qproj wikioAMF
The Open Argument Mining Framework (oAMF) is a toolkit from the Centre for Argument Technology (ARG-tech) at the University of Dundee. It splits argument mining into small steps, such as segmenting text, extracting propositions and finding the relations between them, and lets any module for a step be swapped for another. Every module reads and writes xAIF, the format Quell’s argument maps use.
Contents
How it works
Each module is a small web service in a Docker container. It receives an xAIF document, adds its own layer of analysis and returns the document. A pipeline is a graph of modules: a turn separator feeds a segmenter, the segmenter feeds a propositionaliser, and so on. Pipelines can be built with the Python library (oamf on PyPI), in a drag-and-drop editor based on n8n, or through a web interface. Modules are either deployed locally from their GitHub repositories or called on ARG-tech’s servers. New modules start from a template project, AMF_NOOP.1
Fixing the format and keeping the modules independent is the point of the design. One step can be replaced or compared without touching the rest of the pipeline, and modules built on different theories of argument can still be combined.2
Modules
The framework launched with seventeen modules:3
| Step | Modules | Approach |
|---|---|---|
| Turns | DTSG | Splits a transcript into speakers’ turns. |
| Segments | DSG, TARGER, DSS | Divide text into locutions. DSS prompts DeepSeek-R1 1.5B with a few examples. |
| Anaphora | DARJ | Resolves pronouns in locutions. |
| Propositions | SPG, CPJ | Extract the proposition each locution asserts. CPJ is a cascade that resolves anaphora, checks for reported speech and reconstructs the subject. |
| Relations | DAMG, DTERG, SARIM, ARIR, TARGER-AM, DRIG, DSRM | Find support, conflict and rephrase between propositions. ARIR is a fine-tuned RoBERTa classifier over pairs (none, inference, conflict, rephrase), DRIG a fine-tuned DialoGPT, DSRM DeepSeek-R1 1.5B prompted with examples. |
| Schemes | WSCR, PDSCZ | Label inferences with Walton’s schemes, or with pragma-dialectics’ comparison, symptomatic and causal argumentation. |
| Statement type | PTCR | A fine-tuned RoBERTa classifier: fact, value or policy. |
Three tools sit alongside the modules. CASS compares two xAIF maps of the same text and reports macro F1, accuracy, kappa, U-alpha and text similarity. WSTT transcribes audio with Whisper, and SV draws a map as SVG.
Performance
ARG-tech evaluated three pipelines on a corpus of student essays (AAEC) that none of the modules were trained on. The best found argument components with an F1 of 56.32 and relations with 32.65. The best end-to-end systems they were compared with, trained on the same corpus, reached 76.55 and 54.66. On the harder cross-corpus relation task, one oAMF pipeline scored 47.37, ahead of the previous best of 42.00.4 Finding relations remains the weak step, as it is across the field.
The modules differ greatly in speed. For a single relation-identification sample, DSRM, the DeepSeek module, took 19.461 seconds, and DTERG, which is BERT-based, took 0.345 seconds.5
In Quell
See also
- Argument mining — the research field oAMF serves.
- Inference Anchoring Theory and xAIF — the dialogue model behind xAIF.
- Stage-wise evaluation — how candidate approaches are to be scored.
References
- Gemechu, D., Ruiz-Dolz, R., Gorska, K., Moslemnejad, S., Maguire, E., Zografistou, D., Jo, Y., Lawrence, J., & Reed, C. (2025a). The Open Argument Mining Framework. In Proceedings of the 63rd Annual Meeting of the ACL (Volume 3: System Demonstrations), 318–328. ACL Anthology.
- Gemechu, D., Ruiz-Dolz, R., Lawrence, J., & Reed, C. (2025b). Practical Solutions to Practical Problems in Developing Argument Mining Systems. In Proceedings of the 12th Argument Mining Workshop, 100–106. ACL Anthology.
External links
- oAMF on GitHub — module list and documentation.
- oamf on PyPI.
- oAMF web interface.
What links here: Argument mining