Unsupervised Relation Extraction

[gonzobrandon]

I see the great tutorial from @svlandeg on relation extraction - where (to my understanding) we supply a categorical list of relations to train and then predict against. For example: {binding, inhibition, activation} as a list of possible relations in the video example.

How does this method scale to something like thousands of categories or with one-shot learning? What I'm aiming to try is to find a relation between two (NER) entities with either no pre-knowledge of the linguistic relationship or a subset of corpus tokens nearby to query against. An example: "Today, Kevin Quinn won the New York City marathon" ...in this case the word "won" would be the specific relationship I'm after. But what if the training model has not seen this specific relationship (verb, in this case). I believe this is called Open-IE-style. I could reduce the relations down to perhaps 1,000-5,000 key phrases and words if needed. Certainly, calculating token distance via embeddings can be done for out-of-vocab or out-of-corpus words.

Any papers or implementations would be very appreciated. If any engineers or consultants are available to work on this, please reach out.

[svlandeg]

Hi!

I got confused with the terminology of your post for a second, so just to clarify, within spaCy code & docs, we define:

• entity linking as the process of linking a textual mention (e.g. "President Obama") to a unique database identifier, e.g. "Q76".
• relation extraction as the process of determining whether or not two (or more) entities are in a semantic relation as expressed by the sentence context, and if so, labeling that relation.

I understand your question is about relation extraction, though the title threw me off ;-)

Anyway, all that to then give you a disappointing answer that personally, I haven't really looked into this "open ended" style of relation extraction yet. But I'd be interested in hearing from others in the community if they have any relevant pointers!

[gonzobrandon]

Thanks @svlandeg. Yes, the title was off (fixing). Ultimately my pipeline includes NER -> NEL and the last step is indeed the "open ended" relation extraction.

I have had some success with implementing the Hearst Patterns. Some of the sources I turned to were:

[1] Hearst Patterns Revisited: Automatic Hypernym Detection from Large Text Corpora seems to have had a similar experience to my own.

[2] Hearst Patterns with Spacy pip package (mmichelsonIF/hearst_patterns_python)

[3] Introduction to Information Extraction using Python and spaCy A good overview of some approaches. (Thanks Walid from UBIAI)

Some implementation thoughts on improving on the rule/pattern approach:

1. Training a model: I would naturally like to train a neural network with spacy/prodigy that could take some number of example documents and like was done in the Emerson example, as input provide {Sentence Context, Sentence Dependency Tree, POS Tags} from which perhaps there is enough of a pattern to generalize. This is what I am working on. I have yet to implement the annotation tool in either Prodigy or another tool, such as
   UBIAI but I will start this very soon.

2. Extending Emerson Project: I thought about a very large extension of the Emerson project and including thousands or tens of thousands of token/phrases that summarize the relation well, but annotating and scaling that model might be challenging to say the least.

3. I think the open-ended RE problem is tractable because a given sentence context (or some n number of sentences before) contain the verb or adjective token or span that contain the best relation...So, we're not searching from zero-knowledge or trying every possible relation possible in the language.

4. There is a lot of research and datasets available for extracting triples. Some of them try to restrict the relation to a hybrid identifier that can be adapted to OWL/RDF data format such as has_buyer etc - however could be equally useful to extract a natural language representation.

   • Borcan. Python NLP Tutorial: Information Extraction and Knowledge Graphs (2020)

   • Sakor et. al. Old is Gold: Linguistic Driven Approach for Entity and Relation Linking of Short Text (2019)

   • Gozwami et.al. Unsupervised Relation Extraction from Language Models using Constrained Cloze Completion (2020)

Wold love community input on this topic. Thank you!