Inferring Implicit Event Locations from Context with Distributional Similarities

Jin-Woo Chung, Wonsuk Yang, Jinseon You, Jong-Chan ParkIJCAI 2017 · DOI

TL;DR

This paper tackles the challenge of automatically identifying where events happen when the location isn't explicitly stated in the text. They propose a novel system that leverages distributional similarities between event descriptions to infer these implicit locations, achieving promising results that significantly improve upon previous approaches.

What Problem Does This Paper Address?

In many real-world applications, like tracking disease outbreaks or natural disasters, knowing the location of an event is crucial. While some texts explicitly state event locations (e.g., "a bomb exploded in Boston"), many documents don't. Instead, location information is often implied or understood from context. Previous research has largely focused on extracting explicitly stated locations, missing a significant portion (nearly 40%) of inferable location information. This gap makes it difficult to get a complete picture of events, especially when dealing with large volumes of unstructured text. This paper aims to bridge this gap by developing a system that can infer these implicit event locations.

Methods

The authors propose a two-step system:

  1. Detection of Explicitly Stated Locations: This step first identifies candidate location expressions (like noun phrases and adverbial phrases) and then detects explicit event-location links using locative semantic roles (ARGM-LOC and ARGM-DIR) from the OntoNotes corpus.
  2. Detection of Implicit Locations: This is the core innovation.
  3. Intra-sentential: They identify implicit links within the same sentence using three factors:
  4. Events sharing the same first argument (e.g., if "Annan met with Arafat in Gaza" and another event involves Annan, it's likely in Gaza too).
  5. Locative prepositional phrases modifying event arguments.
  6. When the event agent itself is a large entity like a country or city (e.g., "Canada is saying goodbye...").
  7. Inter-sentential: They extend explicitly stated and implicitly identified locations to other sentences. This is achieved by:
  8. Utilizing coreference relations to link events with shared arguments across sentences.
  9. Crucially, employing distributional similarities between event descriptions. The hypothesis is that if two events are described using similar language, they are likely to occur in the same location. They use word embeddings (gensim, word2vec) trained on a large corpus to measure these similarities between event words, event-location pairs, and first arguments.

Datasets

The primary dataset used is a corpus annotated for event-location relations, previously presented in Chung et al. (2015). This corpus is based on 48 documents from CNN broadcast news articles (OntoNotes Release 5.0).

A significant contribution is the newly added annotations for containment and coreference relations among location expressions. This was necessary because the original corpus lacked information about hierarchical spatial relationships (e.g., "Serbia contains Belgrade") and coreference (e.g., "Belgrade" and "the Yugoslavian capital" referring to the same place), which are crucial for robust evaluation. They achieved good inter-annotator agreement (0.79) on these new annotations.

Key Insights and Ideas

  • Distributional Similarity as a Proxy for Spatial Proximity: The core idea that similar linguistic descriptions of events imply similar locations is powerful. This allows the system to "spread" known locations to unknown events.
  • Implicit Information is Abundant: The paper highlights that a substantial portion of event-location information is not explicitly stated, underscoring the need for inference systems.
  • Combining Intra- and Inter-sentential Inference: A multi-stage approach, first handling within-sentence clues and then propagating information across sentences, is effective.
  • The Importance of Contextual Features: Using combinations of distributional similarities (event-event, event-location, first argument) rather than just one type yields better results, suggesting that a holistic view of event descriptions is beneficial.
  • Coarse-grained Locations as Fallback: When explicit cues are missing, linking events to larger, more general locations (like countries or cities) or the nearest recognized location proves to be a reasonable strategy.

Key Results

The system achieves an F1-score of 0.58 for inferring implicit event locations. While this might seem modest, the authors note it's comparable to state-of-the-art classifiers for intra-sentential spatiotemporal relations (around 0.60 F1).

  • Intra-sentential methods show high precision but low coverage.
  • Inter-sentential methods, especially those leveraging distributional similarities, significantly improve overall performance.
  • Using maximum or average values of distributional similarities across different event element pairs (event-event, event-location, first argument) performs better than using single types.
  • The coarse-grained method significantly boosts performance in a "relaxed match" setting, suggesting its utility when precise location identification is difficult.

Limitations and Future Work

  • Error Propagation: A key limitation is that errors in earlier steps (e.g., incorrect explicit location detection) can cascade and propagate to neighboring events, leading to further mistakes.
  • Temporal Gaps: Significant temporal distances between events can hinder inference, even with high distributional similarity.
  • Lack of Temporal Features: The current system does not exploit temporal relations, which are often correlated with spatial ones.
  • Future Directions:
    • Investigating "spatial transition" among multiple events.
  • Exploiting other contextual features like event coreference, narrative schemas, and sub-event hierarchies.
  • Grounding identified spatial expressions to actual geolocations for more practical applications.

Summary updated

BibTeX
@inproceedings{chung2017inferring,
  author = {Jin-Woo Chung and Wonsuk Yang and Jinseon You and Jong-Chan Park},
  title = {{Inferring Implicit Event Locations from Context with Distributional Similarities}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/136},
}
Figure