Generating Training Data for Semantic Role Labeling based on Label Transfer from Linked Lexical Resources

Generating Training Data for Semantic Role Labeling based on Label Transfer from Linked Lexical Resources Silvana Hartmann⋄, Judith Eckle-Kohler⋄, Iryna Gurevych⋄‡ ⋄ UKP Lab, Technische Universit¨at Darmstadt ‡ UKP Lab, German Institute for Educational Research http://www.ukp.tu-darmstadt.de Abstract We present a new approach for generating role-labeled training data using Linked Lexi- cal Resources, i.e., integrated lexical resources that combine several resources (e.g., Word- Net, FrameNet, Wiktionary) by linking them on the sense or on the role level. Unlike resource-based supervision in relation extrac- tion, we focus on complex linguistic anno- tations, more specifically FrameNet senses and roles. The automatically labeled train- ing data (www.ukp.tu-darmstadt.de/ knowledge-based-srl/) are evaluated on four corpora from different domains for the tasks of word sense disambiguation and seman- tic role classification. Results show that classi- fiers trained on our generated data equal those resulting from a standard supervised setting. 1 Introduction In this work, we present a novel approach to automati- cally generate training data for semantic role labeling (SRL). It follows the distant supervision paradigm and performs knowledge-based label transfer from rich external knowledge sources to large corpora. SRL has been shown to improve many NLP appli- cations that rely on a deeper understanding of seman- tics, such as question answering, machine translation or recent work on classifying stance and reason in online debates (Hasan and Ng, 2014) and reading comprehension (Berant et al., 2014). Even though unsupervised approaches continue to gain popularity, SRL is typically still solved using supervised training on labeled data. Creating such labeled data requires manual annotations by experts,1 1Even though crowdsourcing has been used, it is still prob- resulting in corpora of highly limited size. This is especially true for the task of FrameNet SRL where the amount of annotated data available is small. FrameNet SRL annotates fine-grained semantic roles in accordance with the theory of Frame Seman- tics (Fillmore, 1982) as illustrated by the following example showing an instance of the Feeling frame including two semantic roles: HeExperiencer feltFeeling no sense of guiltEmotion in the betrayal of personal confidence. Our novel approach to training data generation for FrameNet SRL uses the paradigm of distant supervi- sion (Mintz et al., 2009) which has become popular in relation extraction. In distant supervision, the over- all goal is to align text and a knowledge base, using some notion of similarity. Such an alignment allows us to transfer information from the knowledge base to the text, and this information can serve as labeling for supervised learning. Hence, unlike semi-supervised methods which typically employ a supervised clas- sifier and a small number of seed instances to do bootstrap learning (Yarowsky, 1995), distant supervi- sion creates training data in a single run. A particular type of knowledge base relevant for distant supervi- sion are linked lexical resources (LLRs): integrated lexical resources that combine several resources (e.g., WordNet, FrameNet, Wiktionary) by linking them on the sense or on the role level. Previous approaches to generating training data for SRL (F¨urstenau and Lapata, 2012) do not use lex- …

Content truncated to 500 words