Entity/Event-Level Sentiment Detection and Inference

Lingjia DengNAACL 2015

TL;DR

This paper proposes a novel system for detecting both explicit and implicit sentiments expressed towards entities and events in text. By incorporating inference rules into joint prediction models, the system aims to go beyond simply identifying stated opinions to inferring underlying sentiments, particularly those related to events that positively or negatively affect entities.

What Problem Does This Paper Address?

Current sentiment analysis and opinion mining research largely focuses on extracting explicit sentiments. However, opinions are often expressed implicitly, requiring inference to understand the full sentiment landscape. This paper addresses the challenge of detecting and inferring these implicit sentiments, especially when they are tied to specific entities or events. The authors highlight that while explicit sentiments are straightforward, implicit sentiments require deeper analysis, often involving understanding how events impact entities (e.g., a "defeat" event negatively impacts a "bill"). This capability is crucial for applications like question answering systems that need to understand nuanced opinions.

Methods

The proposed approach involves a multi-stage process:

  • Explicit Sentiment Extraction: The system leverages existing tools to extract opinion and source spans. A key contribution here is a proposed classification model to identify "eTargets" (entities or events that are the direct target of an opinion). This model considers features related to opinion spans and target spans, addressing cases where target spans might be structurally absent.
  • Implicit Sentiment Inference: Explicit sentiments serve as input for inferring implicit sentiments. The core of this inference is a joint prediction model that combines probabilistic calculations of various ambiguities (e.g., sentiment source, target, polarity) under the constraints of inference rules.
  • Inference Rules: The paper builds upon previously developed rules for inferring sentiments related to "+/-effect events" (events that have a positive or negative impact on a theme). These rules are formalized using first-order logic.
  • Joint Prediction Framework: To handle the complex interdependencies between ambiguities, the authors propose using Markov Logic Networks (MLN) or Probabilistic Soft Logic (PSL). MLN allows for weighted first-order logic formulas to represent probabilistic graphical models, while PSL can handle numerical values for atoms, offering more flexibility.

Datasets

  • MPQA 3.0: The paper proposes creating a new corpus, MPQA 3.0, by extending the existing MPQA 2.0 corpus. MPQA 3.0 will include "entity-target" and "event-target" (eTarget) annotations, which are crucial for the proposed entity/event-level sentiment analysis.
  • KBP Sentiment Track: The NIST Knowledge Base Population (KBP) Sentiment track is mentioned as a relevant benchmark and a source of corpora for collecting information regarding sentiments expressed toward named entities.

Key Insights and Ideas

  • Implicit Sentiment is Key: The paper strongly emphasizes that understanding implicit sentiments is crucial for a comprehensive sentiment analysis system, moving beyond surface-level explicit opinions.
  • "+/-Effect Events" as a Bridge: The concept of "+/-effect events" is a central idea, providing a structured way to link events to their impact on entities and infer sentiments.
  • Joint Prediction for Ambiguity Resolution: Recognizing that various sentiment-related ambiguities are interdependent, the authors advocate for a joint prediction framework (MLN/PSL) to simultaneously resolve these ambiguities.
  • Handling Structurally Absent Features: The proposed classification model for eTargets addresses a practical challenge where features might not be present for all samples, a common issue when integrating diverse resources.

Key Results

  • Feasibility of Annotation: An agreement study for the eTarget annotation task showed an average F-measure of 0.82, indicating the feasibility of creating the proposed MPQA 3.0 corpus.
  • Pilot Study Success: A pilot study using a simplified version of the inference rules and an Integer Linear Programming (ILP) framework demonstrated the potential of the joint inference approach.

Limitations and Future Work

  • Scope of Proposal: The current proposal focuses on sentiments expressed by the writer and sentiments towards +/-effect events. Sentiments expressed by other entities and beliefs/arguments are noted as out of scope for this specific proposal but are identified as future exploration.
  • Future Corpus Annotation: The authors plan to continue annotating the MPQA corpus to further develop MPQA 3.0.
  • Framework Extension: The proposed framework is designed to be extensible to jointly analyze sentiments and beliefs in the future.
  • Choice of Framework: While MLN is proposed for its ability to directly handle first-order logic rules, the paper also discusses the limitations of ILP and the potential of PSL for handling numerical atom values. Exploration of other graphical models like LDA is also mentioned.

Summary updated

BibTeX
@inproceedings{deng2015entity,
  author = {Lingjia Deng},
  title = {{Entity/Event-Level Sentiment Detection and Inference}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
}
Figure