AMIE 3.5.1: Symbolic Rule Mining for Knowledge Graphs

• AMIE 3.5.1 is a rule-mining engine for knowledge graphs that extracts Horn-style rules to enable symbolic reasoning and provide inputs for neural pipelines.
• It employs metrics such as support, head coverage, and standard confidence to evaluate rule quality under constraints like a maximum of 3 atoms.
• Integrated in systems like Rule2Text and GraphMIRAcles, AMIE 3.5.1 generates rules that enhance natural language explanation and candidate triple inference.

Searching arXiv for papers mentioning AMIE 3.5.1 to ground the article. AMIE 3.5.1 is a rule-mining engine for knowledge graphs that is used to discover Horn-style logical rules from relational triples. In recent arXiv literature, it appears as the symbolic component in pipelines for rule-based inference and for natural-language explanation of learned rules, rather than as a standalone predictor. These downstream uses describe AMIE 3.5.1 as a mature rule-learning system with “comprehensive metrics for rule evaluation,” and they treat it as compatible with benchmark knowledge graphs ranging from Freebase-derived corpora to large biomedical graphs (Shirvani-Mahdavi et al., 31 Jul 2025, Shirvani-Mahdavi et al., 14 Aug 2025).

1. Rule discovery role and operational scope

AMIE 3.5.1 is used to extract logical regularities from a knowledge graph in the form of Horn rules. In the Rule2Text line of work, those mined rules are the primary symbolic objects later verbalized by LLMs. In the WikiKG90M-LSC report by Team GraphMIRAcles, the same rule-mining function appears under the name “AMIE 3,” with the implementation taken from the GitHub code at lajus/amie; there, the miner is an upstream component that generates Horn rules from the training graph and feeds them into a downstream inference pipeline (Cai et al., 2021).

A central point in these papers is that AMIE 3.5.1 is not itself the explanation model, the embedding model, or the final ranking model. In Rule2Text, it supplies the rules that are later converted into readable natural-language explanations. In GraphMIRAcles, it supplies symbolic constraints and candidate triples that are later consumed by ComplEx-CMRC, ensembling, and knowledge distillation. This places AMIE 3.5.1 in the category of symbolic structure induction rather than end-task prediction (Shirvani-Mahdavi et al., 31 Jul 2025, Cai et al., 2021).

2. Formal rule language and rule-quality measures

The rule formalism associated with AMIE 3.5.1 is the standard Horn-clause form

$$B_1 \land B_2 \land \ldots \land B_n \Rightarrow H,$$

or equivalently $\overrightarrow{B} \Rightarrow H$, where the body consists of atoms $B_1,\ldots,B_n$ and the head is $H$. An atom is written as $r(h,t)$, corresponding to a triple $(h,r,t)$, with subject and/or object allowed to be variables. The rule fires when all body atoms can be instantiated in the knowledge graph (Shirvani-Mahdavi et al., 14 Aug 2025).

The Rule2Text papers also summarize the scoring notions used to filter or rank mined rules. Support is defined as the number of distinct $(subject, object)$ pairs in the head across all valid instantiations of the rule. Head coverage is support divided by the number of facts in the head relation. Standard confidence is support divided by the number of body instantiations. In the reported AMIE 3.5.1 setups, these are not merely descriptive metrics: they are used for rule selection, annotation, and evaluation (Shirvani-Mahdavi et al., 31 Jul 2025, Shirvani-Mahdavi et al., 14 Aug 2025).

These works also report explicit mining constraints. AMIE 3.5.1 is run with default settings “for optimized performance,” while imposing a minimum head coverage of 0.1, a minimum standard confidence of 0.1, and a maximum number of atoms of 3. A plausible implication is that the version, as used in these studies, is tuned for short, high-yield Horn rules rather than deeper multi-hop logical templates (Shirvani-Mahdavi et al., 14 Aug 2025).

3. Empirical scale in recent arXiv studies

The recent literature shows AMIE 3.5.1 being used at substantially different scales depending on the target knowledge graph. In Rule2Text, the mined rule sets span standard benchmarks and large industrial-scale or biomedical graphs. In GraphMIRAcles, the graph is sufficiently large that rule mining is performed on sampled subgraphs rather than on the full training graph at once.

Dataset or run	Reported rules	Source
FB15k-237	6,320	(Shirvani-Mahdavi et al., 14 Aug 2025)
FB-CVT-REV	14,355	(Shirvani-Mahdavi et al., 14 Aug 2025)
FB+CVT-REV	2,965	(Shirvani-Mahdavi et al., 14 Aug 2025)
ogbl-biokg	145,114	(Shirvani-Mahdavi et al., 14 Aug 2025)
WikiKG90M-LSC merged from five sampled subgraphs	11,716	(Cai et al., 2021)

For WikiKG90M-LSC, the five sampled subgraphs yielded 7,179, 4,981, 8,026, 3,903, and 5,999 rules respectively, which were then merged into a final rule set of 11,716 rules. After confidence filtering, the report retained 2,062 rules with confidence $> 0.95$ and 1,464 rules with confidence $> 0.99$. The same report states that rules were mined with length no longer than 3 (Cai et al., 2021).

These counts indicate two distinct operating regimes. On moderately sized benchmark graphs, AMIE 3.5.1 can be treated as a direct rule miner over the whole dataset. On graphs at WikiKG90M scale, the same basic functionality remains usable, but only through decomposition into sampled subgraphs. This suggests that the practical deployment envelope is strongly conditioned by graph size and by the downstream system’s tolerance for partial rule coverage.

4. Integration into symbolic-neural inference pipelines

In GraphMIRAcles, AMIE-derived rules are integrated into a hybrid workflow. The reported sequence is: first train the basic embedding model ComplEx-CMRC on the training set; then mine logical rules from the training graph with AMIE 3; then use the mined rules to augment data and improve inference, together with ensembling and knowledge distillation. AMIE therefore acts as a source of high-confidence symbolic patterns rather than as the main predictive model (Cai et al., 2021).

The report gives a concrete example of a mined Horn rule:

$$\text{livesIn}(h, p) \land \text{marriedTo}(h,w) \Rightarrow \text{livesIn}(w,p).$$

It also describes a more generic two-body form,

$\overrightarrow{B} \Rightarrow H$0

and defines new-triple predictions for relation $\overrightarrow{B} \Rightarrow H$1 through adjacency matrices:

$\overrightarrow{B} \Rightarrow H$2

Here, $\overrightarrow{B} \Rightarrow H$3 is the adjacency matrix for relation $\overrightarrow{B} \Rightarrow H$4, the matrix product captures paths matching the rule body, and subtracting $\overrightarrow{B} \Rightarrow H$5 removes already-known triples so that only unseen candidates remain (Cai et al., 2021).

This use of AMIE 3.5.1 is operationally important because it shows how mined rules can be converted into candidate-triple generators. The report further states that these generated triples are used to fine-tune the basic model and to “make prediction for unseen data.” At the same time, it does not provide a direct ablation isolating AMIE alone while holding the rest of the system fixed. Accordingly, the numerical contribution of the rule miner is supported indirectly, through the full-system performance and the documented rule statistics, rather than through a standalone AMIE-vs-no-AMIE comparison (Cai et al., 2021).

5. Interpretability, Rule2Text, and rule verbalization

A second major downstream use of AMIE 3.5.1 is interpretability. Rule2Text takes Horn rules mined by AMIE 3.5.1 and asks LLMs to convert them into natural-language explanations. The prompts do not use the bare rule alone: they include the rule, a concrete rule instance, dataset-specific background on predicate syntax, and, in stronger variants, variable entity types. The papers report that providing variable type information significantly improved explanation quality, and that Chain-of-Thought prompting further improved performance, with Gemini 2.0 Flash yielding the strongest overall results in the reported experiments (Shirvani-Mahdavi et al., 31 Jul 2025, Shirvani-Mahdavi et al., 14 Aug 2025).

The interpretability studies also clarify the kinds of rule structures AMIE 3.5.1 produces. They explicitly distinguish 2-atom rules, 3-atom rules, rules over binary relations, rules involving mediator nodes, and rules involving concatenated relations. The last category is especially salient in Freebase-derived graphs, where relation labels can become long concatenations of path fragments. These syntactic properties materially affect downstream comprehensibility: the papers state that 3-atom rules are harder than 2-atom rules, and that concatenated relations are harder than binary relations (Shirvani-Mahdavi et al., 31 Jul 2025).

The evaluation framework around these explanations is also revealing about the character of AMIE outputs. Human judges rate explanation correctness, clarity, missed entities/relations, hallucinated entities/relations, and rule logicalness. A plausible implication is that AMIE 3.5.1 generates rule sets that are semantically useful but not automatically human-readable, especially when the underlying knowledge graph uses idiosyncratic schema labels. The Rule2Text agenda is therefore not a replacement for AMIE, but a post hoc interpretability layer built around its mined rules (Shirvani-Mahdavi et al., 14 Aug 2025).

6. Constraints, limitations, and nomenclature

The constraints reported for AMIE 3.5.1 are consistent across the recent downstream literature. The most explicit is the maximum number of atoms of 3, which bounds the complexity of the mined Horn clauses. The Rule2Text papers explicitly note that future work should address “more complex rules beyond AMIE’s capabilities,” making the 3-atom regime both a practical design choice and a limitation of scope (Shirvani-Mahdavi et al., 14 Aug 2025).

Other limitations are empirical rather than formal. On very large graphs such as WikiKG90M, AMIE-based rule mining was not run on the full graph at once; instead, five sampled subgraphs were mined separately and then merged. The same report does not disclose AMIE-specific pruning parameters beyond the length limit and confidence-based filtering, and it does not report a support threshold. This suggests that practical use at extreme scale may rely on engineering workarounds not visible in simplified benchmark settings (Cai et al., 2021).

A final source of confusion is nomenclature. The label “AMIE” is overloaded across arXiv literatures. In autonomous-vehicle dialogue research it denotes Automated-vehicle Multimodal In-cabin Experience; in clinical dialogue research it denotes Articulate Medical Intelligence Explorer. AMIE 3.5.1, by contrast, refers to the knowledge-graph rule-mining engine used to discover Horn rules and to provide symbolic inputs for systems such as Rule2Text and GraphMIRAcles. Distinguishing these usages is essential, because the shared acronym does not imply methodological continuity across those domains (Okur et al., 2018, Tu et al., 2024).

7. Position within current research practice

Within current arXiv practice, AMIE 3.5.1 occupies a specific methodological niche: it is a symbolic rule discovery layer that is routinely paired with neural components rather than positioned against them. In GraphMIRAcles, it augments neural link prediction by generating rule-derived candidate triples and symbolic pseudo-labels. In Rule2Text, it supplies the structured logical inputs that LLMs attempt to explain. In both cases, its value lies in producing compact Horn-rule abstractions from large relational graphs (Cai et al., 2021, Shirvani-Mahdavi et al., 31 Jul 2025).

This division of labor also clarifies what AMIE 3.5.1 is not. It is not a dialogue system, not an LLM, not a text generator, and not a KG embedding model. Its outputs are rules, together with support-, coverage-, and confidence-like statistics that downstream systems use for filtering, augmentation, and interpretation. The recent literature therefore portrays AMIE 3.5.1 less as an end-user application than as infrastructure for symbolic reasoning over knowledge graphs (Shirvani-Mahdavi et al., 14 Aug 2025).