Ansys Medical Device Regulatory

A Hybrid Approach to Information Retrieval and Answer Generation for

Regulatory Texts Are Inherently Long And Com-

plex, presenting significant challenges for infor-

Techniques To Extract Relevant Information From

large regulatory corpora.

With The Traditional Bm25 Algorithm To Achieve

both semantic precision and lexical coverage.

Cessing Tools For Compliance-Driven Applica-

tions in regulatory domains.

Information Retrieval (Ir) Systems Are Concerned

with efficiently querying large corpora to retrieve relevant results. Traditional systems, such as search engines, often depend on term-frequency statistical methods like tf-idf, which measures the importance of a term in a document relative to its frequency in

The Corpus . Bm25 , A Well-Established Rank-

ing function, builds on similar principles to provide a scalable and effective retrieval framework. How-

Where The Semantics Often Outweigh Simple Term

matching. Regulatory content is particularly challenging due

To Its Specialized Terminology And Nuanced Lan-

guage.

Synonyms, Paraphrasing, And Domain-

specific jargon frequently obscure the relationship

Between Queries And Relevant Documents, Reducing

the effectiveness of lexical retrieval methods. Semantic search addresses these limitations by us-

Embeddings, Capturing The Semantic Meaning Of

the text in a condensed high-dimensional space .

Matches, Grouping Related Content Together Even

with different terminology.

Like Bert , Have Introduced High-Quality Con-

textual embeddings for words, sentences, and para-

Graphs Which Can Be Leveraged In Semantic Search

applications.

Despite These Advances, Building An Effective Ir

system for regulatory texts poses unique challenges.

Domain-Specific Knowledge Required For Accurate

retrieval in specialized fields. Fortunately, various

That These Base Models Can Be Fine-Tuned To Close

this gap . In this paper, we present a hybrid information re- trieval system that integrates both lexical and se- mantic approaches to address the limitations of tra-

Retrieval Augmented Generation (Rag) System

that leverages the hybrid retriever to provide com-

Prehensive And Accurate Answers To User Queries

using a Large Language Model (LLM).

Through Extensive Experiments, We Demonstrate

that the hybrid retriever achieves superior perfor-

31

Proceedings of the 31st International Conference on Computational Linguistics, pages 31–35

System Effectively Synthesizes Retrieved Content,

delivering detailed responses that address the com-

Contributions Aim To Advance Regulatory Informa-

tion retrieval and lay the foundation for more ef-

Fective Question-Answering Systems In Specialized

domains.

The Development Of An Effective Information Re-

trieval (IR) system for regulatory content requires

Related Queries. These Systems Must Return A Set

of ranked passages from the corpus that accurately address the compliance aspects of a given question.

A Widely-Used Algorithm That Ranks Results Based

on query term frequency and other statistical fea- tures. While BM25 is effective for lexical retrieval, it struggles to capture semantic relationships, par-

Ticularly In Regulatory Domains Where Terminology

often varies for the same concepts.

Our Approach Enhances Bm25 By Integrating A Text

embedding model, enabling semantic matching. This hybrid system identifies semantically relevant content that BM25 alone might overlook, offering a significant advantage in handling the complexities of regulatory language.

Dataset

The dataset used for this study, ObliQA, consists of 27,869 regulatory questions extracted from 40 doc- uments provided by Abu Dhabi Global Markets.

Vices Within The European Economic Area, Making

the dataset highly relevant for compliance-related tasks .

The Dataset Is Divided Into Three Subsets: Train-

ing (22,295 questions), testing (2,786 questions),

The Relevant Information Needed To Answer It. The

data is stored in JSON format, where each entry in- cludes the question, associated passages, and their metadata. An example is shown below.

A Bert-Based Model Trained On General-Purpose

data. The fine-tuning process employed a loss func-

Lowed By Pooling And Normalization Layers. To

better capture semantic nuances in regulatory texts,

We Increased The Embedding Dimension From 384

to 512.

Former Library . The Model Was Trained Over 10

epochs with a batch size of 64, using a learning rate

Of 2X10−4 To Preserve The Model’S General-Purpose

knowledge while fine-tuning it for the domain. The

Was Employed, Assuming All Unpaired Examples In

the batch as negatives, which is particularly suited for scenarios with positive pairs only.

Precision (Amp) To Reduce Memory Usage And

enhance training speed.

Able On Hugging Face Hub, Alongside The Complete

implementation in our GitHub repository.

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.