TL;DR
This paper introduces an unsupervised approach to automatically build a lexicon of adjectives that exhibit "grading" – a semantic relationship where one adjective's meaning is defined by the intensity of another (e.g., "small" vs. "tiny"). By clustering adjectives based on their co-occurrence with adverbs, the authors lay the groundwork for a comprehensive grading lexicon, a resource currently lacking in computational linguistics.
What Problem Does This Paper Address?
While relationships like synonymy and antonymy are well-documented in linguistic resources, the phenomenon of "grading" (also known as intensification or magnification) has received less attention. Grading describes how adjectives can be understood in terms of their intensity relative to each other, like "small" and "tiny," or "warm" and "hot." This relationship is crucial for understanding nuances in language, with applications in areas like question answering and ontological representations. However, there's a significant lack of dedicated resources for identifying and cataloging adjectives that participate in grading. Existing resources like WordNet are often too broad or don't capture the specific nuances of grading. This paper aims to address this gap by developing an initial step towards building such a lexicon.
Methods
The core of the paper's methodology is an unsupervised clustering approach. The authors first construct a co-occurrence matrix using Google syntactic ngrams, focusing on adverbs that are syntactically dependent on adjectives (specifically, those in an advmod dependency relation). They experiment with several "feature sets" for this matrix:
- Manual: A curated set of adverbs including degree modifiers, "remarkably adverbs," subject-oriented adverbs, and viewpoint/time adverbs (84 features total).
- Manual Collapsed: A simplified version of the manual set, grouping adverbs by type (5 features total).
- Top 1000 Advs: The 1000 most frequent adverbs in the corpus, regardless of their dependency on adjectives.
This co-occurrence matrix is then weighted using Positive Pointwise Mutual Information (PPMI). Finally, k-means clustering (with k=2) is applied to the adjectives to split them into two groups: those presumed to be gradable and those presumed to be non-gradable. To further refine the "gradable" group and identify adjectives that might not truly have "scalemates" (adjectives on the same scale), they employ anomaly detection using the Local Outlier Factor (LOF) algorithm.
Datasets
- Google Syntactic Ngrams Corpus: Used to build the co-occurrence matrices.
- Collins COBUILD Dictionary: Used for evaluation. Adjectives were classified as "classifying" (roughly non-gradable) or "qualitative" (roughly gradable). The authors filtered their datasets to align with these classifications, handling cases with conflicting senses.
- Wikipedia List of Adjectival and Demonymic Forms for Countries and Nations: Used as a heuristic to identify and exclude demonyms (like "American," "Swedish") which are reliably non-gradable.
Key Insights and Ideas
- Grading vs. Gradability: The paper highlights an important distinction: "gradability" (ability to be modified by "very" or used in comparatives) is not the same as participating in the "grading" semantic relation. For example, "lukewarm" might not be considered gradable but still participates in grading, while "spotted" is gradable but doesn't express grading with other words.
- Unsupervised Approach is Viable: Given the scarcity of labeled data for grading, an unsupervised clustering approach is presented as a practical and promising first step.
- Adverb Co-occurrence as a Signal: The frequency and type of adverbs an adjective co-occurs with serve as a strong indicator of its participation in grading.
- Manual Feature Engineering Not Always Superior: Surprisingly, the more manually curated feature sets did not consistently outperform the simpler "Top 1000 Advs" set, suggesting that broader co-occurrence patterns can be highly effective.
- Anomaly Detection for Refining Scales: LOF proved useful in identifying potential outliers within the gradable group, though its effectiveness for definitively identifying adjectives without scalemates is questioned.
Key Results
The clustering step achieved F₁ scores between .78 and .82 on the COBUILD dataset. While these scores are lower than a previous supervised approach by Hatzivassiloglou and Wiebe (2000), the authors emphasize their higher recall, suggesting their broader definition of gradable adjectives captures more instances. The anomaly detection step identified words like "able" and "logical" as potential outliers, though their classification by COBUILD as "classifying" (non-gradable) suggests a broader interpretation of "grading" in their method. Demonyms were correctly excluded with 100% accuracy.
Limitations and Future Work
- Manual Feature Selection: The authors acknowledge that their current feature selection process still involves significant manual effort, which they aim to reduce.
- Evaluation Challenges: The reliance on the COBUILD dictionary for evaluation has limitations, as it uses a narrower definition of gradability and may misclassify certain words (e.g., "far"). There's a lack of a gold standard dataset for grading.
- LOF Effectiveness: The effectiveness of LOF for identifying adjectives without scalemates is questioned, suggesting further exploration of anomaly detection or other methods is needed.
- Broader Feature Sets: Future work could explore more semantically related feature sets and potentially biclustering techniques to group adjectives and adverbs simultaneously.
- Cross-Lingual Applications: The authors note that their approach, particularly the lack of reliance on manual feature engineering, holds promise for applications in other languages.
- Lexicon Expansion: The ultimate goal is to expand this initial step into a comprehensive lexicon of grading adjectives.
Summary updated