Abstract
Due to the commonality in natural language, negation focus plays a critical role in deep understanding of context. However, existing studies for negation focus identification major on supervised learning which is timeconsuming and expensive due to manual preparation of annotated corpus. To address this problem, we propose an unsupervised word-topic graph model to represent and measure the focus candidates from both lexical and topic perspectives. Moreover, we propose a document-sensitive biased PageRank algorithm to optimize the ranking scores of focus candidates. Evaluation on the *SEM 2012 shared task corpus shows that our proposed method outperforms the state of the art on negation focus identification.
Introduction
Identifying the "focus" of a negation—the specific part of a sentence being denied—is a fundamental challenge in natural language processing. For instance, in the sentence, "He didn't stop until he got to Jackson Hole," the negation focus is the temporal condition "until he got to Jackson Hole." Accurately pinpointing this focus is essential for deep contextual understanding.
Historically, this task has been approached through supervised learning, which requires extensive, manually annotated datasets. Such methods are not only resource-intensive but often overlook the broader discourse context—the "importance" of words within the document's overall theme. This research introduces an unsupervised approach that leverages discourse context to identify negation focus without the need for labeled training data.
Methodology: The Word-Topic Graph Model
The core of the proposed approach is the Word-Topic Graph Model (WTGM), which represents focus candidates through two interconnected layers:
- The Word Layer: This layer models lexical relationships. Nodes represent content words (nouns, verbs, adjectives, etc.), and edges are established based on co-occurrence frequencies and semantic similarity derived from WordNet.
- The Topical Layer: This layer captures the underlying themes of the document using Latent Dirichlet Allocation (LDA). It establishes links between words and topics based on the probability of a word belonging to a specific topic and the relevance of that topic to the document.
Document-Sensitive PageRank
To rank focus candidates, the authors developed a Document-Sensitive Biased PageRank algorithm. Unlike standard PageRank, which assumes a uniform probability for "teleporting" to any node, this version biases the random walk toward words that are highly relevant to the document's specific topic distribution. This reflects the intuition that a word is more likely to be the focus of negation if it is central to the document's primary themes. Finally, semantic roles in a sentence are ranked based on the average PageRank scores of their constituent words.
Experimental Setup
The model was evaluated using the *SEM 2012 shared task corpus, which includes 3,544 instances of negation focus annotations. To capture the necessary discourse context, the authors mapped these instances back to their original sources in the Wall Street Journal section of the Penn TreeBank, allowing the model to analyze the sentences preceding and following each negation.
Results and Discussion
The experimental results demonstrate the efficacy of integrating topical information and document-sensitive ranking. The full model (WTGM + DS) achieved an accuracy of 69.39%, significantly higher than the word-layer-only baseline (52.61%).
Previous analysis indicated that 77% of negation focuses can be identified using contextual clues. By bridging lexical and semantic gaps through the topical layer, the WTGM connects related concepts even when they do not appear in the same local window. Crucially, this unsupervised approach outperformed existing state-of-the-art supervised systems.
As shown in the comparison, the proposed model surpassed the performance of B&M (2011) and Zou et al. (2014) by a statistically significant margin ($p < 0.01$). This suggests that the inherent thematic structure of a document provides a powerful signal for identifying negation focus, often exceeding the predictive power of features used in supervised classifiers.
Conclusion
This work demonstrates that unsupervised methods can effectively identify negation focus by leveraging discourse context and topic modeling. While the model excels at capturing global context, the authors acknowledge that approximately 17% of samples remain difficult to identify without considering local syntax. Future work will focus on integrating local syntactic information, such as dependency trees, into the graph model to further refine performance.
Summary updated