Plausible Reasoning Based on Qualitative Entity Embeddings

Steven Schockaert, Shoaib JameelIJCAI 2016 · OpenReview

Abstract: Formalizing and automating aspects of human plausible reasoning is an important challenge for the field of artificial intelligence. Practical advances, however, are hampered by the fact that most forms of plausible reasoning rely on background knowledge that is often not available in a structured form. In this paper, we first discuss how an important class of background knowledge can be induced from vector space representations that have been learned from (mostly) unstructured data. Subsequently, we advocate the use of qualitative abstractions of these vector spaces, as they are easier to obtain and manipulate, among others, while still supporting various forms of plausible reasoning.

Introduction

Artificial Intelligence often struggles with "plausible reasoning"—the ability to draw conclusions that are highly likely given background knowledge, even if they aren't strictly logically certain. For instance, if one knows that the BBC and ITV are British broadcasters regulated by Ofcom, it is plausible to infer that other British broadcasters share this regulatory status.

Traditionally, AI has relied on two paradigms: rigid symbolic logic (like WordNet taxonomies) and numerical vector embeddings (like word2vec). While taxonomies are often too shallow to capture nuanced relationships, vector embeddings are "black boxes" that are difficult to interpret or manually correct. This paper proposes a middle ground: Qualitative Entity Embeddings.

Methodology: From Vectors to Qualitative Rankings

The proposed framework abstracts high-dimensional entity embeddings into qualitative, interpretable rankings. The process involves several key steps:

  1. Inducing Vector Spaces: Starting with high-dimensional vectors learned from unstructured data.
  2. Identifying Salient Directions: Instead of using all mathematical dimensions, the system identifies specific "directions" corresponding to meaningful properties, such as "chronology" for historical civilizations or "violence" for movies.
  3. Qualitative Abstraction: Raw coordinates (e.g., 0.872) are converted into rankings (e.g., "more recent than"). These rankings are more robust to noise and easier for humans to validate.
  4. Interval-Based Concepts: Concepts are represented as intervals within these rankings. An entity belongs to a concept if its rank falls within specific bounds across relevant dimensions.
Visualization of historical civilizations mapped within a domain-specific subspace.

Plausible Reasoning and Betweenness

A core contribution of this work is the use of geometric "betweenness" for interpolation. If an entity like a "harpsichord" is positioned between a "piano" and a "guitar" in the qualitative space, the system can plausibly infer that properties shared by the endpoints likely apply to the entity in the middle. This allows for flexible reasoning that mimics human common sense.

Resolving Ontological Conflicts

The framework is particularly useful for merging conflicting databases. Different knowledge bases often use different terminologies or hierarchies (e.g., one defining a rugby ball as a "ball" and another as a "subclass of football"). By using conceptual neighborhoods in the embedding space, the system can "weaken" these rigid definitions to find a consistent middle ground.

Example of resolving terminological conflicts between concepts like football, soccer, and rugby.

Key Insights and Results

  • Robustness: Qualitative rankings are more stable than raw coordinates and allow domain experts to easily tweak the model.
  • Efficiency: The paper provides a formal method to check if one concept is "between" others using simple interval arithmetic, which is computationally more efficient than complex geometric calculations like convex hulls.
  • Accuracy: Previous experiments cited by the authors show that classifiers using these qualitative rankings can match the accuracy of those using full, high-dimensional numerical vectors.

Limitations and Future Work

While promising, the method currently relies on identifying "salient directions," which sometimes requires manual intervention or sophisticated pattern extraction from text. Additionally, representing concepts as simple hyperrectangles (intervals) can occasionally lead to false positives. Future research will focus on automatically selecting relevant rankings based on specific reasoning contexts, such as distinguishing between a "coastal" context and a "geological" context for the same entity.

Summary updated

Abstract

Formalizing and automating aspects of human plausible reasoning is an important challenge for the field of artificial intelligence. Practical advances, however, are hampered by the fact that most forms of plausible reasoning rely on background knowledge that is often not available in a structured form. In this paper, we first discuss how an important class of background knowledge can be induced from vector space representations that have been learned from (mostly) unstructured data. Subsequently, we advocate the use of qualitative abstractions of these vector spaces, as they are easier to obtain and manipulate, among others, while still supporting various forms of plausible reasoning.
BibTeX
@inproceedings{schockaert2016plausible,
  author = {Steven Schockaert and Shoaib Jameel},
  title = {{Plausible Reasoning Based on Qualitative Entity Embeddings}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=HyE2s4MdWB},
}
Figure