Rhetoric, Writing, And Anexact Architecture: The Experiment Of Natural Language Processing (Nlp) And Computer Vision (Cv) In Architectural Design

Yuxin Lin University of Michigan

Language has the power to create novel, generative architectures that could change the shape of the world, and writing aids the iterative process of creation (Steenson, 2017). Greg Lynn believes that geometry is rhetoric in architectural writing and uses “anexact yet rigorous” forms to establish a new geometric understanding in architecture (Lynn, 1998). With Artificial Intelligence’s help, verbalizing innate and intuitive space feelings and seeking anexact forms to do architectural design becomes practical. This article aims to demonstrate (1) the ability of AI to do architectural writing through Natural Language Processing and (2) its potential to interpret the texts into anexact forms under the topic of the coexistence between Computer Vision and humans. The original contribution includes (1) proving that AI can synthesize the sense of order inside the texts that describe different spatial forms and (2) giving architecture the ability to measure and visualize uncertain, heterogeneous writing characteristics through anexact forms. The selection of neural networks that can achieve NLP and do image-based translation is critical in this article. Also, the content and structure of the dataset are under consideration since a post-carbon future of the architecture requires sustainability and multifunctionality. Furthermore, human-machine collaboration is inevitable and necessary to turn unexpected outputs into conceptualized designs. Decision-making about using AI-generated results requires further investigation. The whole process is from inexact to exact and finally anexact. GPT-2, which generates synthetic text samples in response to the arbitrary input (Radford et al., 2019), was utilized in this paper. The collection of (1) architecture form dataset and (2) swarm form dataset describes sustainable and multifunctional spaces in nature helps to fine-tune GPT-2. The model then generates coherent and continuous descriptions of space by looking for inter-relationships to organize the compositions, merging two forms. Attentional Generative Adversarial Networks (Xu et al., 2018), known as AttnGAN, can translate written form into visual output, allowing to interrogate shape through language (del Campo, 2021; Yuan et al., 2019). The algorithm here helps visualize the dialogues from GPT-2 so every sentence can be related to an image containing complex shapes and forms which imply space. Then, human intelligence will interpret every image into 3D assembly unit models and eventually piece them together as the final result. The presented project, Anexact Building, combines texts, visualizes machine logic, and artificial intelligence communicating with human intelligence. AI successfully generated realistic architectural descriptions containing sustainable and multifunctional texts by analyzing the natural form language and breaking them apart along the joints to understand and recombine them through text. Also, AI-generated anexact shapes, along with the combination of a bottom-up form generation process operated by machines and a top-down approach that we make decisions as architects, helps to shape new spaces and generate new architectural styles. The next generation of the algorithm could translate input prompts into scene images directly, and image discrimination can be automated by semantic segmentation or object detection. Even more, AI could generate 3D models automatically based on arbitrary text inputs in the future. References: del Campo, M. (2021). Architecture, Language and AI-Language, Attentional Generative Adversarial Networks (AttnGAN) and Architecture Design. Lynn, G. (1998). Folds, bodies & blobs: collected essays. Books-by-architects. [Bruxelles]: La Lettre volée. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language Models are Unsupervised Multitask Learners. Steenson, M. W. (2017). Architectural intelligence: How designers and architects created the digital landscape. The MIT Press. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., & He, X. (2018). AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE. Yuan, Y., Chen, Y., Liu, Z. (2019). Architecture of {AI} Language. Unpublished master’s thesis, Taubman College of Architecture and Urban Planning, University of Michigan, Ann Arbor, Michigan.

Keywords: Rhetoric And Writing; Natural Language Processing; Computer Vision; Gpt-2; Attngan; Human-Computer Interaction; Architectural Design; Post-Carbon; Sdg3; Sdg11

View Full Paper