Technical language automation with industry
Before there were themes there were projects, each of which put a text classifier into a working system and left a question its paper does not contain.
OneThe projects
Technical text, the terse, abbreviation-heavy language of repair logs, cost documents and support tickets, is where I have done most of my applied work, always with a partner who needed the classifier to run. Avionics maintenance text with GE Aerospace, where TF-IDF and pre-trained word embeddings were compared inside sequence classifiersa for predicting the maintenance action from the repair narrative [1]. Bills of quantities in construction, where free-text cost lines were mapped onto the International Construction Measurement Standard for an Innovate UK programme, with a £1.2M UWE share, and the trained model was released publicly for the industry to use [2]. University helpdesk tickets, classified so that the auto-reply could be tailored to the problem [3]. Farmers' complaints in Egypt, routed through a cloud decision-support system built with Fayoum University under a Newton-Mosharafa programme [4].
TwoWhat they left behind
The aerospace work left the observation that a single repair generates several records, written by different people at different stages of the workflow, and that which of them a model reads is settled before any model comparison starts; that observation grew into the records theme. The construction work made the labels the interesting part: the categories came from a published measurement standard rather than from the data, so the fit between what a cost line says and what the standard can express was itself the problem being solved. The helpdesk work tied the classes to the replies the service could send, so the decision the system had to take shaped the label set from the start. Working inside these systems, rather than on benchmark data, is what turned measurement into my research question.
ThreeWhere they could restart
- The construction classifier, rebuilt: The code is public [2]. Put a current fine-tuned encoder against the published baseline on the same task and report what four more years of language models actually buy on short, telegraphic cost lines. MSc
- A record audit in a new industry: Take a domain with a working classifier, map which records of a case exist, who writes each one and at what stage, and which of them the model reads. It is the fieldwork behind the records theme, done fresh where nobody has looked. MSc
- Helpdesk replies judged at the decision: Tailored auto-replies evaluated on whether the problem was resolved, not on whether the class was right, extending the label-set question the helpdesk project opened [3]. MSc or PhD
Notes
a An embedding represents each word as a vector of numbers, placed so that words used in similar ways sit near one another; TF-IDF is the older approach of weighting each word by how distinctive it is. A sequence classifier reads those vectors in order to assign the whole passage a category. ↩
References
- Mayhew, P., Ihshaish, H., Deza, J. I. and del Amo, A. (2023). Maintenance automation using deep learning methods: a case study from the aerospace industry. Artificial Neural Networks and Machine Learning, ICANN 2023, LNCS 14263, 295-307. doi:10.1007/978-3-031-44204-9_25
- Deza, J. I., Ihshaish, H. and Mahdjoubi, L. (2022). A machine learning approach to classifying construction cost documents into the International Construction Measurement Standard. arXiv:2211.07705. Code: github.com/ihshaish/BoQ-classifier-ICMS
- Nicholls, R., Fellows, R., Battle, S. and Ihshaish, H. (2022). Problem classification for tailored helpdesk auto-replies. ICANN 2022. doi:10.1007/978-3-031-15937-4_37
- Munir, K., Ghafoor, M., Khafagy, M. and Ihshaish, H. (2022). AgroSupportAnalytics: a cloud-based complaints management and decision support system for sustainable farming in Egypt. Egyptian Informatics Journal, 23(1), 73-82. doi:10.1016/j.eij.2021.06.002