Excitatory or Inhibitory: A New Semantic Orientation Extracts Contradiction and Causality from the Web

Chikara Hashimoto, Kentaro Torisawa, Stijn De Saeger, Jong-Hoon Oh, Jun’ichi KazamaEMNLP 2012

TL;DR

This paper introduces "Excitation," a novel semantic orientation that classifies predicates as excitatory, inhibitory, or neutral. By automatically acquiring this "Excitation" knowledge from the web, the authors demonstrate its effectiveness in extracting millions of contradiction and causality pairs, and subsequently generating plausible, novel causal hypotheses.

What Problem Does This Paper Address?

Understanding the relationships between events described in text is fundamental for many Natural Language Processing (NLP) tasks, such as question answering. For instance, to answer "What ruined the crops in Japan?", a system needs to recognize that "Fukushima nuclear power plant caused radioactive pollution" implies a causal link and that "contaminate crops" contradicts "preserve crops." While existing work has explored semantic orientations like good/bad or positive/negative, there's a need for a more nuanced understanding of how events activate or suppress each other. This paper aims to fill this gap by proposing and acquiring a new semantic orientation, "Excitation," to better extract contradictions and causal relationships from vast amounts of web text.

Methods

The core of this paper is the introduction and acquisition of the "Excitation" semantic orientation.

  • Excitation: Predicates are classified into three categories:
  • Excitatory: The referent of the argument is activated or enhanced (e.g., cause X, preserve X).
  • Inhibitory: The referent of the argument is deactivated or suppressed (e.g., ruin X, contaminate X, prevent X).
  • Neutral: Neither excitatory nor inhibitory (e.g., consider X, related to X).
  • Excitation Template Acquisition: The authors propose a bootstrapping method to automatically acquire Excitation templates from the web. This method leverages "natural, language independent constraints on narrative structures." Specifically, it analyzes co-occurrences of templates in sentences, considering:
  • Connectives: Distinguishing between AND/THUS-type (suggesting consistency) and BUT-type (suggesting inconsistency).
  • Noun Pairs: Classifying noun pairs as "positively-associated" (PNPs) where one noun facilitates the other's emergence, or "negatively-associated" (NNPs) where one suppresses the other.
  • Constraint Matrix: A matrix that defines expected relationships between template polarities and noun pair types (e.g., excitatory templates connected by AND/THUS-type connectives with PNPs).
  • Bootstrapping: The process starts with a small set of manually identified seed templates and iteratively expands the template network, calculating Excitation values using a "spin model" approach (similar to physics) until convergence.
  • Contradiction Extraction: Contradiction pairs are extracted by identifying templates with opposite Excitation polarities that share the same argument noun and have the same part-of-speech for their predicates. These pairs are then ranked using a score that combines the absolute Excitation values of the templates and their distributional similarity.
  • Causality Extraction: Causality pairs are extracted from sentences using non-causal connectives (like "and" but not "thus") where two templates with strong Excitation tendencies co-occur. The assumption is that if one event is strongly activated or suppressed, it can have a causal effect on another. These pairs are ranked by the product of the absolute Excitation values of their templates.
  • Causality Hypothesis Generation: This method generates novel causal hypotheses by taking an extracted causality pair and replacing its phrases with their contradiction counterparts. For example, if "increase in crime ⇒ heighten anxiety" is a causality, and "decrease in crime | increase in crime" is a contradiction, then "decrease in crime ⇒ diminish anxiety" can be generated as a hypothesis. These hypotheses are then ranked.

Datasets

  • Corpus: 600 million Japanese web pages, parsed by KNP (Kawahara and Kurohashi, 2006).
  • Seed Templates: 46 manually prepared seed templates (36 excitatory, 10 inhibitory).
  • Evaluation Data: Manually annotated samples of templates and argument nouns, labeled as excitatory, inhibitory, neutral, or undecided.

Key Insights and Ideas

  • "Excitation" as a Fundamental Semantic Property: The paper posits that "Excitation" (activation/suppression) is a core semantic orientation, distinct from sentiment (good/bad) and activity (active/passive), that is crucial for understanding event relationships.
  • Web-Scale Knowledge Acquisition: The proposed method demonstrates that robust semantic knowledge, like contradiction and causality, can be automatically acquired from massive, uncurated web data by leveraging linguistic constraints.
  • Contradiction as a Bridge to Causality: The novel approach of generating causality hypotheses by inverting existing causal relationships using contradiction pairs highlights an interesting interplay between these semantic relations.
  • Universality of Constraints: The authors believe the discourse constraints used for Excitation acquisition are universal, suggesting the method's potential applicability across languages.

Key Results

  • Excitation Template Acquisition: The method successfully acquired a large number of Excitation templates, outperforming baselines like ALLEXC and SIM in precision for both excitatory and inhibitory templates.
  • Contradiction Extraction: Over one million contradiction pairs were extracted with an estimated 70% precision. The method significantly outperformed a random baseline and a baseline that didn't use Excitation for ranking.
  • Causality Extraction: Approximately 500,000 causality pairs were extracted with about 70% precision. The proposed method (PROPcaus) significantly outperformed a frequency-based baseline (FREQ).
  • Causality Hypothesis Generation: One million plausible causality hypotheses were generated with 57% precision.

Limitations and Future Work

  • Neutral Templates: While the method classifies templates as excitatory or inhibitory, a significant portion of templates are likely neutral, and their handling could be further refined.
  • Negation Inversion: The paper acknowledges that the inversion of Excitation for negated predicates (e.g., "don't preserve X" being inhibitory) might seem counter-intuitive and plans to address this in future work.
  • Ambiguity in Template Usage: The authors note that some templates can be excitatory in one context and inhibitory in another (e.g., "improve X" depending on whether X is a function or a disorder), and their current method struggles with this nuance.
  • Future Work: The authors plan to make their acquired knowledge resources publicly available and investigate further applications of Excitation, particularly in improving question answering systems.

Summary updated

BibTeX
@inproceedings{hashimoto2012excitatory,
  author = {Chikara Hashimoto and Kentaro Torisawa and Stijn De Saeger and Jong-Hoon Oh and Jun’ichi Kazama},
  title = {{Excitatory or Inhibitory: A New Semantic Orientation Extracts Contradiction and Causality from the Web}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2012},
}
Figure