Argument Role Labelling Question-Answer Pair Data Augmentation

One sentence summary: Augment question-answer pairs with semantic and syntactic constraints to solve the argument role labeling task.

[What] This work explores the QA augmentation for events in news articles. Our altimate goal is to automatically extract structured events from unlabeled unstructured articles. In this work, we focus on a subtask: extract QA pairs for the arguments given an event. The synthetic data then severs as augmented data to train the QA event extraction model.

They were married in Spain.
Arument Role LabelQuestionAnswer
Person-ArgWho is married?They
Place-ArgWhere is someone married?Spain

[Why] Traditionally, to extract structured event QA pairs from an article, the steps are: ask questions → extract answers. However, to augment event QA pairs, we propose a reversed approach: extract answers → ask questions. This approach takes advantage of the transfer ability of QA datasets and circumvents the high proportion (~60%) of no-answer questions.

See our paper and code.

PROC2PDDL: Predicting Domain Definitions Based on Natural Language for Symbolic Planning

One sentence summary: Reasoning on the changes of entities and states when performing an action in symbolic language.

[What] This work explores the causal reasoning of events. It treats causal links as the pre- and post-conditions between actions, represented in the symbolic language. To implement a sequence of actions, the pre-conditions should be satisfied before executing each action, and the post-conditions will be acquired after the action. We propose to generate the pre- and post-conditions of actions in three steps: text identification, entity-state extraction, and entity-state translation. Then, a PDDL solver finds a path to achieve a goal.

# an example of PDDL action go;

(: action go
  parameters: avatar_1, location_1, location_2 
  pre-conditions: (avatar_1 at location_1) (location_1 connect to location_2) # an action go requires an avatar 1 at location 1 and location 1 connects to location 2
  post-conditions: (avatar_1 at location_2) # the result is an avatar 1 at location 2

[Why] The core theory supporting this approach is tracking the dynamics of entities and their attributes (more granular than an action itself), such as 'avatar 1' with attribute 'at location 1', in a series of unit actions. It decomposes the accomplishment of a complicated goal into the fulfilment of a series of reusable unit actions, e.g. 'go'.

# an example of changes of attributes of an entity 'avatar 1';

init: (avatar_1 at location_1)
go: (avatar_1 at location_2) # avatar_1's location attribute changes from location_1 to location_2
get: (avatar_1 has item_1) # avatar_1's inventory attribute adds item_1
goal: (avatar_1 has item_1)

See our paper and code.

Human-in-the-loop Schema Induction

One sentence summary: Four stages event schema induction with GPT-3 and human curation.

[What] This work explores a human-computer interactive way to build an event schema. Event Schema is a structured graph representation of events, analogical to knowledge graph. To clearly guide a machine and human for schema induction, we divide the task into four sequential stages: step generation, node extraction, graph construction, and node grounding.

# a toy sample of an event schema.
# temporal relation: searching event happens before find event.
# co-reference of entities: subject: person_1 and object: item_1 are shared in two events.

event_searching(subject: person_1 ; object: item_1) → event_find(subject: person_1 ; object: item_1)

[Why] Our approach is proposed based on the observations that automatic schema induction can be productive but vulnerable in quality; whereas manual schema induction can be accurate but onerous. We combine the work of humans and computers to generate event schemas in high efficiency (quality and quantity).

The video recording for the paper.

See our paper, code, and poster.