Multilingual FrameNet Tutorial at COLING 2018 (Santa Fe, NM)

• Frame Semantics and FrameNet (Miriam)

  • Frames, Frame Elements, and Lexical units
  • Annotation and Reports
  • Frame Relations and Frame Element Relations
  • Construction Grammar and Constructicons

• The Multilingual FrameNet Project (Collin)

  • FrameNets in Languages other than English
  • Variations in Goals and Practices
  • Alignment across Languages
    • Alignment Algorithms

• Parallel Annotation on TED Talk (Michael)

  • Web Annotation Tool (Torrent et al. 2016)
  • Cross-lingual Framing Differences
  • Trade-offs in Representation
  • Constructions implicit in FrameNet Annotation
  • Metaphor and Metonymy in FrameNet % and their connection to MLFN's Ted Talk annotation
  • Mental Spaces: Negation and Conditionals % and their connection to MLFN's Ted Talk annotation

• Uses of Multilingual FrameNet (Swabha)

  • Automatic Semantic Role Labeling
    • Automatic Semantic Role Labeling (English)
    • Automatic Semantic Role Labeling (Other Languages)
    • Multilingual Applications

• Representing Semantics in Frames and Constructions

  • Frame-based Knowledge Representation
  • Interactions between Frames and Constructions
  • Implications for Multilingual Applications

Since 1997, the FrameNet Project at the International Computer Science Institute (http://framenet.icsi.berkeley.edu) has been building a richly detailed lexical database of contemporary English, implementing the theory of Frame Semantics, developed by the late Prof. Charles Fillmore and colleagues (Fillmore 1977, 1982}. aathe ICSI project has defined semantic frames, frame elements (roles) in these frames, and lexical units (word senses) which evoke the frames, extracted text from corpora and annotated the instances of these lexical units in the texts. The ICSI FrameNet lexical database currently contains 1,224 semantic frames, each of which has an average of 9.7 frame elements (FEs), and conprise 13,639 lexical units (LUs). There are 202,229 manually annotated instances of these lexical units, each containig annotation of the FEs that appear in the sentence[FN: Some FEs which are not instantiated in the sentence may also be annotated,cf. Ruppenhofer et al. 2010 Sec. 3.2.3].

All of this research has been done on the core vocabulary of English, but the researchers have frequently considered the obvious question: to what extent are the semantic frames created for English appropriate for analyzing other lanaguages. Fortunately, inspired by the work at ICSI, a number of related projects have been developing frame semantic lexical databases for roughly a dozen languages, which vary in size, methodology, and availability. In all cases, the new projects have taken the ICSI (English) frames as a starting point, although some have ahered more closely to the example of English. In general, these projecs have found that a large proportion of the target-language words fit comfortably in those frames.

The FrameNet team has now embarked on a Multilingual FrameNet project, developing alignments across many of these FrameNets, seeking a better understanding of cross-linguistic similarities and differences in frame structure. Alignment on the frame level is often quite easy, as many projects have kept names or ID numbers which refer to the ICSI frames. Going beyond such connections, we will also use techniques such as multilingual word vectors (Hermann & Blunsom 2014) to cluster and align lexical units across languages.