static).