An Advanced Press Review System Combining Deep News Analysis and Machine Learning Algorithms

Danuta Ploch, Andreas Lommatzsch, Florian SchultzeACL 2016

TL;DR

This paper introduces an extensible press review system that leverages deep news analysis and machine learning to automate the complex task of media monitoring for a university's press department. By integrating various text mining techniques, the system aims to provide a comprehensive and insightful overview of media coverage, going beyond simple keyword matching to understand nuances like named entities, sentiment, and event clustering.

What Problem Does This Paper Address?

In today's media-saturated world, understanding how companies and institutions are perceived is crucial. Manually creating press reviews, which involves sifting through vast amounts of news and social media content, is a time-consuming and challenging task. This is particularly true for organizations like universities, which generate a lot of public interest across various media channels. The sheer volume of information, coupled with the need to identify specific entities (like professors or faculties), track events, and gauge sentiment, makes automated solutions highly desirable. Existing commercial systems often lack transparency in their algorithms and rely heavily on human expertise, while research systems may focus on specific aspects rather than a holistic approach. This paper addresses the need for an integrated, extensible system that can handle the complexities of modern media analysis.

Methods

The core of the system is an open framework designed to integrate diverse information sources and machine learning algorithms. It employs a flexible processing pipeline that enriches crawled documents with various text mining techniques. Key components include:

  • Crawlers: These gather potentially relevant documents and tweets from news portals, search engines, RSS feeds, and Twitter.
  • Validation: A rule-based classifier, trained on manually labeled data, filters out irrelevant documents, with a particular focus on handling abbreviations common in academic settings.
  • Deduplication: The Rabin fingerprint algorithm is used to efficiently identify near-duplicate documents, crucial for handling content that appears across multiple sources.
  • Named Entity Detection (NED): Leveraging Stanford CoreNLP, this component identifies and disambiguates persons, professors, and faculties mentioned in articles. It builds a comprehensive representation of individuals by resolving co-references both within and across documents.
  • Assignment of Faculties: By creating a register of faculty affiliations based on university websites, news articles can be assigned to specific faculties.
  • Event Detection: A two-stage clustering approach using Canopy and k-means algorithms groups news articles related to a single event, improving the accuracy of event identification.
  • Sentiment Analysis: The system incorporates two sentiment analysis components: a lexicon-based approach using the SentiWS dictionary and a machine learning approach employing a Multi-nominal Naive Bayes classifier trained on annotated sentences.
  • Visualization: A web-based user interface provides two main views: a "Newsletter" or live view for recent news and a "Newsarchive" view for historical data, statistics, and event visualization.

Datasets

The paper mentions the use of a manually labeled dataset for training the validation component's rule-based classifier. For sentiment analysis, a training dataset of approximately 2,400 randomly selected sentences from crawled documents was annotated with positive, negative, or neutral sentiment. No new benchmarks were introduced; rather, existing resources like the SentiWS sentiment dictionary and Stanford CoreNLP were utilized.

Key Insights and Ideas

  • Handling Abbreviations is Crucial: The paper highlights the significant challenge of dealing with common abbreviations (like "TUB" for "Technische Universität Berlin") in news articles and emphasizes the need for robust filtering mechanisms.
  • Integrated Approach for Deeper Understanding: By combining multiple text mining techniques (deduplication, NED, sentiment analysis, event detection), the system moves beyond simple keyword matching to provide a richer, more nuanced understanding of media coverage.
  • Person-Centric Faculty Assignment: Assigning news to faculties based on identified individuals is an innovative way to measure media response at a granular level within an institution.
  • Two-Stage Event Clustering: The use of Canopy for initial clustering followed by k-means for refinement is a clever strategy to handle event detection in news streams.
  • Dual Sentiment Analysis: Employing both lexicon-based and machine learning approaches for sentiment analysis provides a more comprehensive and potentially more accurate assessment, especially given the objective nature of news reporting.

Key Results

The paper focuses on demonstrating the system's functionality and its ability to provide comprehensive media analysis. While specific quantitative results like accuracy scores for individual components are not detailed in this system demonstration paper, the authors showcase the system's capabilities through its web interface. They illustrate how the system can:

  • Display live news streams and historical archives.
  • Visualize aggregated sentiment scores for universities.
  • Group news articles by identified events.
  • Show the distribution of news related to different university faculties.
  • Provide sentiment scores for individual sentences within documents.

The system successfully integrates various advanced text mining and machine learning algorithms to create a functional press review portal tailored for a university's press department.

Limitations and Future Work

The authors acknowledge several limitations and outline promising future directions:

  • User Studies: They plan to conduct comprehensive user studies to optimize algorithms based on user feedback and needs.
  • Expanding Data Sources: Continuous work is planned to incorporate more data sources, such as blogs and additional RSS feeds.
  • Multilingual Support: The current system primarily focuses on German texts, and improved support for documents in other languages is a key future goal.
  • Advanced Summarization and Tracking: Future work includes integrating algorithms for summarizing multiple documents related to events and for tracking the evolution of topics and sentiments over longer time frames.
  • Deep Neural Networks: The authors aim to create an extended entity dataset and train a deep neural network for improved person identification.

Summary updated

BibTeX
@inproceedings{ploch2016advanced,
  author = {Danuta Ploch and Andreas Lommatzsch and Florian Schultze},
  title = {{An Advanced Press Review System Combining Deep News Analysis and Machine Learning Algorithms}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2016},
}
Figure