MasonNLP System: Modular NLP Framework

• MasonNLP is a modular natural language processing framework that unifies rule-based and neural methods for robust information extraction from engineering, medical, and social media texts.
• It leverages a multi-stage pipeline integrating classical parsing, transformer models, and retrieval-augmented generation to produce structured outputs like SysML diagrams and clinical orders.
• The system demonstrates competitive performance across benchmarks, offering scalable and adaptable solutions for diverse, data-intensive applications.

The MasonNLP System is a comprehensive, modular NLP framework for information extraction and structured modeling from unstructured textual sources, with a particular emphasis on engineering, medical, and social media domains. Its design unifies classical rule-based extraction (syntactic parsing, chart parsing, information extraction) and modern neural architectures (transformer models, retrieval-augmented LLMs) to automate tasks such as entity extraction, requirements modeling, clinical order detection, medical question answering, and semantic textual relatedness. The system’s evolution traces connections from early constraint-based parsing pipelines to prompt-engineered LLMs and multimodal retrieval-augmented generation, yielding strong baseline and competitive results in diverse information extraction benchmarks.

1. NLP Foundations and Architectural Principles

MasonNLP incorporates a multi-stage, modular pipeline architecture. Early system modules derive from structured parsing and symbolic processing, such as chart parsing with constraint-augmented context-free grammars (CFGs) following Earley’s algorithm. These modules tokenize, parse, and classify objects of interest (e.g., noun phrases, entities, functions, attributes) using a dictionary-augmented CFG and well-formedness constraints, including agreement checks (e.g., number-agreement(NP, VP)) and selective compound noun recognition via pattern matching (Macdonell et al., 2014).

Subsequent system generations embrace neural architectures for semantic representation, exploiting transformer-based LLMs (e.g., BERT, RoBERTa, domain-adapted variants such as BioBERT, MentalBERT) that are fine-tuned or domain-pretrained for specific applications, including detection of depression symptoms, semantic relatedness, and medical extraction (Ramachandran et al., 2023, Sakib et al., 2023).

The MasonNLP architecture supports both monolithic and ensemble workflows. Recent iterations integrate large, general-purpose instruction-tuned LLMs (e.g., LLaMA-4 17B) with prompt engineering and minimal in-context learning to provide structured outputs for clinical and engineering texts without domain-specific retraining (Karim et al., 12 Oct 2025, Karim et al., 12 Oct 2025). Key system features include:

• Layered modularity: Separate processing for tokenization, syntactic parsing, term management, knowledge base population, and diagram generation.
• Extensibility: Incorporation of both statistical feature engineering and neural embeddings, supporting adaptation to new domains with minimal reconfiguration.
• Schema-driven extraction: Definition of structured output formats for downstream informatics (e.g., tuple-based schema for medical orders, SysML diagrams).

2. Extraction Methodologies and Technical Processes

The system supports both syntactic and semantic extraction, progressing from rule-driven to data-driven methods:

• Classical Extraction: Early modules employ sentence tokenization, constraint-based chart parsing, noun phrase extraction, PCFGs, and assignment of terms to semantic categories (entity, function, attribute). A term management module facilitates automated and interactive term curation and knowledge base construction. Entity and relationship extraction utilize grammatical rules and interactive refinement by domain experts (Macdonell et al., 2014).
• Statistical Term Scoring: For systems engineering tasks, MasonNLP employs term frequency–inverse document frequency (tf–idf) weighting and WordNet-based semantic depth scoring for nouns; key phrase scoring aggregates these metrics using the formula

$$\lambda_{p,k} = (\Sigma w_{t,k} / N_p) + (\Sigma h_{t,k} / N_p) + \text{count}_{p,k}$$

where $w_{t,k}$ is tf–idf, $h_{t,k}$ is one-complement normalized depth, and $N_p$ is the number of terms in phrase $p$ (Zhong et al., 2022).

• Open Information Extraction (OpenIE): Relationship extraction leverages semantic role labeling, noun-based relational extraction, and confidence thresholding ($\sigma_{\text{relationship}}$) to populate a connection schema (Zhong et al., 2022).
• Neural Embeddings for Classification and Ranking: For noisy and large-scale social media corpora, the system uses transformer-derived embeddings (e.g., MentalBERT) and cosine similarity for ranking text segments against reference queries or symptom descriptions (Sakib et al., 2023).
• Prompt Engineering for LLMs: Recent clinical and multimodal VQA modules rely on structured prompt engineering with in-context demonstrations and explicit schema definitions for tuple extraction, ensuring output consistency in complex dialogues (Karim et al., 12 Oct 2025, Karim et al., 12 Oct 2025).
• Retrieval-Augmented Generation (RAG): In multimodal VQA, textual and visual exemplars are indexed (FAISS, MiniLM/CLIP), retrieved by similarity (with balanced $\alpha$-weighted fusion), and prepended to input prompts to provide clinical grounding and minimize hallucinations (Karim et al., 12 Oct 2025).

3. Structured Output and Diagram Generation

MasonNLP supports automated generation of structured diagrams and knowledge representations from textual input, crucial for systems engineering, requirements analysis, and knowledge-intensive domains:

• SysML Diagram Generation: The pipeline automatically extracts entities and relationships from unstructured technical documents, mapping high-confidence key phrases to SysML block elements and categorizing inter-block relations as composite, generalization, or reference using string overlaps and WordNet-based semantic relations. Augmentation applies abstraction (removal of low-score words) and semantic enrichment from hypernym/hyponym relations (Zhong et al., 2022).
• Clinical Order Schema Extraction: In clinical dialogue transcripts, the model outputs tuples in the schema $$(\text{order\_type}, \text{description}, \text{reason}, \text{provenance})$$, where provenance is a set of turn IDs linking each order to supporting utterances (Karim et al., 12 Oct 2025).
• Medical Wound VQA Schema: For medical VQA, both free-text answers and structured wound attributes are generated, conforming to a schema

$$o_e = (\text{resp}_e, \text{loc}_e, \text{type}_e, \text{thick}_e, \text{color}_e, \text{drainAmt}_e, \text{drainType}_e, \text{infect}_e)$$

ensuring machine-readability and compatibility with health informatics systems (Karim et al., 12 Oct 2025).

Automated diagramming leverages PlantUML and GraphViz backends, supporting rapid visualization for user feedback loops and iterative refinement.

4. Performance, Benchmarking, and Evaluation

Empirical results across multiple tasks and domains are as follows:

• Classical Parsing/Extraction: Term extraction precision and recall in SysML diagram tasks achieve rates from 55% to >90% (precision) and 50%–82% (recall), with relationship mapping accuracy between 64%–85%, varying by domain and input heterogeneity (Zhong et al., 2022).
• Sentiment and Information Extraction in Social Media: Information extraction density and efficiency in Lithium NLP—an architecture influencing MasonNLP’s social media modules—yields 2.8 times more entities per kilobyte and 22 ms per 1 KB throughput, with an F1-score of 73% for entity disambiguation (Bhargava et al., 2017).
• Depression Symptom Detection: A dual-stage filter using RoBERTa and LSTM achieves 92%/97% validation accuracy respectively in filtering candidate sentences, with subsequent MentalBERT-based ranking producing AP = 0.035, R-Precision = 0.072, Precision@10 = 0.286, NDCG@1000 = 0.117 (majority voting) (Sakib et al., 2023).
• Structured Medical Order Extraction: A prompt-engineered, few-shot LLaMA-4 17B model achieved an average F1 score of 37.76, with especially strong gains in reason and provenance extraction compared to zero-shot or smaller model baselines (Karim et al., 12 Oct 2025).
• Medical Visual QA: Lightweight RAG with LLaMA-4 produced an average aggregate score of 41.37% and competitive dBLEU, ROUGE, and BERTScore figures, ranking 3rd of 19 in the MEDIQA-WV 2025 shared task; ablation confirms the contribution of multimodal retrieval (Karim et al., 12 Oct 2025).
• Tool-Calling and Modular Interaction: Replacing programmatic JSON tool calls with a natural language selection framework (YES/NO per tool) improves tool call accuracy by 18.4 percentage points and reduces output variance by 70%, particularly for open-weight LLMs (Johnson et al., 16 Oct 2025).

Evaluation protocols are tailored to each domain, employing field-specific F1 (including ROUGE-1 and MultiLabel F1), cosine similarity for embedding comparison, Spearman correlation for semantic textual relatedness, and metrics reflecting both free-text and structured accuracy.

5. Comparative Analysis and System Evolution

MasonNLP’s design is informed by both legacy and contemporary NLP systems:

• Early Systems: The integration of chart parsing, constraint-driven extraction, and interactive knowledge curation draws on foundational prototypes for requirements analysis and interactive term management (Macdonell et al., 2014).
• Neural Pipelines and Ensembles: The adoption of transformer models, statistical ML methods (ElasticNet, linear regression), unsupervised embeddings, and weighted ensemble approaches reflects best practices for high-dimensional, multilingual, and cross-lingual settings (Goswami et al., 22 Mar 2024).
• RAG and Prompt Engineering: Advancements in prompt engineering, schema-driven data annotation, in-context learning, and lightweight RAG frameworks enable the deployment of domain-agnostic LLMs to specialized tasks without retraining, positioning MasonNLP as a scalable system for domain adaptation (Karim et al., 12 Oct 2025, Karim et al., 12 Oct 2025).
• Tool-Calling Innovations: Employing natural language interfaces for modular tool invocation offers robustness and makes MasonNLP adaptable to multi-domain and safety-critical tasks, surpassing rigid function-calling approaches in accuracy and variance (Johnson et al., 16 Oct 2025).

Comparisons with contemporary systems underscore MasonNLP’s balance between automation (full-pipeline integration and rapid visualization), openness to manual curation (expert-in-the-loop refinement), and modular extensibility for emerging tasks in engineering and medicine.

6. Practical Applications and Impact

MasonNLP is leveraged in diverse domains, with applications and benefits including:

• Systems Engineering: Automation of SysML diagram creation from technical documentation, resulting in more standardized and comprehensive system models and mitigating human error and inefficiency in early-stage design (Zhong et al., 2022).
• Clinical Informatics: Structured extraction of medical orders, wound assessment responses, and clinical knowledge from dialogic transcripts and medical images, supporting decision support, documentation, and telemedicine workflows (Karim et al., 12 Oct 2025, Karim et al., 12 Oct 2025).
• Social Media Mining: Extraction of mental health signals, sentiment, and social-informational entities from noisy, user-generated content using transformer ensembles and domain-adaptive techniques (Sakib et al., 2023, Ramachandran et al., 2023, Bhargava et al., 2017).
• Tool-Enhanced LLM Agents: Modular tool selection in customer service and mental health agents, with reduced overhead and improved accuracy via natural language tool-calling (Johnson et al., 16 Oct 2025).

By combining rigorous classical NLP, data-driven neural techniques, retrieval augmentation, and practical modular design, MasonNLP delivers robust and adaptable solutions for real-world structured information extraction, meeting the requirements of both engineering and clinical stakeholders.