Knowledge Base Population from German Language Text Leveraging Graph Embeddings

1. Präsentation von Marco Lehner

Problem Description

Reporters lose time for investigating topics twice and performing tedious and easily automatable tasks.

Readers having a hard time finding  (background) information on topics of interest.

Newsrooms can't use their collected information in its entirety since it's not machine readable.

Proposed Solution:
Knowledge Base Population

Knowledge base population systems add extracted information to existing Knowledge Bases (KB).

Knowledge Extraction often consists three steps:

  1. Entity Recognition
  2. Relation Extraction
  3. Entity Linking

Challenges

Absence

Extracted Entities are not part of a given KB.

 

Latency

Knowledge needs to be served to users in real-time.

Evolving Information (not tackled)

Change of facts over time needs to be represented in the KB.

Architecture

Background

  • >88 Millionen Knoten
  • >1 Terrabyte
  • Based on OS rdf4j
  • Support of Rank Calculation
  • Low Latency
  • Bulk Loading
  • Optimized for graphs >1B triples

Input

Any text document in german language.

Best results on documents with many entities.

NER

Python wrapper for Stanford Core NLP.

  • Support of 66 human languages
  • High precision in German NER tasks
  • High speed

Candidate Generation

  1. Get surface form of entity
  2. Search KB for matching labels
  3. Create Tupel
    • ID
    • Label
    • Node embedding
    • String similarity
    • PageRank

Local score

}

Candidate Selection

  1. Select the tuple with highest local score for each candidate
  2. Create set of those tuples
  3. Calculate variance of embeddings
  4. Randomly change tuples in the set
  5. If variance is lower, replace old best set
  6. Break after n iterations or converging global scores.

Global Score

Weighted product of embedding variance and local score.

Entity Classification

  1. Identify causing tuple
  2. Mark tuple as NIL
  3. Recalculate set's global score without tuple

(Baseline)

While the set's global score is below a given threshold

Relation Extraction

  1. Identify sentence with more than one entity.
  2. Get dependendy tree of this sentence
  3. Use lowest verb in tree containing two entities as relation.

(Baseline)

KB Population

  1. Add new nodes to graph
  2. Add new relations to graph
  3. Recalculate GraphRank
  4. Recalculate embeddings of new and neighboring nodes.

(Baseline)

Calculation of embeddings

Parravicini uses node2vec to calculate embeddings. It is not possible to get embeddings for new nodes.

 

Therefore GraphSAINT is (probably) used to calculate the embeddings. GraphSAGE trains an encoding function on a subset of the graph which later calculates embeddings for unseen parts of the graph. The structure of the graph has to remain stable.

Todos

  • Implement GraphSAINT
  • Research German relation extraction frameworks
  • Develop evaluation concept

Demo

master.marco-lehner.de

Knowledge Base Population from German Language Text Leveraging Graph Embeddings

By redadmiral

Made with Slides.com

Knowledge Base Population from German Language Text Leveraging Graph Embeddings

  • 55

More from redadmiral