TL;DR
This paper introduces TweeNLP, a novel platform that curates and visualizes Twitter discussions related to Natural Language Processing (NLP). By integrating with the NLPExplorer scientific literature search engine, TweeNLP bridges the gap between informal online conversations and formal research, offering a unique way to explore the NLP community's pulse.
What Problem Does This Paper Address?
The NLP community, like many others, increasingly uses online platforms for communication. While Twitter offers broad reach and accessibility for announcements, discussions, and informal research exchanges, it lacks the structured organization and long-term memory of dedicated forums. This makes it difficult to track trends, connect discussions to specific research papers, and gain a comprehensive overview of the community's evolving interests. Existing NLP research portals focus on published literature but miss the rich, real-time discourse happening on social media.
Methods
TweeNLP employs an automatic data curation pipeline to gather and process tweets. Key components include:
- Tweet Classifier: A BERT-base classifier fine-tuned to categorize tweets into six topics: New Paper Announcements, Call for Papers, Reading Materials & Tutorials, Career Opportunities, Talks & Seminars, and Others. This classification is crucial for organizing the vast amount of Twitter data.
- Conference Page Builder: Identifies tweets related to ongoing conferences to build dedicated conference pages.
- CFP Timeline Builder: Extracts submission deadlines from "Call for Papers" tweets to create a chronological timeline of upcoming events.
- Paper Tweet Linker: Matches tweets to research papers by analyzing paper titles and URLs, enabling the embedding of relevant tweets on research paper pages within NLPExplorer.
The system integrates with NLPExplorer, a portal that visualizes NLP research volume based on the ACL Anthology.
Datasets
- Twitter Data: Curated using the Twint library, focusing on tweets with the hashtag #NLProc and conference-specific hashtags (e.g., #acl2020, #emnlp2020). As of April 2021, the dataset contained 19,395 tweets collected since October 2017.
- ACL Anthology: Used to obtain conference and journal names, paper titles, and URLs for linking tweets to research papers.
Key Insights and Ideas
- Bridging the Gap: TweeNLP effectively connects the informal, real-time discussions on Twitter with the structured, formal world of NLP research papers, creating a more holistic view of the field.
- Community Memory: The platform acts as a collective memory for the NLP community, preserving and organizing discussions that might otherwise be lost in the fast-paced Twitter feed.
- Beyond Formal Metrics: By analyzing Twitter activity, TweeNLP offers insights into community sentiment, emerging trends, and the informal impact of research, complementing traditional citation-based metrics.
- Discoverability: The Tweet Explorer and Conference Visualizer features enhance the discoverability of relevant NLP content, from new papers and job opportunities to conference schedules and popular discussions.
Key Results
- The Tweet Classifier achieved a test accuracy of 75% on a manually annotated dataset.
- The system curated 19,395 NLP tweets as of April 2021, spanning over 1300 days.
- TweeNLP organizes tweets into six distinct categories, with "New Paper Announcements" and "Others" being the most prevalent.
- The platform provides visualizations of conference activity, popular papers, and a timeline of submission deadlines.
Limitations and Future Work
- Scope: Currently, TweeNLP is limited to NLP papers within the ACL Anthology.
- Bias: The current system prioritizes popular tweets based on likes and retweets, which might bias the representation of ideas by overemphasizing influential users.
- Future Work: The authors suggest extending the system to include papers from other major conferences (NeurIPS, ICLR, CVPR) and other research domains. They also aim to address the bias in popularity metrics by developing methods to create a more inclusive representation of community perspectives.
Summary updated