cs.CLOct 8, 2025
SaveReasoning for Hierarchical Text Classification: The Case of Patents
Organizations: University of Cambridge · National Science Library, Chinese Academy of Sciences · Department of Information Resources Management, School of Economics and Management, University of Chinese Academy of Sciences
Abstract
Hierarchical text classification (HTC) assigns documents to multiple levels of a pre-defined taxonomy. Automated patent subject classification represents one of the hardest HTC scenarios because of professional difficulties and extensive labels. Prior approaches only output a flat label set, which offers little insight into the reason behind predictions. Therefore, we propose Reasoning for Hierarchical Classification (RHC), a novel framework that reformulates HTC as a step-by-step reasoning task to sequentially deduce hierarchical labels. RHC trains large language models (LLMs) in two stages: a cold-start stage that aligns outputs with chain-of-thought (CoT) reasoning format and a reinforcement learning (RL) stage to enhance multi-step reasoning ability. RHC demonstrates four advantages in our experiments. (1) Effectiveness: RHC surpasses previous baselines and outperforms the supervised fine-tuning counterparts by approximately 3% in accuracy and macro F1. (2) Explainability: RHC produces natural-language justifications before prediction to facilitate human inspection. (3) Scalability: RHC scales favorably with model size with larger gains compared to standard fine-tuning. (4) Applicability: Beyond patents, we further demonstrate that RHC achieves state-of-the-art performance on other widely used HTC benchmarks, which highlights its broad applicability.
Figures & tables
| Hierarchical Level | #Labels | Example Label | Example Label Description |
| Section | 8 | A | Human necessities |
| Class | 129 | A01 | Agriculture; forestry; animal husbandry; hunting; trapping; fishing |
| Subclass | 639 | A01C | Planting; sowing; fertilising |
| Group | 7,314 | A01C 3 | Treating manure; manuring |
| Subgroup | 61,397 | A01C 3/06 | Manure distributors, e.g., dung distributors |
Table 1: Example of International Patent Classification (IPC) scheme ( WIPO, 2025 ) .
Figure 1: Overview of our Reasoning for Hierarchical Classification (RHC) method compared to traditional method. While traditional methods only predict labels, RHC deduces labels at each level with justifications.
| Dataset Split | Source | Year | #Section | #Class | #Subclass | #Docs | Text Type |
| US Patents (Train) | USPTO | 2011–2017 | 8 | 117 | 500 | 22,500 | Claims |
| US Patents (Test) | USPTO | 2011–2017 | 8 | 117 | 500 | 2,500 | Claims |
| EU Patents (Test) | EPO | 2024 | 8 | 115 | 437 | 2,845 | Claims |
Table 2: Statistics of our PCD-BD dataset for patent classification.
| US Patents | EU Patents | |||||||||||
| Model / Setting | Section | Class | Subclass | Section | Class | Subclass | ||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | |
| Baselines (SFT) | ||||||||||||
| BERT ( Lee and Hsiang, 2020 ) | 77.1 | 76.0 | 60.8 | 55.0 | 45.5 | 44.1 | 69.1 | 66.7 | 50.8 | 45.8 | 36.9 | 30.8 |
| XLNET ( Haghighian Roudsari et al., 2022 ) | 77.8 | 77.2 | 62.4 | 59.4 | 49.2 | 48.5 | 71.8 | 69.8 | 55.8 | 50.7 | 42.4 | 37.2 |
| BERT-for-Patent ( Chikkamath et al., 2022 ) | 82.1 | 82.1 | 68.7 | 66.7 | 56.7 | 56.1 | 76.1 | 74.6 | 61.6 | 58.0 | 49.6 | 43.9 |
Table 3: Patent subject classification performance (Accuracy & Macro F1) of different models on US and EU patent test sets. The best result of each column is highlighted in bold , while the second-best result is marked underlined .
| Input Claim: 1. A multi-phase clock signal generator, comprising: a delay-locked loop having a first multi-tap adjustable delay line configured to delay a reference signal to provide a plurality of clock signals having different phases relative to the reference clock signal; and a second multi-tap adjustable delay line coupled to the delay-locked loop to receive a delay control signal and a periodic signal having a 50-percent duty cycle and the same period as the reference clock signal, the second multi-tap adjustable delay line configured to delay the periodic signal according to the delay control signal from the delay-locked loop to provide a plurality of clock signals having different phases relative to the periodic signal. |
| Gold Main Label: H03L |
| Model Output: Step 1 — Section Brief Justification: The invention concerns electronic signal generation and control using electronic circuits (delay-locked loop, delay lines). Decision: H . Step 2 — Class Brief Justification: It focuses on fundamental electronic circuitry for generating and manipulating clock signals, which falls under basic electronic circuit arrangements. Decision: H03 . Step 3 — Subclass Brief Justification: The core is a delay-locked loop with delay lines for phase control—i.e., automatic control of signal phase/oscillation frequency. Decision: H03L . |
| H: Electricity |
| H03: Electronic circuitry |
| H03L: Automatic control, starting, synchronisation, or stabilisation of generators of electronic oscillations or pulses |
Table 4: An example of model input and correct output. The reasoning path is correct and well-aligned with the gold IPC labels. The explanations are readable, and the justification can help human examiners assess the decisions.
Figure 2: SFT and RHC results of different Qwen-2.5 model sizes (0.5B, 3B, 7B) on US and EU test sets. Performance differences between RHC and SFT become larger as model size increases.
| Model / Setting | Micro F1 | Macro F1 |
| Previous SOTA | ||
| HCL-MTC ( Zhang et al., 2025 ) | 86.47 | 81.02 |
| Ours | ||
| Qwen-2.5-7B-SFT | 85.15 | 82.54 |
| Qwen-2.5-7B-RHC | 87.13 | 83.82 |
Table 5: Results of hierarchical text classification on the WOS dataset.
| US Patents | EU Patents | |||||||||||
| Model / Setting | Section | Class | Subclass | Section | Class | Subclass | ||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | |
| Llama-3.1-8B-RHC | 83.1 | 82.5 | 70.9 | 66.0 | 59.0 | 57.1 | 75.5 | 73.6 | 61.2 | 55.9 | 48.4 | 41.4 |
| w/o Base SFT | 81.6 | 81.3 | 68.0 | 61.8 | 53.5 | 49.2 | 76.1 | 75.0 | 60.3 | 53.4 | 46.9 | 38.5 |
| w/o Cold Start | N/A | |||||||||||
| w/o GRPO | 75.4 | 74.7 | 58.5 | 50.1 | 45.2 | 39.4 | 71.1 | 70.0 | 54.2 | 46.5 | 40.6 | 32.1 |
Table 6: Ablation study results of patent subject classification on US and EU patents. Base SFT provides the model with basic classification ability. GRPO substantially enhances reasoning-based classification. The format reward ensures that the model generates structured reasoning paths rather than bare labels.
| US Patents | EU Patents | |||||||||||
| Model | Section | Class | Subclass | Section | Class | Subclass | ||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | |
| Qwen-2.5-0.5B-SFT | 80.6 | 80.2 | 66.5 | 63.1 | 53.2 | 52.4 | 74.7 | 72.6 | 57.6 | 52.9 | 44.3 | 39.1 |
| Qwen-2.5-0.5B-RHC | 80.9 | 80.5 | 66.4 | 62.2 | 52.8 | 51.1 | 74.4 | 72.7 | 57.9 | 52.1 | 45.2 | 38.4 |
| Qwen-2.5-3B-SFT | 81.0 | 80.6 | 68.5 | 65.4 | 56.0 | 55.3 | 74.3 | 73.2 | 59.6 | 54.8 | 47.6 | 42.0 |
| Qwen-2.5-3B-RHC | 81.8 | 81.8 | 69.2 | 64.5 | 57.3 | 56.4 | 75.6 | 73.6 | 60.8 | 56.1 | 48.6 | 42.6 |
Table 7: Results of different Qwen-2.5 model sizes (0.5B, 3B, 7B) on US and EU patents. The best result of each column is highlighted in bold .
Figure 3: Qwen-2.5-7B-RHC results of different number of cold start data (1k, 3k, 6k) on US and EU patents.
| Example 1 |
| Input Claim: 1. An isolated regulatory element that drives transcription in a constitutive manner, wherein the regulatory element comprises a nucleotide sequence selected from the group consisting of: a) the nucleotide sequence of SEQ ID NO: 1, 2 or 3; and b) a sequence comprising a functional fragment of the nucleotide sequence set forth in a) 2. An expression cassette comprising a regulatory element operably linked to a nucleotide sequence wherein the regulatory element comprises the nucleotide sequence of claim 1. 3. A plant stably transformed with an expression cassette of claim 2. 4. The plant of claim 3, wherein said plant is a monocot. 5. The plant of claim 3, wherein said monocot is maize, wheat, rice, barley, sorghum or rye. 6. Seed of the plant of claim 3 wherein the seed comprises the expression cassette. 7. A method for selectively expressing a nucleotide sequence in a plant cell, the method comprising: a) transforming a plant cell with an expression cassette, the expression cassette comprising a regulatory element operably linked to a nucleotide sequence wherein the regulatory element comprises a nucleotide sequence selected from the group consisting of: i) the nucleotide sequence set forth in SEQ ID NO: 1, 2 or 3; and ii) a functional fragment of i); and b) growing the plant cell to selectively express the nucleotide sequence. 8. The method of claim 7 wherein the regulatory element initiates expression of the nucleotide sequence in plant tissue. 9. The method of claim 7 further comprising regenerating a stably transformed plant from the plant cell; wherein expression of the nucleotide sequence alters the phenotype of a plant. 10. The plant of claim 9, wherein said plant is a monocot. 11. The plant of claim 9, wherein said monocot is maize, wheat, rice, barley, sorghum or rye. 12. Seed of the plant of claim 9 wherein the seed comprises the expression cassette. |
| Gold Main Label: A01H |
| Model Output: Step 1 — Section Brief Justification: The invention concerns genetic elements (nucleotide sequences, regulatory elements) used to modify gene expression in plants, which falls under biotechnology and genetic engineering. Decision: C . Step 2 — Class Brief Justification: It specifically relates to biotechnology and genetic engineering constructs (expression cassettes, transformed plant cells) used to alter phenotypes. Decision: C12 . Step 3 — Subclass Brief Justification: The core is DNA/RNA sequences and expression constructs (regulatory elements, nucleotide sequences) for transforming plant cells and obtaining stably transformed plants, including seeds and methods of expression. Decision: C12N . |
| A: Human necessities |
| A01: Agriculture; forestry; animal husbandry; hunting; trapping; fishing |
| A01H: New plants or processes for obtaining them; plant reproduction by tissue culture techniques |
Table 8: Examples of model inputs and failed outputs.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| CoT Prompt: |
| You are an expert patent classification specialist. Your task is to predict a patent’s main IPC (International Patent Classification) code step by step, based on patent claims. |
| - Reason through each hierarchical level: Section Class Subclass. |
| - Output exactly one decision per level. |
| - Wrap the code of each level in \box{} . |
| - Provide a brief justification before predicting the code at each step. |
Table 9: Prompts used for hierarchical IPC prediction and CoT data generation.
| CoT Prompt: |
| You are an expert scientific topic classification specialist (Web of Science style). |
| Your task is to predict Level-1 field and Level-2 subfield labels step by step, based on abstracts. You must reason through each hierarchical level: field → subfield. |
| Instructions: |
| - Output exactly one decision per level. |
Table 10: Prompts used for WOS prediction and CoT data generation.
More in this topic
Hierarchical Classification
Hierarchical classification uses relationships among labels to predict categories at multiple levels, from broad groups to fine-grained classes.
See all 89 papers