Jonathan Dunn Jon Beltran de Heredia Maura Burke Lisa Gandy Sergey Kanareykin Oren Kapah Matthew Taylor Dell Hines Ophir Frieder David Grossman Newton Howard Moshe Koppel Scott Morris Andrew Ortony Shlomo Argamon

Language Independent Ensemble Approaches To Metaphor Identification Funded by IARPA via Department of Defense US Army Research Laboratory contract number W911NF-12-C-0021

Metaphor Identification Metaphor is a cognitive process which may serve a key role in structuring conceptual representation and inference.

Illinois Institute of Technology Behavioral Media Networks Illinois Institute of Technology Central Michigan University Brain Sciences Foundation Bar Ilan University Behavioral Media Networks Behavioral Media Networks Georgetown University Georgetown University Massachusetts Institute of Technology Bar Ilan University Illinois Institute of Technology Northwestern University Illinois Institute of Technology

Language Independence The ideal system is language independent so that it can be easily and directly applied to new languages.

Metaphor often surfaces in language. Linguistic Processing

Input Text

Thus, when dealing with natural language text as data, it can be important to know if a piece of language is Literal or Metaphoric.

Lexical Information

(1) Because Literal and Metaphoric language represent different types of knowledge: Literal:

Metaphoric

“Falling into a hole”

“Falling into poverty”

“Demolishing his building”

“Demolishing his theory”

“Flying through restricted air space”

“Flying through his homework”

“Driving me home”

“Driving me insane”

Knowledge about the world

Knowledge about the conceptual system

Candidate Extraction

Ensemble of Identification Algorithms

Linguistic Processing: POS Tagging, Parsing

Language Dependent

Candidate Extraction: Phrase Templates

Language Dependent

Lexical Information: Translated from English

Language Independent

Ensemble of Algorithms: Combined predictions

Language Independent

(2) Because metaphor is central to reasoning about abstract concepts: If you fall into poverty, then you can get stuck in poverty and are unable to climb out. If you demolish a theory, then there is rubble left behind that you can use to build a new one.

Candidate extraction templates are the main language-specific resource required, although they depend upon syntactic parse trees.

If you fly through your homework, then progress is movement towards a destination. If someone drives you insane, then a mental state is a location you can travel to.

These templates allow multi-word phrasal components not adjacent in input text.

Ensemble Methods

Evaluation

The identification algorithms take as input a candidate expression and lexical information for each word in that expression. Lexical information includes Concreteness, Imageability, and Familiarity ratings from the MRC Psycholinguistic Database.

Performance Across Systems, English Domain Interaction: Dunn 2013a, b Source-Target:

Shutova, et al. 2013

Similarity:

Sporleder and Li 2010a, b

Abstractness:

Turney, et al. 2011

Using VU Amsterdam Metaphor Corpus

CPO Classifier (Recall) CPO Classifier (Precision)

Literality Filter 1

Priority Combiner

Majority Vote Combiner

Performance Across Genres, English (Only Ensemble System) Using VU Amsterdam Metaphor Corpus

2

Linguist-annotated corpus, very reliable

CPO Classifier (Accuracy)

Category Profile Overlap Classifier: Looks at amount of overlap between the Source and Target candidates in terms of category information from background Corpus, using lexical ratings to help categorize words.

Performance Across Languages (Only Ensemble System) Using in-house corpus

Random Forests Classifier: [Not part of evaluation]

Looks at degree of compatibility between the source noun and the set of nouns most commonly collocated with the target, operationalized as vectors of lexical properties.

Participant-annotated corpus, less reliable Only one algorithm used within ensemble