Intelligent Investment Research and Financial Sentiment Monitoring with NLP and Big Data

The financial asset‑management industry faces massive unstructured data growth, making information asymmetry a key competitive factor. To address this, the authors present an end‑to‑end intelligent investment‑research platform that leverages natural‑language‑processing (NLP), big‑data pipelines, and deep‑learning models to transform raw textual sources into structured insights for investment decisions.

The system architecture consists of three layers: an application layer exposing over 30 NLP services, a component layer providing core algorithms such as tokenization, named‑entity recognition, and dependency parsing, and a corpus layer that supplies both generic and domain‑specific training data. This modular design enables rapid development of downstream applications.

Two flagship applications are detailed: (1) a large‑scale bid‑tender text analysis system that automatically crawls, extracts, and structures millions of procurement documents, achieving 98% title extraction accuracy and 96% body extraction accuracy; (2) a financial sentiment‑monitoring system for WeChat groups that classifies messages into eleven categories, aggregates information without loss, and supports multi‑dimensional hotspot analysis.

Key model innovations include a tag‑embedding technique that learns distributed representations of HTML tags, an improved Transformer‑based named‑entity recognizer with bigram embeddings and relative positional attention, and a lightweight three‑layer CNN classifier enhanced by position embeddings. For few‑shot scenarios, the authors employ a two‑stage pipeline: a generic NER model followed by a domain‑specific CNN that together achieve high recall (0.97) and precision (0.96) with minimal labeled data.

Extensive experiments demonstrate that text‑augmentation methods (back‑translation, EDA, non‑core‑word replacement) consistently improve F1 scores by 5–9 points, especially when training data are scarce. A three‑stage training strategy—pre‑training word embeddings on billions of tokens, iteratively expanding the training set with high‑confidence predictions, and fine‑tuning on the target data—further boosts performance, yielding up to 48 percentage‑point gains in low‑resource settings.

The paper concludes with reflections on the importance of close collaboration between technical and domain experts, the practicality of lightweight models over heavyweight Transformers for most financial NLP tasks, and future directions such as GPT‑based augmentation and cross‑modal learning.