SURE: A Sentence encoding based approach for Unsupervised Relation Extraction

SURE is an unsupervised system for relationship extraction relying on sentence vector representations (i.e., sentence enocoding). For more details please refer to: The paper

Architecture: system description

---

Dependencies

You need to have Python 3.6 or above and the following libraries installed:

NLTK: http://www.nltk.org/

sKlearn: https://scikit-learn.org/stable/

Numpy: https://numpy.org/

Sentence-Transformers: https://www.sbert.net/

Pandas: https://pandas.pydata.org/

Pytorch: https://pytorch.org/

which you can install issuing the following command:

pip install -r requirements.txt

Usage:

To run the relation extraction system use the following command:

python main.py corpus entity_type1 entity_type2

Config

between_length: 6             # Maximum number of tokens between two entities       
before_after_window: 3        # Maximum number of tokens before the first entity and maximum number of tokens after second entity
similiraty: 0.25              # Cosine similirity threshold during the first iteration
top_similar: 15               # Maximum number of top similar sentences to the query term using cosine similarity
query_term: born in           # Natural language representation for relationship birthPlace

corpus

A sample sentence in the corpus is one sentence per line, with tags identifing the named type of named-entities, e.g.:

<ORG> Consolidated Edison </ORG>, based in <LOC> New York </LOC>, generated more than $7 billion in annual revenue.
The social media platform <ORG> Facebook,Inc.</ORG> announced it was acquiring <ORG>WhatsApp</ORG>, its largest acquisition to date.
<LOC> Herzogenaurach </LOC> is the home of goods company <ORG> Adidas<ORG>.

entity types

Two named entity types are provided initally for a particular relation for example, for rleation headquarterIn we provide ORG (Organization) and LOC (Location).

Dataset used

Dataset	Download
NYT-FB dataset	Download
Wikipedia_Wikidata dataset	Download
English gigaword	Download

Run with basic configuration

python main.py Evaluation/Eval_NYT/data/AnnotatedANLP_NYT__sample.txt  PER LOC 2

Authors

• Manzoor Ali (DICE, Paderborn University)
• Muhammad Saleem (AKSW, University of Leipzig)
• Axel-Cyrille Ngonga Ngomo (DICE, Paderborn University)