Asymmetric Modeling of Directional Entailment

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 236 papers53 conceptsfrozen 2026-10-04

In natural language processing, we often conflate "similarity" with "meaning." If two sentences share many words or have similar vector embeddings, we assume they mean the same thing. However, logic is not a mirror; it is a one-way street. While "A man is walking a dog" implies "An animal is moving," the reverse is certainly not true.

A growing body of research argues that the fundamental failure of many semantic models stems from treating logical relationships as symmetric. To move from simple pattern matching to genuine reasoning, models must explicitly represent the asymmetry of entailment.

The Fallacy of Symmetry

Most distributional models—like those based on cosine similarity or word overlap—are inherently bidirectional. If sentence $A$ is 90% similar to sentence $B$, then $B$ is 90% similar to $A$. As Ido Dagan and Oren Glickman pointed out in their early work on probabilistic entailment, this symmetry is a bug, not a feature.

Ken-Ichi Yokote and Danushka Bollegala demonstrated that textual entailment features should be constructed by centering calculations on the hypothesis rather than the source text. In their view, the goal isn't to see how much the two sentences overlap, but to verify if the specific information required by the hypothesis is fully covered by the premise. The premise can contain a mountain of irrelevant information (e.g., "The tall man in the red hat is walking a golden retriever") and still entail a simple hypothesis ("A man is walking"), but a symmetric similarity metric would penalize the "mismatch" caused by the extra details.

Geometry of Implication

How do we force a vector space—which is naturally built on symmetric distances—to respect the one-way flow of logic?

One influential approach, proposed by Thomas Demeester and Sebastian Riedel, involves lifted rule injection. Instead of treating relations as points in space, they model logical implications as a partial ordering of embedding vectors. If relation $P$ entails relation $Q$, the model enforces a constraint where every component of $P$'s vector must be less than or equal to the corresponding component of $Q$'s vector. This creates a geometric "subsumption" where the premise is literally contained within the boundaries of the conclusion.

Similarly, Naoaki Okazaki and Ran Tian found that training distributional representations with these formal logic constraints allows models to distinguish between causes and effects—relationships that are often co-located in text (and thus "similar") but logically distinct.

Entailment as Structural Subsumption

Beyond vectors, researchers have looked at the internal structure of sentences to find the "direction" of logic. Juri Opitz and Anette Frank proposed treating Natural Language Inference (NLI) as a semantic substructure problem. By converting sentences into Abstract Meaning Representation (AMR) graphs, they measure how much of the hypothesis graph is contained within the premise graph. This asymmetric metric is far more robust against the "spurious cues" that often fool black-box neural models.

This structural view also applies to document processing. Hiroya Takamura and Manabu Okumura reimagined text summarization as a "Facility Location Problem," where summary sentences act as facilities that must "cover" the information in the rest of the document. By modeling the relationship between sentences as asymmetric—where a specific sentence can cover a general one but not vice versa—they produced summaries that better capture the hierarchy of information in a text.

Directed Graphs of Knowledge

In the realm of Knowledge Graphs, the challenge is to move from "link prediction" (finding missing facts) to "entailment graph induction" (finding missing rules). Mohammad Javad Hosseini and Shay B Cohen argue that these two tasks are duals of each other. They use relational Markov chains to capture the directionality of entailment: the probability of transitioning from "acquire" to "own" is high, but the reverse is low.

This directional graph approach has even been applied to modern Large Language Models (LLMs). Longchao Da and Hua Wei recently showed that we can quantify LLM uncertainty by building a directional entailment graph of the model's various responses. If a model generates several specific claims that all entail a single general conclusion, it is likely consistent. If it generates claims that are "similar" but logically contradictory in their directionality, it is likely hallucinating.

Why It Works: The Principle of Containment

The recurring intuition across these 35 papers is that inference is a containment relationship.

  • In logic, the set of worlds where the premise is true must be a subset of the worlds where the hypothesis is true.
  • In linguistics, the context of the hypothesis must be broader than or equal to the context of the source (Idan Szpektor).
  • In student grading, a response is correct if it entails the reference answer, even if it adds extra (asymmetric) information (Omer Levy).

By replacing symmetric similarity with asymmetric containment—whether through partial vector orders, directed graphs, or substructure matching—researchers are building models that don't just recognize that two things are "about the same topic," but understand which one follows from the other.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 49 papers
  2. Version 1 · 2026-10-04 · 36 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_directional_entailment_modeling_411c82f8,
  title = {directional entailment modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/directional-entailment-modeling/rdr_411c82f829274f89924c38aa59840113},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_asymmetric_modeling_of_directional_entai_411c82f8_v1,
  title = {Asymmetric Modeling of Directional Entailment},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-modeling-of-directional-entailment/rdr_411c82f829274f89924c38aa59840113/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}