Semantic Analysis of Massive Text under Multi-Model Strategy

Ze‐Kun Tao, Youwei Zhang, Feiyue Fang, Jing Li, Chuanwei Lu, Hongjian WuProceedings of the International Conference on Information Economy, Data Modeling and Cloud Computing, ICIDC 2022, 17-19 June 2022, Qingdao, China 2022 · DOI

Abstract

The comment text generated by tourists' travel is one of the core contents of the research on semantic analysis of tourist destinations. Considering the phenomenon of fake reviews, simple copying, worthless information and irrelevant content, it prevents tourists from obtaining valuable information. This paper proposes a multi-model fusion strategy to extract meaningful insights from massive, noisy tourist review datasets. By combining density clustering, semantic word embeddings, and probabilistic language models, the authors filter out "worthless" text to build structured knowledge graphs and word clouds that characterize hotel destinations.

Introduction

The explosion of online travel platforms has led to a massive influx of user-generated reviews. However, this data is often "noisy"—riddled with fake reviews, simple copy-pasted text, and irrelevant information. For tourists, this makes it difficult to find authentic feedback; for hotel managers, it obscures actionable insights. The challenge lies in accurately deduplicating content that isn't just an exact copy but is semantically similar, while also identifying which comments actually carry "value" for analysis.

Methodology: The Multi-Model Strategy

The researchers employ a "Multi-Model Strategy" that functions in two main phases: data cleaning (fusion) and feature extraction.

The overall workflow of the multi-model strategy, from pre-processing to visualization.

Data Cleaning and Fusion

To ensure only high-quality data remains, the authors use a voting algorithm across four distinct models:

  • DBSCAN (Density-Based Clustering): Groups similar texts based on density. It effectively handles non-convex data sets, keeping the longest review in a cluster and discarding the rest.
  • Word2Vec: Converts reviews into high-dimensional vectors. By calculating the distance between these vectors, the system identifies and removes semantically redundant sentences.
  • Simhash: A fingerprinting technique that converts text into a bitstring. It uses Hamming distance to quickly identify near-duplicate documents efficiently for large datasets.
  • N-Gram Language Model: Calculates the "Perplexity" (PPL) of sentences. The authors discard the bottom 25% of data (highest PPL) to filter out gibberish or bot-generated text.

Feature Extraction and Visualization

Once the data is cleaned, the authors apply TF-IDF and Latent Dirichlet Allocation (LDA) to extract keywords and latent topics such as service, location, and facilities. Sentiment analysis is performed using the BosonNLP library. The final output is a Knowledge Graph (Neo4j) consisting of "triples" (Subject-Keyword-Hotel) that visualizes the relationships between specific hotel characteristics and room types.

Dataset Description

The study utilizes a dataset from the 9th National "Teddy Cup" Data Mining Challenge (Question C), comprising 25,000 comment entries across 50 different hotels.

Sample of the raw hotel review dataset including name, time, content, and room type.

The attributes include hotel name, timestamp, review content, and room type (e.g., standard room, superior room). This structured format allows for granular analysis of how specific room types are perceived by guests.

Key Insights and Results

The core innovation is the ensemble deduplication approach. While Simhash handles near-duplicates, Word2Vec addresses semantic similarity, and N-Gram perplexity ensures linguistic quality. Combining these via a voting strategy creates a robust filter that successfully reduced noise in the 25,000-comment dataset.

The system successfully generated a Neo4j knowledge graph, transforming messy reviews into a navigable map of hotel attributes. For instance, a Word Cloud for "Hotel A01" identified top-tier concerns like "services," "front desk," "traffic," and "breakfast," providing clear actionable feedback for management.

Conclusion and Future Work

While effective, the current model relies on a single data source. Future research will focus on multimodal fusion, incorporating images and videos from reviews to create a more comprehensive profile for tourist destinations. Additionally, refining the voting algorithm's weights will improve handling of diverse "worthless" text styles across various social media platforms.

Summary updated

Abstract

The comment text generated by tourists' travel is one of the core contents of the research on semantic analysis of tourist destinations. Considering the phenomenon of fake reviews, simple copying, worthless information and irrelevant content, it prevents tourists from obtaining valuable information
BibTeX
@inproceedings{tao2022semantic,
  author = {Ze‐Kun Tao and Youwei Zhang and Feiyue Fang and Jing Li and Chuanwei Lu and Hongjian Wu},
  title = {{Semantic Analysis of Massive Text under Multi-Model Strategy}},
  booktitle = {Proceedings of the International Conference on Information Economy, Data Modeling and Cloud Computing, ICIDC 2022, 17-19 June 2022, Qingdao, China},
  year = {2022},
  doi = {10.4108/eai.17-6-2022.2322807},
}
Figure