PrimeNet: A Framework for Commonsense Knowledge Representation and Reasoning Based on Conceptual Primitives

PrimeNet is a scalable, generalizable knowledge base which enables effective commonsense and real-world knowledge representation and facilitates commonsense reasoning including inferences.

Overview

Inspired by the theory of conceptual primitives, PrimeNet is constructed by comprising a small core of primitive commonsense concepts and relations, linked to a much more extensive base of factual knowledge instances.

Case Study

Primitives are organized in a multi-level hierarchy that allows to iteratively generalize words and multiword expressions into increasingly general primitives. Top-level primitives (aka superprimitives) are then defined by means of first-order logic.

More cases on primitive generalization:

Evaluation

1. The evaluation of Human Assessment (detailed in Section 4.1) can be reproduced using the facts from PrimeNet V3. Download all facts in PrimeNet V3 from this link.

2. The evaluation of Distributional Representations (detailed in Section 4.2) can be reproduced using the released files in the Evaluation folder.

   The code of refining word embeddings is available in this link.

   We released the refined representations using different knowledge bases (i.e., WordNet, FrameNet, ConceptNet, and PrimeNet). All files can be downloaded using this link.

   glove.6B.50d.primenet.txt glove.6B.50d.conceptnet.txt glove.6B.50d.wordnet.txt glove.6B.50d.framenet.txt glove.6B.300d.primenet.txt glove.6B.300d.conceptnet.txt glove.6B.300d.wordnet.txt glove.6B.300d.framenet.txt word2vec.cbow.300d.primenet.txt word2vec.cbow.300d.conceptnet.txt word2vec.cbow.300d.wordnet.txt word2vec.cbow.300d.framenet.txt

   Example of evaluating on word similarity task:

   python eval_wordsim.py refined_vectors/glove.6B.50d.conceptnet.txt

   The output is:

   0.4427091438043214 0.6743373528926029 0.7073685164959393 0.5974265070145617 0.3761684502800549 0.2355190563306831 0.27269399402968747 0.48538634242377177

Functions (PrimeNet SDK)

In primenet_graph.py, we define over 20 functions to explore PrimeNet. The input, output, and description of each function is explained on Table 3 in the paper.

We show the examples of using these functions:

primenet = read_json_file("primenet_fact_v3.json")
graph = Graph(primenet)

# generate all nodes and edges in PrimeNet:
graph.nodes()
graph.edges()

graph.get_number_of_nodes()
# output: 1409185
graph.get_number_of_edges()
# output: 3056776
graph.density()
# output: 3.0786328229340862e-06

graph.what_is("hammer")
# output: hammer --isA--> hand_tool
graph.get_node_degree("hammer")
# output: 24
graph.explain(node="hammer",relation="isA")
# output: hammer --isA--> hand_tool --isA--> tool
graph.what_can_be("hammer")
# output: [['isA', 'hand_tool'], ['isA', 'industrial_equipment'], ['isA', 'power_tool'], ['isA', 'tool'], ['mannerOf', 'beat'], ['usedFor', 'hit'], ['madeOf', 'forged_metal'], ['partOf', 'gunlock'], ['partOf', 'piano_action'], ['IntentionOf', 'build'], ['IntentionOf', 'demo'], ['Afford', 'breaking_glass'], ['Afford', 'break_glass'], ['Afford', 'nail_board'], ['Afford', 'break_wall'], ['Afford', 'break_window'], ['Afford', 'strike_nail'], ['Afford', 'force_nail_into_board'], ['Afford', 'break_fragile_objects'], ['Afford', 'nail_nails'], ['Afford', 'strike_with_great_force'], ['Afford', 'nail_nail'], ['Afford', 'hit_nail'], ['Afford', 'drive_in_nails']]
graph.generalize("hammer")
# output: [('Afford', 'BREAKING_GLASS'), ('mannerOf', 'STRIKE'), ('usedFor', 'HIT'), ('IntentionOf', 'BUILD'), ('isA', 'TOOL'), ('partOf', 'ARSENAL'), ('madeOf', 'FORGED_METAL')]

graph.relation_exist(node="hammer",relation="Afford")
# output: True
graph.relation_types("hammer")
# output: ['madeOf', 'Afford', 'IntentionOf', 'usedFor', 'partOf', 'mannerOf', 'isA']

graph.get_node_with_relation(node="hammer",relation="isA")
# output: 'hand_tool'
graph.get_node_with_relation(node="hammer",relation="usedFor")
# output: 'hit'
graph.find_last_nodes("hit_nail")
# output: ['hammer-->Afford-->hit_nail']
graph.get_all_node_with_relation(node="hammer",relation="Afford")
# output: ['breaking_glass', 'break_glass', 'nail_board', 'break_wall', 'break_window', 'strike_nail', 'force_nail_into_board', 'break_fragile_objects', 'nail_nails', 'strike_with_great_force', 'nail_nail', 'hit_nail', 'drive_in_nails']

graph.find_path(start_node="hammer",end_node="tool")
# output: [('START', 'hammer'), ('isA', 'hand_tool'), ('isA', 'tool')]
graph.find_all_paths(start_vertex="hammer",end_vertex="tool")
# output: [[('START', 'hammer'), ('isA', 'hand_tool'), ('isA', 'tool')], [('START', 'hammer'), ('isA', 'power_tool'), ('isA', 'tool')], 002: [('START', 'hammer'), ('isA', 'tool')], [('START', 'hammer'), ('mannerOf', 'beat'), ('isA', 'band'), ('isA', 'record'), ('isA', 'tool')]...]
graph.get_path(start_node="hammer",end_node="tool")
# output:'hammer --isA--> hand_tool --isA--> tool'

graph.get_similarity(node1="hammer",node2="pen",relation="Afford")
# output: 0.11111
graph.get_similarity(node1="hammer",node2="brick",relation="Afford")
# output: 0.66667

graph.add_node("a_new_node")
graph.add_edge(("a_new_node", "isA","testing_example"))