TL;DR
This paper introduces a novel task: discovering "emerging entities" (like new products or people) in microblogs as they first appear, before they become widely known or registered in knowledge bases. They propose a "time-sensitive distant supervision" method that leverages early-stage microblog posts to automatically create training data, achieving high precision and detecting entities significantly earlier than their Wikipedia registration.
What Problem Does This Paper Address?
Keeping up with the constant influx of new information is crucial for many applications, from social trend analysis to market research. This includes identifying "emerging entities" – things like new products, people, or events that are just starting to gain traction. Existing approaches often rely on knowledge bases (KBs) like Wikipedia. However, there's a significant delay before new entities are added to these KBs, and only the most notable ones make the cut. This means many emerging entities, especially "long-tail" ones (less frequent but diverse), are missed. Previous methods that try to detect "out-of-KB" entities often end up with a lot of noise (non-emerging entities) and struggle with entities that share names with existing ones (homographs). Furthermore, creating labeled datasets for these tasks is labor-intensive and dependent on the specific KB used.
Methods
The core of the proposed method is a novel "time-sensitive distant supervision" approach. Instead of manual annotation, it leverages the inherent temporal nature of microblogs and existing KBs to automatically generate training data.
Here's how it works:
- Candidate Emerging Entity Collection: They start with entities registered in Wikipedia and identify those that appeared relatively recently. To avoid immediate homographs, they filter out entities that already have a high volume of mentions in the microblog data within their early period.
- Emerging Context Collection: For these candidate entities, they collect microblog posts from the period just before their Wikipedia registration. These posts are hypothesized to contain "emerging contexts" – language used when the writer assumes the reader isn't yet familiar with the entity.
- Prevalent Context Collection (Negative Examples): To prevent the model from overfitting and to help it distinguish between emerging and established entities, they also collect microblog posts from a later period (e.g., one year after registration) for the same entities. These serve as "prevalent contexts."
- Sequence Labeling: The collected emerging and prevalent contexts are used to train sequence labeling models (like CRF and LSTM-CRF). These models learn to identify spans of text that represent emerging entities. The BIOES tagging scheme is used for labeling.
Datasets
- Training Data: The authors created a large-scale dataset from their own Twitter archive, which has been compiled since March 2011.
- Entities: Wikipedia articles registered between March 2012 and December 2015 (Japanese version).
- Contexts: Tweets collected before the Wikipedia registration date (emerging contexts) and tweets collected one year later (prevalent contexts).
- This resulted in 19,604 entities with 222,092 tweets (emerging and prevalent contexts combined).
- Evaluation Data: Daily Japanese retweets from three distinct periods in 2017 (March 21st, June 14th, and September 1st) were used as test sets.
- Reference List for Recall: A list of 13,386 emerging entities was constructed by identifying entities that appeared frequently on Twitter and were later registered in Wikipedia.
Key Insights and Ideas
- "Emerging Contexts" are Key: The fundamental idea is that people use specific language when introducing something new, signaling its emergence. This linguistic cue is more reliable than simply checking if an entity is in a KB.
- Time-Sensitive Supervision is Powerful: Automatically generating training data by exploiting the temporal relationship between microblog mentions and KB registration is a significant advancement, reducing reliance on manual annotation.
- Microblogs are Rich Sources: Microblogs are ideal for this task because they are timely and capture a wider variety of emerging entities, including long-tail and homographic ones, earlier than traditional news or KBs.
- Homographs are Manageable: The method can effectively distinguish emerging entities even when they share names with existing ones, a common challenge for previous approaches.
- Early Detection is Possible: The ability to detect entities significantly before they are registered in Wikipedia highlights the value of this approach for real-time applications.
Key Results
- High Precision: The proposed LSTM-CRF model achieved an average precision of 83.2% for the top 500 discovered emerging entities across the test sets. This significantly outperformed baselines (under 20%).
- High Recall: The method detected 80.4% of emerging entities that were eventually registered in Wikipedia.
- Early Detection: A remarkable 92.8% of detected entities were identified earlier than their Wikipedia registration date, with an average lead time of 578 days.
- Discovery of Long-Tail and Homographic Entities: The method successfully identified entities that are less frequent and those with ambiguous names, which are often missed by other methods.
- LSTM-CRF Superiority: The LSTM-CRF model generally outperformed the CRF model, likely due to its ability to model longer contextual dependencies.
Limitations and Future Work
- Limitations:
- The method struggles with entities that do not exhibit clear "emerging contexts" in microblogs.
- The presence of prevalent entities in the reference list used for evaluation could slightly affect performance metrics.
- The current method does not explicitly handle entities that are already well-known but might be experiencing a resurgence or new phase of popularity.
- Future Work:
- The authors plan to extend their work to perform semantic typing of emerging entities, leveraging the rich contextual information captured.
- Exploring applicability to other microblogging platforms like Weibo.
Summary updated