K-SRL: Instance-based Learning for Semantic Role Labeling

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 599–608, Osaka, Japan, December 11-17 2016. K-SRL: Instance-based Learning for Semantic Role Labeling Alan Akbik IBM Research Almaden 650 Harry Road, San Jose CA 95120, USA akbika@us.ibm.com Yunyao Li IBM Research Almaden 650 Harry Road, San Jose CA 95120, USA yunyaoli@us.ibm.com Abstract Semantic role labeling (SRL) is the task of identifying and labeling predicate-argument structures in sentences with semantic frame and role labels. A known challenge in SRL is the large num- ber of low-frequency exceptions in training data, which are highly context-specific and difficult to generalize. To overcome this challenge, we propose the use of instance-based learning that performs no explicit generalization, but rather extrapolates predictions from the most similar in- stances in the training data. We present a variant of k-nearest neighbors (kNN) classification with composite features to identify nearest neighbors for SRL. We show that high-quality predictions can be derived from a very small number of similar instances. In a comparative evaluation we experimentally demonstrate that our instance-based learning approach significantly outperforms current state-of-the-art systems on both in-domain and out-of-domain data, reaching F1-scores of 89,28% and 79.91% respectively. 1 Introduction Semantic role labeling (SRL) is the task of annotating predicate-argument structures in sentences with shallow semantic information. One prominent labeling scheme for the English language is the Proposi- tion Bank (Palmer et al., 2005), which annotates predicates with frame labels and arguments with role labels (see Figure 1 for examples). Frame labels disambiguate the predicate meaning in the context of the sentence. Role labels roughly correspond to simple questions (who, when, how, why, with whom) with regards to the disambiguated predicate. SRL has been found useful for a wide range of applications such as information extraction (Fader et al., 2011), question answering (Shen and Lapata, 2007; Maqsud et al., 2014) and machine translation (Lo et al., 2013). Current state-of-the-art SRL approaches train classifiers with bags of features (Johansson and Nugues, 2008; Bj¨orkelund et al., 2009; Choi and Palmer, 2011) to predict semantic labels for each constituent in a sentence. These approaches typically employ classifiers such as logistic regression or SVM that learn for each feature a measure of impact on the classification decision and abstract away from local contexts in specific training examples. Local bias. We argue that such approaches are not ideal for SRL due to a strong local bias of features within specific contexts. Low-frequency examples in SRL are often not noise to be abstracted away, but rather correspond to exceptions that require explicit handling. For example, consider the task of argument labeling: Arguments that are syntactically realized as passive subjects are typically labeled A11. However, there exist numerous low-frequency exceptions to this rule. For instance, passive subjects of certain frames (such as the frame TELL.01) are most commonly labeled A2. See Figure 1 for an example. Other examples of local bias include different types of diathesis alternation which affect specific frames and argument types (Kipper et al., 2008), the syntactic …

Content truncated to 500 words