2009/10/23
2009/07/13
2009/05/25
Information Extraction: Algorithms and Prospects in a
2009/05/23
2009/04/13
2009/04/07
reading group
Labeling" on Monday the 6th of April (from 4:00 to 5:00)
2009/03/31
2009/03/29
Tree Kernels for semantic role labeling
Tree Kernels for semantic role labeling (32 pages)
0. Abstract
This paper is based on kernel functions to model parse tree properties in kernel based machines.
1. Introduction
Recently attention: automatic labeling of semantic roles. It requires:
1. Detection of the target predicate
2. Detection and classification of constituting arguments.
Previous work:
- Machine learning
- Linking theories semantic and syntax
Tree kernel for:
- Parsing re-ranking
- Relation extraction
- Name entity recognition.
Recent work with high labeling accuracy,
Joint inference on predicate-argument structure. For this -> extract features from sentence syntactic parse tree.
1- Feature extraction.
2- Combination with traditional attribute value models.
3- Feature engineering using kernel combinations and marking strategies.
4- Tree kernels….
5- Identify the most important structured features: highest-weighted features -> better classifiers.
“Tree kernels should always be applied”.
Limitations:
a) Poor impact on boundary detection.
b) Limited contribution to overall accuracy.
Tree kernels for srl: types:
- Canonical mappings
- Feature extraction functions
Carefully engineered tree kernels always increase accuracy.
2. Automatic shallow semantic parsing
3. Tree kernels
4. State of the art architecture for SRL
5. Structured feature engineering
6. Experiments
7. Discursion and conclusions
2009/03/28
Semantic Role Labeling: An Introduction to the Special Issue
® Sentence + designated verb: SRL identify the arguments of the verb. And label them with semantic roles.
® Steps:
¯ 1 Pruning.
¯ 2 Local scoring
¯ 3 Joint scoring
(4) Fixing common errors
Pruning
® Filtering the set of arguments for a given predicate.
¯ Any subsequence of words in the sentence is an argument candidate.
® Xue and Palmer (2004)
Joint scoring
® Good structure of labeled arguments.
® Arguments do not overlap,
® Core arguments do not repeat, etc.
® Re-ranking
® Probabilistic methods.
Conditional random fields
SRL architecture:
® Combination of systems and input annotations.
¯ Increase robustness,
¯ Gain coverage
¯ Reduce effects of parse errors
® One my combine:
¯ Output of independents srl basic systems
¯ Outputs from same srl s. changing input annotations or parameters.
Gaing of 2-3 F1 points.
® Joint labeling…
® Dependency parsing
® Combine parsing and srl in a single step. …
® Characterize candidate argument
phrase type, headword, …
® Characterize verb predicate +cntx
lemma, voice,
® Characterize the relation.
Syntactic + semantic.
Left/right position of the constituent with respect to the verb…
® Recall:
® 81% argument identification
® 95% assigned correct semantic role.
® SemEval-2007
® Disambiguation of 50 verbs.
® FrameNet
® 40 frames
® F1=92% asigning semantic roles
® F1 83% segmenting and labeling arg.
® Complete analysis of semantic roles on unseen texts,
® Precision 60s
® Recall 30s
® SRL relies on syntactic structure.
® Output by statistical parser 90% matching.
® Is common to use parser trees.
® Gold-standard trees.???
® Most of the errors are by having incorrect syntactic constituents.
® SRL relies on syntactic structure.
® Output by statistical parser 90% matching.
® Is common to use parser trees.
® Gold-standard trees.???
® Most of the errors are by having incorrect syntactic constituents.
® CoNLL-2005 Brown corpus annotated
® Performance drop below 70%
® They clamed errors are in assigning the semantic roles rather than identification of argument boundaries.
® Spanish and catalan, CESS-ECE corpus.
® 86% disambiguation predicates
® 83% labeling arguments.
® chinese
® semEval-2007 25K SENTENCE
® 80% core elements.
Top performing team use machine learning techniques