Simple Statement Knowledge Bases (SSKB)

Knowledge Base loading and annotation facilities

The sskb library provides easy access to Natural Language Knowledge Bases (KBs), and tools to facilitate annotation.

It exposes available KBs as sequences of simple statements. For example (from ProofWiki):

"A '''set''' is intuitively defined as any aggregation of 
objects, called elements, which can be precisely defined in 
some way or other."

Each statement is accompanied of relevant metadata, in the form of premises necessary for the statement to be true, and named entities associated with the respective KB.

SSKB is built upon the Simple Annotation Framework (SAF) library, which provides its data model and API. This means it is compatible with saf-datasets annotators.

Installation

To install, you can use pip:

pip install sskb

Usage

Loading KBs and accessing data

from sskb import ProofWikiKB

kb = ProofWikiKB()
print(len(kb))  # Number of statements in the KB
# 146723

print(kb[0].surface)  # First statement in the KB
# A '''set''' is intuitively defined as any aggregation of objects, called elements, which can be precisely defined in some way or other.

print([token.surface for token in kb[0].tokens])  # Tokens (SpaCy) of the first statement.
# ['A', "''", "'", 'set', "''", "'", 'is', 'intuitively', 'defined', 'as', 'any', 'aggregation', 'of', 'objects', ',', 'called', 'elements', ',', 'which', 'can', 'be', 'precisely', 'defined', 'in', 'some', 'way', 'or', 'other', '.']


print(kb[0].annotations)  # Annotations for the first sentence
# {'split': 'KB', 'type': 'fact', 'id': 337113631216859490898241823584484375642}


# There are no token annotations in this dataset
print([(tok.surface, tok.annotations) for tok in kb[0].tokens])
# [('A', {}), ("''", {}), ("'", {}), ('set', {}), ("''", {}), ("'", {}), ('is', {}), ('intuitively', {}), ('defined', {}), ('as', {}), ('any', {}), ('aggregation', {}), ('of', {}), ('objects', {}), (',', {}), ('called', {}), ('elements', {}), (',', {}), ('which', {}), ('can', {}), ('be', {}), ('precisely', {}), ('defined', {}), ('in', {}), ('some', {}), ('way', {}), ('or', {}), ('other', {}), ('.', {})]

# Entities cited in a statement
print([entity.surface for entity in kb[0].entities])
# ['Set', 'Or', 'Aggregation']

# Accessing statements by KB identifier
set_related = kb[337113631216859490898241823584484375642] # All statements connected to this identifier

print(len(set_related))
# 40

print(set_related[10].surface)
# If there are many elements in a set, then it becomes tedious and impractical to list them all in one big long explicit definition. Fortunately, however, there are other techniques for listing sets.

# Filtering ProofWiki propositions
train_propositions = [stt for stt in kb 
                      if (stt.annotations["type"] == "proposition" and stt.annotations["split"] == "train")]

print( train_propositions[0].surface)
# Let $A$ be a preadditive category.

print("\n".join([prem.surface for prem in train_propositions[0].premises]))
# Let $\mathbf C$ be a metacategory.
# Let $A$ and $B$ be objects of $\mathbf C$.
# A '''(binary) product diagram''' for $A$ and $B$ comprises an object $P$ and morphisms $p_1: P \to A$, $p_2: P \to B$:
# ::$\begin{xy}\xymatrix@+1em@L+3px{
#  A
# &
#  P
#   \ar[l]_*+{p_1}
#   \ar[r]^*+{p_2}
# &
#  B
# }\end{xy}$
# subjected to the following universal mapping property:
# :For any object $X$ and morphisms $x_1, x_2$ like so:
# ::$\begin{xy}\xymatrix@+1em@L+3px{
#  A
# &
#  X
#   \ar[l]_*+{x_1}
#   \ar[r]^*+{x_2}
# &
#  B
# }\end{xy}$
# :there is a unique morphism $u: X \to P$ such that:
# ::$\begin{xy}\xymatrix@+1em@L+3px{
# &
#  X
#   \ar[ld]_*+{x_1}
#   \ar@{-->}[d]^*+{u}
#   \ar[rd]^*+{x_2}
# \\
#  A
# &
#  P
#   \ar[l]^*+{p_1}
#   \ar[r]_*+{p_2}
# &
#  B
# }\end{xy}$
# :is a commutative diagram, i.e., $x_1 = p_1 \circ u$ and $x_2 = p_2 \circ u$.
# In this situation, $P$ is called a '''(binary) product of $A$ and $B$''' and may be denoted $A \times B$.
# Generally, one writes $\left\langle{x_1, x_2}\right\rangle$ for the unique morphism $u$ determined by above diagram.
# The morphisms $p_1$ and $p_2$ are often taken to be implicit.
# They are called '''projections'''; if necessary, $p_1$ can be called the '''first projection''' and $p_2$ the '''second projection'''.
# {{expand|the projection definition may merit its own, separate page}}

Available datasets: e-SNLI (ESNLIKB), ProofWiki (ProofWikiKB), WorldTree (WorldTreeKB).