Negation, Contrast and Contradiction in Text Processing

Sanda M. Harabagiu, Andrew Hickl, V. Finley LacatusuAAAI 2006

Abstract

Detecting contradictions across multiple text sources is a critical challenge for modern Natural Language Processing (NLP). This paper introduces a comprehensive framework that combines negation processing, automated antonym discovery through "antonymy chains," and the classification of contrastive discourse relations. By extending traditional Textual Entailment (TE) logic, the proposed system effectively identifies conflicting information, significantly improving performance in tasks such as Question Answering (QA) and Multi-Document Summarization (MDS).

Introduction

In an era of massive data, information from different sources often conflicts. While traditional NLP systems excel at identifying similarities (paraphrases) or logical flows (entailment), they frequently struggle with incompatibility. For instance, if one source reports a successful missile test while another claims it never occurred, a standard system might fail to flag the discrepancy. Contradictions are often subtle, involving complex paraphrases or semantic oppositions rather than simple "Yes/No" binaries. This research identifies three primary drivers of contradiction: negation, antonymy, and contrastive discourse structures.

System Architecture

The authors propose a dual-view architecture integrated into a Textual Entailment (TE) system. This approach treats contradiction as the "flip side" of entailment, leveraging existing frameworks to identify disagreements.

  • View 1 (Negation Removal): This pathway "reverses" the problem by identifying and removing negations to see if the underlying propositions entail one another. If a statement like "A did not happen" is transformed into "A happened" and subsequently entails another statement "A occurred," the original pair is flagged as contradictory.
  • View 2 (Feature-based Classification): This pathway extracts a wide array of features—including lexical alignment, dependency parses, and semantic/pragmatic information—and processes them through a C5.0 decision tree classifier to predict contradictions.
Figure 1: System architecture of the Textual Entailment system for contradiction detection.

Semantic Opposition and Antonymy Chains

A key innovation of this work is the development of "Antonymy Chains." Recognizing that standard databases like WordNet have limited direct antonym pairs (e.g., hot/cold), the authors utilized WordNet’s "IS-A" and "GLOSS" relations to discover deeper semantic oppositions. By traversing definitions, the system can link words like "preclude" to "be able," even if they are not direct antonyms.

Additionally, the system treats the recognition of CONTRAST discourse relations as a classification problem. By analyzing textual alignment, the system can identify disagreements even when explicit cue words like "but" or "although" are absent.

Figure 2: Illustration of alignment and antonymy chains between contrasting sentences.

Evaluation and Results

The system was evaluated using a modified version of the PASCAL RTE 2006 dataset, Wiktionary for antonym discovery, and a custom negative question set for QA applications.

Key findings include:

  • Antonymy Detection: The system correctly identified antonyms 65% of the time for words with WordNet data and achieved 33% accuracy for words with no prior antonym data using the chain method.
  • Contradiction Accuracy: The system achieved 75.63% accuracy for contradictions involving overt negation. However, accuracy dropped to 62.55% for contradictions based purely on paraphrases, highlighting the ongoing difficulty of semantic-based detection.
  • Question Answering: A hybrid strategy combining negation detection, set intersection, and entailment achieved 55.33% accuracy on negative questions, outperforming single-method approaches.

Conclusion

This research demonstrates that contradiction detection is a viable extension of textual entailment. While overt negations are relatively easy to process, "paraphrased contradictions" remain a significant hurdle. Future work will focus on scaling semantic chains and improving alignment modules to handle the diverse linguistic expressions of opposition found in real-world text collections.

Summary updated

BibTeX
@inproceedings{harabagiu2006negation,
  author = {Sanda M. Harabagiu and Andrew Hickl and V. Finley Lacatusu},
  title = {{Negation, Contrast and Contradiction in Text Processing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}

Loading related content...

Figure