Event Vec

What it is

This project aims to vectorize event mentions in text. Earlier implementations use a version of the skip-gram model to vectorized event mentions. This project aims to incorporate temporal information in to the process of event vectorization.

Initial Design

We learn the event vectors by feeding through a network that takes the vectorized versions of the parts of two event mention as inputs and outputs the temporal relationship between them. An additional output is the confidence of distribution.

Parts of the code

Data Ingestor

This modular component can be swapped out for other components that read from different types of sources such as Wikipedia or ebooks.

Coreference resolution

This modular component can be swapped out for other components that resolve coreferences within the same text.

Event extractor

This component, based on dependency parses and other hand-created rules, will extract the verb, subject, object etc for a given event mentions.

Date extractor

This component based on Spacy's entity recognizer and hand-created rules will extract date mentions in the text and attempt to normalize them

Event relationship extractor

Using Spacy's dependency parser, the relationships between events or events and dates are extracted. This is learned from TIMEBANK1.2.

Vectorizer

The network as described in the intial design.

Utils

The component that is able to learn the relationships between entity and dates from the TIMEBANK data.

Testing

The vectors will be tested against the tasks of

• event similarity
• event cloze
• event year prediction

And the implementations of Modi and Granroth-clark for event vectorization.

Also added will be an ablation study that is able to find the impact of

• coreference resolution
• relationship extractor and the different methods for it.

Downstream usage

For this to be useful, one can

• train with their own data
• feed events in a structured format and obtain event vectors
• feed sentences and obtain events in an structured format and their vectors.

Demo

To quickly see the project and what it is trying to achieve. There will be a demo which can take two events and output the similarity. Take an event and output the similar events from the trained data.

Running on Jade

PYTHONPATH=REMOTE_CODE_FOLDER_PATH/event-vec ENV=JADE RUNCONFIGID=41 python3.8 REMOTE_CODE_FOLDER_PATH/event-vec/eventvec/server/entry_points/main.py

Running Locally

REQUEST_ID=1234 PROJECT_ID=event-vec RUNCONFIGID=21 ENV=JADE PYTHONPATH=jade_front/event-vec/code/event-vec python jade_front/event-vec/code/event-vec/eventvec/server/entry_points/main.py