We propose a simple, syllogism-centric scheme for argument mining that builds full arguments by chaining “atomic” syllogisms. Unlike taxonomies that distinguish “claim” vs. “premise,” we collapse local (intermediate) conclusions into analysis units and reserve the “conclusion” label only for the document's final outcome. This choice aligns with polysyllogistic reasoning and reduces label ambiguity. We formalize the scheme with concise guidelines for marking atomic links and test it on U.S. corporate reorganization cases under I.R.C. §368. In passage classification with a Linear SVC over several embeddings (TF-IDF, SBERT, Legal-BERT, ModernBERT) and an LLM classifier (GPT-5-mini), collapsing intermediate conclusions into improves macro-F1 by 7-15 points over the 5-class setup across embeddings (Table 1).
Data and annotation.
We collected 40 U.S. corporate reorganization cases (1k–10k words), focusing
on I.R.C. §368(a)(1)(A),(B),(C),(D),(F) and excluding (E),(G) to limit statutory variety. This report includes
human expert annotations on 26 documents, containing a total of 333 valid classified passages. We
define a passage as a span of text that does not necessarily align with sentence boundaries, which can
be subjective and detached from logical units. This ensures that each passage is annotated according to
its distinct role within the argument. The annotator, a second-year law student, worked independently
using a modified version of the annotation software Label Studio. Their task was to select spans of text
consisting of atomic claims, and connect them according to a syllogistic grammar drawn from Gardner
and Bartholomew [
Passages are labeled as analysis, rule, conclusion (final only), background facts (BF), and
procedural history (PH). Links connect premises (rules/analyses) to their immediate conclusions, forming
chains. Several similar argument chain approaches have been proposed for annotating cases and
tackling argument mining tasks [
CEUR Workshop
ISSN1613-0073
TF-IDF SBERT Legal-BERT Modern-BERT GPT-5-mini
BF 0.70 0.65 0.81 0.78 0.64 0.14
Conclusion they serve as supporting text, not active components of the argument. An example of a valid syllogism annotation is provided in Figure 1.
The primary improvement of our approach lies in focusing on the functional role within the argument, rather than isolating individual roles, and employing a recursive method to build these structures. Our dataset aims to create a consistent and closely deductive structure that mimics human legal reasoning and provides an indication of logical deduction at each step of the argument construction.
Conclusion: We accept petitioner’s position
Rule: To accept petitioner’s position, we would have to find that the stock had a value of $652,180
Analysis: The record shows that the
preferred stock issued by the petitioner is $652,180
Models. We compare (i) a Linear SVC [
Results and Discussion. Table 1 shows macro F1 across embeddings and the LLM classifier. Moving
from 5 classes to the 4-class variant (collapsing final conclusions into analysis) yields consistent gains:
+0.15 (TF–IDF), +0.10 (SBERT), +0.09 (Legal-BERT), +0.07 (ModernBERT), and +0.10 (GPT-5-mini).
LegalBERT attains the highest average macro F1 among embeddings in the 4-class setup (0.82) as well as in
the 5-class setting (0.73). We include a random baseline for reference. We hypothesize that removing
the ambiguous “final conclusion” label reduces overlap with analysis, improving separability when
passages are judged out of context. As a secondary contribution, the background facts and procedural
history provide span-level section annotations of U.S. case law, which are formally structured in ECHR
[
Limitations and Future Work. Our evaluation uses a single annotator over 26 opinions and classifies passages independently; richer context (e.g., graph models) may recover information lost at sentence scope by modeling relations between passages and allow visual reasoning about entailment and counterfactuals. Future work includes structured prediction over chains, explicit entailment links, and information retrieval to test argument completion.
Declaration on Generative AI. During the preparation of this work, the authors used Grammarly in order to: Grammar and spelling check, paraphrase and reword. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.