Conversational Construction Site Management & Reporting: Extracting Structured Data from Informal Natural-Language Messages
Abstract
Objective: Construction site teams report material movement, labour use, expenses, payments and progress in short, informal messages, which are usually converted into structured records by hand. This paper proposes and evaluates a conversational pipeline that turns such messages into validated operational records.
Methodology: The system normalizes language, extracts intents and entities, maps them to construction schemas, matches entities against application state, validates the result and assigns a confidence value before a record is created. It was tested on a synthetic benchmark of 300 records in English, MarathiEnglish code-mixed, transliterated Marathi and informal or noisy text.
Results: On the 75-record held-out test set, intent accuracy was 61.33%, entity macro F1 41.47%, entity micro F1 45.01% and exact match 1.33%. On a 12-record multi-action subset with 17 ground-truth intent instances, intentinstance accuracy was 70.59% and action F1 82.35%. Limitations: The evaluation is a laboratory study on synthetic data. The results are not claimed to hold on real construction sites, nor to outperform large language models.
Originality: The paper offers a construction-specific pipeline for informal, multilingual and multi-action site messages, together with a benchmark, a configuration ablation and an error analysis.
KEYWORDS: Construction management; Natural language processing; Information extraction; Conversational systems; Code-mixed language; Entity matching; Intent detection; Marathi
Full Text:
PDF 76-92Refbacks
- There are currently no refbacks.