SemiSupervised Semantic Role Labeler Based on SemiSupervised Semantic

Semi-Supervised Semantic Role Labeler Based on “Semi-Supervised Semantic Role Labeling via Structural Alignment” by Furstenau and Lapata, 2011 Advisors: Prof. Michael Elhadad and Mr. Avi Hayoun By Efrat Hazani

Frame Semantics �Frame Semantics is a theory that relates linguistic semantics to knowledge and experience �The meaning of words depend on contexted experiences

For example: �“I always eat cereal for breakfast” – breakfast : first meal of the day �“Breakfast is served at any time” – breakfast: a particular combination of foods typically eaten at breakfast

What is a frame? �Structured representation of concept �Identifies the experience as a type, and gives structure and meaning to the relationships, objects and events within the experience �A word represents a category of experience, and thus evokes a frame of semantic knowledge �A word can evoke different frames

Frame Elements �Frame evoking element (FEE) – the word (or lexical unit) which evokes the frame �Frame elements (FEs) – words which have semantic roles in the frame �Semantic roles describe the relations between a predicate and its arguments �Semantic roles are independent from syntactic relations

For example: CAUSE_HARM Agen t Victim Body_part “Lee punched John in the eye” APPLY_HEAT Agent Food Heating_instrume nt “She was frying eggs on a camp stove”

Frame. Net A project building a lexical database of English that is both human- and machinereadable, based on annotating examples of how words are used in actual texts.

Frame. Net – what is it good for? Provide a unique training dataset for semantic role labeling, used in applications such as: • information extraction • machine translation • event recognition • sentiment analysis

The Goal of the Project Create a larger collection of annotated sentences

The General Idea �Input: ◦ A set L of sentences labeled with frames and roles (seed corpus) ◦ A set U of unlabeled sentences (expansion corpus) �For every unlabeled sentence u: ◦ Find the most similar labeled sentence l ◦ Annotate u according to l’s annotation. �For every labeled sentence l: ◦ Return the k best newly annotated sentences according to l.

Measuring Similarity �Sentences are represented by dependency graphs �An alignment between two graphs: ◦ A partial injective function σ : M → N ∪ {} ◦ Domain M - a labeled graph ◦ Range N - an unlabeled graph ◦ x ∈ M isaligned to x’ ∈ N by σ, iff σ(x) = x’ �Find a graph M with the best alignment to the unlabeled graph N

Measuring Similarity An example: Impactor IMPACT Impacte e “His back thudded against the wall” “The rest of his body thumped against the front of the cage”

Measuring Similarity Calculating score for alignment σ : � - the lexical similarity between x and σ(x) (a value between 0 and 1) � - is the grammatical relation between x 1 and x 2 equal to the grammatical relation between σ(x 1) and σ(x 2) (0 or 1)

Measuring Similarity Calculating score for alignment σ : � α - the relative weight of syntactic similarity compared ≈ 0. 55) to lexical similarity (optimal value � C - normalizing factor

Summary �We want more sentences labeled with semantic roles �Expand the set of annotated sentences by: ◦ For every unlabeled sentence, finding an optimal alignment with some labeled sentence ◦ Projecting annotation from the labeled sentence to the unlabeled sentence