Using the interest theory of rights and Hohfeldian taxonomy to address a gap in machine learning methods for legal document analysis

The paper addresses a core gap in legal NLP: while ML is used for tasks like named entity recognition, sentiment analysis, and predicting court outcomes, it does not explicitly capture the principal legal dimensions upon which law operates—rights and duties and their relations. The study proposes leveraging legal theory—the interest theory of rights and Hohfeld’s first-order taxonomy (rights–duties; privileges–no-rights)—to define principal dimensions suitable for machine learning abstractions. The research questions are: (1) Can ML detect these principal legal dimensions in sentences? (2) What ethical implications arise from using such technology? The authors motivate the need for explainability aligned with legal reasoning and propose an approach that combines a philosophical heuristic (a Golden Rule–style harm-aversion test) with language models to detect interests/duties and to classify Hohfeldian relations. Experiments use tailored LM methods and sentence transformers on UK religious discrimination policy texts.

Two main theories of rights are contrasted: interest theory (rights protect essential human interests) and will theory (rights confer control over others’ duties). The study adopts interest theory because it avoids confounds related to capacity, enforceability, and political discretion, enabling a more fundamental representation. Hohfeld’s taxonomy offers irreducible, logically necessary first-order relations—rights–duties and privileges–no-rights—separated from normative/political entanglements. Alternative models (e.g., Honoré, Terry, Salmond) can conflate relations or omit key correlatives. While Hohfeld’s scheme has faced critiques, it remains widely regarded for analytical clarity. Epistemologically, the paper positions its approach at a common denominator across legal formalism, positivism, realism, and critical legal studies: law structures power relations via rights and duties tied to interests. Hohfeld’s semiotic system is theory-neutral regarding interest vs will theories, and interest theory can be consistent with legal positivism. Prior related NLP/legal informatics work includes attempts to annotate Hohfeldian relations and to use norm/FrameNet-based approaches, which have struggled with scalability and mapping, suggesting potential advantages for distributional ML methods.

The methodology comprises three studies using language models and custom embeddings to operationalize interest-theory duty allocation and classify Hohfeldian relations. Study 1 (masked language modeling): The authors encode a harm-aversion heuristic via ALBERT by masking a key token in transformed sentences (e.g., “a man would [MASK] like to be murdered”) and assessing whether the model predicts negation for typically harmful acts and affirmation for beneficial ones. They generate 200 test sentences (100 typically wanted, 100 typically unwanted) using random word generators and evaluate ALBERT’s top prediction against expected polarity. Study 2 (custom sentence embedding formulations): Using the Universal Sentence Encoder, they compare test sentences to constructed synonym/antonym poles that encode wanted vs unwanted (e.g., require vs despise, happy vs unhappy, demand continue vs demand stop, wish continue vs wish stop). They build composite vectors by adding/subtracting these poles and compute cosine similarity with test sentences to assign wanted/unwanted labels. They also test an alternative where four pole vectors are used independently as features; a PCA reduces dimensionality followed by logistic regression. A dataset of 100 simple SVO sentences (previously collected with consent) is used. Study 3 (classifying Hohfeldian first-order relations): They compile 200 labelled sentences from UK anti-religious discrimination policy: 100 rights/duties and 100 privileges/no-rights (expert-labelled). Sentences are embedded using multiple sentence-transformer models (e.g., paraphrase-mpnet-base-v2, all-mpnet-base-v2, all-MiniLM-L12-v2, bert-base-nli-mean-tokens, stsb-distilbert-base, all-distilroberta-v1). UMAP projections visualize clustering, and logistic regression evaluates separability. They also fine-tune/classify with SetFit on an 80:20 split and test a LegalBERT model (legalbert-large-1.7M-2). Hyperparameters include batch size 16, 20 iterations, and 4 epochs.

- Study 1 (ALBERT masked LM): Overall accuracy 86.5% on 200 sentences. Of 100 wanted sentences, 99% correctly predicted; of 100 unwanted sentences, 74% correctly predicted (26% misclassified). Ambiguities arose from randomly generated contexts (e.g., occupational contexts for “embraced,” “caressed,” “brunched”). - Study 2 (custom USE embeddings): Composite-vector approach achieved 79.5% accuracy on 200 sentences (22% of wanted and 19% of unwanted misclassified per confusion matrix). Using four pole vectors as independent features with PCA + logistic regression yielded 72.0% accuracy. - Study 3 (Hohfeld classification): Logistic regression on transformer embeddings produced accuracies ranging from 79.3% to 89.7% across models. Fine-tuned LM classification (SetFit) achieved: paraphrase-mpnet-base-v2 92.5%, all-mpnet-base-v2 90.0%, all-MiniLM-L12-v2 90.0%, bert-base-nli-mean-tokens 87.5%, stsb-distilbert-base 85.0%, all-distilroberta-v1 85.0%. LegalBERT (legalbert-large-1.7M-2) reached 91.7%. - The best-performing model was paraphrase-mpnet-base-v2 (92.5%), slightly outperforming the legal-domain-pretrained model.

Findings demonstrate feasibility of capturing principal legal dimensions with ML: operationalizing a harm-aversion heuristic to assign duties/interests and distinguishing subtle Hohfeldian first-order relations. Performance levels (72–86.5% for interest-theory duty allocation tasks; up to 92.5% for Hohfeldian classification) suggest that language models encode sufficient relational cues to map legal constructs central to rights analysis. The approach can support explainability by grounding predictions in explicit rights/duties vs privileges/no-rights mappings rather than black-box outputs. Surprisingly, a general-purpose sentence transformer (paraphrase-mpnet-base-v2) outperformed a legal-domain model, possibly reflecting imprecisions in legal texts or the strengths of SBERT-based semantic similarity training. Ethically, embedding rights/duties and harm-aversion may help mitigate risks associated with opaque predictive analytics by making legal relations explicit and reviewable, supporting augmented intelligence where humans remain final decision-makers. The method aligns with diverse jurisprudential theories by focusing on power relations and interests that underpin legal rights and duties.

The paper shows that ML can identify core legal dimensions in text: (1) sentences that warrant duty allocation under interest theory via an operationalized ethical heuristic, and (2) Hohfeldian first-order relations, distinguishing rights–duties from privileges–no-rights with high accuracy using sentence transformers. This contributes a principled, explainable legal-analytic layer that can inform downstream legal NLP tasks. Future work includes handling more complex syntax with NER and coreference resolution, expanding and diversifying corpora, refining masked-LM prompting (e.g., via constrained options), and leveraging relational structure (e.g., RotatE for opposites/correlatives) and semi-supervised clustering to improve Hohfeldian labeling and scalability.

Limitations include: (1) Heuristic sensitivity to sentence templates and perspectives (e.g., arrest scenarios require reframing to the underlying harm principle); masked LM outputs can be syntax-dependent and require constrained options to avoid mispredictions. (2) The Study 2 dataset was small and potentially biased (limited authors, SVO simplicity), and personality/writing differences may influence results. (3) Perfect Hohfeldian allocations are challenging for current LMs; better exploitation of correlative/opposite structures and knowledge-graph-style embeddings (e.g., RotatE) may help. (4) Generalization to complex, real-world legal texts will require integration of NER, co-reference resolution, and larger, diverse datasets.