Provide LLM alternative for Relationship/Meaning Extraction

[dominik-pichler]

Instead of spliting the nodes with a LLM, other possibilites (syntactic dissemblement etc.) should be investigated

[dominik-pichler]

maybe libraries like SpaCy, NLTK or Stanford NLP help

[dominik-pichler]

Steps:

1) Coreference Resolution:

Pronounce will be replaced by reference entities.

2) Named Entity Recognition

Identify/Recognize all mentioned entities in the text

Solve Entity Disambiguation & Entity Linking Elon's vs Elon -> is the same but will not be recognized as the same thing

Co-Occurrence Graphs Evaluate the co-occurence of terms in a subtask

3) Relationship Extraction

This could happen in multiple ways:

1. Rule Based with grammatical dependencies/rules (Resource Description Framework)

2. NLP Model (just as we did with the LLM right?), in Addition TACRED or Wiki80 could be helpful.

   found @ https://neo4j.com/blog/text-to-knowledge-graph-information-extraction-pipeline/

[dominik-pichler]

For 1): https://towardsdatascience.com/coreference-resolution-in-python-aca946541dec#:~:text=Coreference%20resolution%20is%20the%20NLP,to%20the%20same%20underlying%20entities.

[dominik-pichler]

Had a problem with coreference lib neuralcoref due to version conflicts. Due to time constraints, I'll take care of this last

[dominik-pichler]

2) Where is the meaning of the sentences hidden? Or is it enough to ask where the ideas are hidden? Actually, I would first need a good working definition of ideas. In this context, ideas are to be understood as concepts that a subject creates in order to make decisions for action. In other words, a premise on the basis of which the world is interpreted. This interpretation is then the foundation on which action decisions are made

In this case, it is a matter of extracting acting relationships. In particular, relationships from (subject-verb-subject)^n. Example: X greets Z)

To be separated from sentient relations ( X feels y about Z) which should come in the next step!

Problem: How do you depict actions spanning across multiple sentences ?

Resources here: https://towardsdatascience.com/named-entity-recognition-with-nltk-and-spacy-8c4a7d88e7da

[dominik-pichler]

Question: In the english language: Is the gramar structured in order? Meaning, give the following Wordtpyes: S1 - a1 - S2 - a2 - S3, could a1 also effect S3 or is it always only affecting S2.

Meaning, what are our chunk patterns ?

[dominik-pichler]

also, the question is, wether chunked subjects (including their attributes in the nodes) should be used ? It would make things easier if I could just work with NP & VDBs, this will likely increase the number of nodes ...

[dominik-pichler]

Quick train of thought: Is it not enough to extract the nouns and link them via sentences ?

The question is, if it is smarter to work with NP's or pure NNs? or, if I can maybe get both ? 🌊

[dominik-pichler]

or to have multiple graphs:

1) NP 2) NNs without attributes 3) NNs with attributes

[dominik-pichler]

Question: In the english language: Is the gramar structured in order? Meaning, give the following Wordtpyes: S1 - a1 - S2 - a2 - S3, could a1 also effect S3 or is it always only affecting S2.

Meaning, what are our chunk patterns ?

No, at least in german there are many examples that break this assumption, so I guess it is the same for english! "Slowly, eaten by the bear, the human starts to cry"

[dominik-pichler]

Further Overview of LLM alternatives

[dominik-pichler]

For now, I'll stop once the Rule Bases Methods work

[dominik-pichler]

so there then will be two versions: 1) LLM based Meaning/Knowledge Extraction 2) RB Meaning/Knowledge extraction -> Nouns will be nodes, corresponding sentences will be edges

[dominik-pichler]

so the first version is working. It just needs some cleaning