Abductive reasoning
Abductive reasoning (also called abduction,[1] abductive inference,[1] or retroduction[2]) is a form of logical inference that seeks the simplest and most likely conclusion from a set of observations. It was formulated and advanced by American philosopher Charles Sanders Peirce beginning in the latter half of the 19th century.
Abductive reasoning, unlike deductive reasoning, yields a plausible conclusion but does not definitively verify it. Abductive conclusions do not eliminate uncertainty or doubt, which is expressed in retreat terms such as “best available” or “most likely”. While inductive reasoning draws general conclusions that apply to many situations, abductive conclusions are confined to the particular observations in question.
In the 1990s, as computing power grew, the fields of law,[3] computer science, and artificial intelligence research[4] spurred renewed interest in the subject of abduction.[5] Diagnostic expert systems frequently employ abduction.[6]
Abduction is considered the third form of reasoning, distinct from deduction and induction. It involves inferring the best possible explanation for a set of observed facts or evidence.
The lecture covers the following key points:
1. Abduction is likened to solving a mystery, where you have evidence and need to figure out what caused it.
2. There can be multiple possible explanations (solutions) for an abductive problem, unlike deduction which typically has one solution.
3. Explanations in abduction need to satisfy the criteria of covering (explaining all observations) and parsimony (meeting some simplicity standard).
4. Different types of abduction problems are discussed: logic-based, set covering, probabilistic, and consistency-based.
5. An example of geospatial abduction is presented, where the goal is to locate weapons caches based on IED attack locations, using a set covering approach.
6. Probabilistic horn abduction, which combines logic and probabilities, is explained with examples from perception tasks like scene understanding.
7. Challenges in abduction are highlighted, such as where to obtain the required domain knowledge or rules, which limited wider adoption initially.
The lecture aims to provide a comprehensive understanding of abductive reasoning, its types, approaches, examples, and historical context.
Abductive reasoning is the process of making educated guesses to provide explanations for observations. Although many applications require the use of knowledge for explanations, the utilization of abductive reasoning in conjunction with structured knowledge, such as a knowledge graph, remains largely unexplored. To fill this gap, this paper introduces the task of complex logical hypothesis generation, as an initial step towards abductive logical reasoning with KG. In this task, we aim to generate a complex logical hypothesis so that it can explain a set of observations. We find that the supervised trained generative model can generate logical hypotheses that are structurally closer to the reference hypothesis. However, when generalized to unseen observations, this training objective does not guarantee better hypothesis generation. To address this, we introduce the Reinforcement Learning from Knowledge Graph (RLF-KG) method, which minimizes differences between observations and conclusions drawn from generated hypotheses according to the KG. Experiments show that, with RLF-KG’s assistance, the generated hypotheses provide better explanations, and achieve stateof-the-art results on three widely used KGs.
This paper introduces the task of complex logical hypothesis generation for abductive reasoning on knowledge graphs. Given a set of observations (entities) and a knowledge graph, the goal is to generate a logical hypothesis that best explains or describes the observations.
Get imbaresearchlabs’s stories in your inbox
Join Medium for free to get updates from this writer.
The main contributions are:
1. Formulating the task of abductive logical reasoning on knowledge graphs by generating complex logical hypotheses to explain observations.
2. Proposing a generation-based approach using transformer models trained on sampled observation-hypothesis pairs to tackle the challenges of knowledge graph incompleteness and the complexity of logical hypotheses.
3. Introducing a reinforcement learning method called RLF-KG (Reinforcement Learning from Knowledge Graph) that incorporates feedback from the knowledge graph to further improve the hypothesis generation model’s ability to provide accurate explanations for unseen observations.
4. Evaluating the proposed methods on three knowledge graph datasets (FB15k-237, WN18RR, DBpedia50), showing the effectiveness of the generation approach combined with RLF-KG over baselines in generating high-quality logical hypotheses that align well with the observations.
The paper frames abductive reasoning on knowledge graphs as a hypothesis generation problem and provides an effective generation-based solution enhanced by reinforcement learning from knowledge graph feedback.
Abductive reasoning aims to find plausible explanations for an event. This style of reasoning is critical for commonsense tasks where there are often multiple plausible explanations. Existing approaches for abductive reasoning in natural language processing (NLP) often rely on manually generated annotations for supervision; however, such annotations can be subjective and biased. Instead of using direct supervision, this work proposes an approach for abductive commonsense reasoning that exploits the fact that only a subset of explanations is correct for a given context. The method uses posterior regularization to enforce a mutual exclusion constraint, encouraging the model to learn the distinction between fluent explanations and plausible ones. We evaluate our approach on a diverse set of abductive reasoning datasets; experimental results show that our approach outperforms or is comparable to directly applying pretrained language models in a zeroshot manner and other knowledge-augmented zero-shot methods.
Here are the key points summarizing the paper:
- The paper proposes LiPoR (Likelihood learning with Posterior Regularization), an approach to perform abductive commonsense reasoning without requiring direct supervision from plausibility annotations for explanations.
- LiPoR fine-tunes a pretrained language model by maximizing the marginal likelihood of the outcome given the context and explanations. It then enforces mutual exclusivity between plausible and implausible explanations via posterior regularization.
- LiPoR outperforms zero-shot pretrained language models and achieves comparable performance to knowledge-augmented zero-shot methods on several abductive reasoning datasets.
- The paper analyzes LiPoR’s ability to preserve multiple plausible explanations at test time, unlike models trained on plausibility annotations which tend to collapse to one explanation.
- LiPoR can leverage automatically generated (and noisier) explanation candidates and outperform strong zero-shot baselines in this setting.
- The analysis shows LiPoR’s limitations in handling distractor explanations that break the assumption of a uniform prior over explanations given just the context.
In summary, the paper introduces an unsupervised approach called LiPoR for abductive commonsense reasoning that avoids the need for manually annotated plausibility labels by exploiting the mutual exclusivity constraint between plausible and implausible explanations.
