Using cross-entity inference to improve event extraction Reading Notes

To do what?

This article mainly uses Object Type consistency to predict event types, event members (participants), and member roles. It depends on the following three phenomena:

1. Entities of the same type are often involved in similar events. These events often use consistent or similar triggers.

2. in events of the same type, entities of the same type normally appear together with similar participants.

3. Members of the same type play the same role in similar events.

 

How to do?

Follow these steps to perform step-by-step reasoning:

Step 1: predict the event type and mark the trigger words that are given to the same type of entities.

Step 1: Identify members in an event by the prior entity type, event type, and trigger word given in step 2.

Step 2: determine the Member role by specifying the entity type, event type, trigger word, and member in an event.

 

The main framework of cross-entity reasoning is as follows:

Training process:

We use 549 ace texts as the training corpus, 10 other news texts for parameter debugging, and 40 news texts for testing.

A1: For each Entity description, we use this description as a query for 50 documents from the web. Then we select 50 keywords and use TFIDF to determine the weight to describe the object background. Create a spatial vector model (VSM) for each type of Entity description ).

A2: cluto clustering tool is used to divide entity types into child types.

A3: extract the following features to prepare for SVM-based classifier.

A4: argument classifier, role classifier, and report event classifier are implemented based on SVM classifier.

 

Test process:

T1: A sentence must contain at least one entity.

T2: Create object background with A1

T3: the clustering obtained by cluto and the K-Nearest Neighbor method are used to determine the object subtype.

T4: feature extraction is the same as A3

T5: Perform step-by-step cross-entity reasoning to detect the trigger word, determine the event type, member category, role category, and report event category.

 

What shoshould I do?

1. Establishes information retrieval and the VSM spatial vector model used in the object background.
2. Cluto Clustering
3. Specific Feature Extraction
4. SVM-based classifier