Presentation is loading. Please wait.

Presentation is loading. Please wait.

Progress report on Semantic Role Labeling

Similar presentations


Presentation on theme: "Progress report on Semantic Role Labeling"— Presentation transcript:

1 Progress report on Semantic Role Labeling
Abdul-Lateef Yussiff nd June, 2011 COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

2 Introduction Semantic analysis of a sentence is a very important step towards understanding of natural Language There are two approaches Deep semantic analysis : given the current state of the art technology and human skills, it is too complex to follow this approach Shallow semantic analysis: Process of annotating the predicate argument structure in text with semantic labels (Jurafsky et al., 2005) COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

3 Some widely recognized Semantic Roles
Description Examples Agent Initiator of action The pilot landed the plane as lightly as a feather. Patient Affected by action, undergoes change of state John broke the window. Theme Entity moving or being “located” John threw the ball. The picture hangs above the fireplace Experiencer Perceives action but not in control Lee noticed the cat slip through the partially open door Beneficiary For whose benefit action is performed The Smiths rented an apartment for their son. Instrument Intermediary/means used to perform an action He shot the wounded buffalo with a rifle. Location Place of object or action The band played on the stage. COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

4 PropBank Tags Arguments of a verb are labeled ARG0 to ARG5
ARG0 is the AGENT ARG1 is the PATIENT Adjunctive arguments ARGM-Loc for locatives ARGM-TMP for temporal ARGM-MNR for manner COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

5 Propbank’s representation of semantic role
[The pilot A0] landed predicate [the plane A1] as lightly as a feather ARG-M. JohnA0 broke predicate [the window A1]. HeA0 shot predicate [the wounded buffaloA1] with a rifle. [The acquisition ARG1] was completed predicate [in September ARG-TMP] COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

6 Semantic Role Labeling and Machine Learning
The task is to select a constituent’s semantic role with respect to a given predicate. To find the best semantic analysis of an entire sentence, it is recommended to extend beyond taking a decision on single constituent because the constituents are interdependent on each other. The basic approach is to use a syntactic parser as input to the system. COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

7 Goal of this Research Goal is to use a machine learning sequence tagging approach to identify and label constituents with semantic roles Conditional Random Field (CRF) was chosen because of its advantages over all other sequence tagging techniques. COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

8 Experimental features
Word POS The syntactic class or category of the words Phrase type It indicates the syntactic categories of the phrase expressing the semantic roles. Phrase type are extracted automatically from parse trees generated by the parser Examples are Noun Phrase (NP), Prepositional Phrase (PP), etc. COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

9 Example Confidence in the pound is widely expected to take another sharp dive if trade figures for September, due for release, fail to show a substantial improvement from July and August’s near- record deficit. ARG-0: Confidence in the pound Predicate: take ARG-1: Another Sharp dive ARG-M: IF trades figures for September, due for release, fail to show a substantial improvement from July and August’ near-record deficit COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

10 Sequence Tagging with Conditional Random Field
Training Stage F1 F2 ….FN Label Confidence in the pound is widely expected to take another sharp dive if trade figures for September, due for release, fail to show a substantial improvement from July and August’s near- record deficit. Confidence NN B-NP B-A0 in IN B-PP I-A0 the DT B-NP I-A0 pound NN I-NP I-A0 is VBZ B-VP O widely RB I-VP O expected VBN I-VP O to TO I-VP O take VB I-VP PRED Sequence Tagging with Conditional Random Field SRL Output COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

11 How to train the SRL with Mallet’s SimpleTagger
The Machine learning tools used in this exercises is Mallet, machine learniing toolkit provided by the Umarst. To train the file, I used the following command Java –cp “/root/lateef/mallet-2.0.6/class: /root/lateef/mallet-2.0.6//lib/mallet-deps.jar” cc.mallet.fst.SimpleTagger --train true –fully-connected false -- weights sparse --model-file wordmodel /root/lateef/semanticRole/test.txt After executing the above command, it produces a model file named wordmodel COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

12 After creating the training of the sample data which created the model file, we need to test and evaluate our sample model file with the following command --model-file wordposmodel test1.txt Rockwell International Corp. 's Tulsa unit said it signed a tentative agreement extending its contract with Boeing Co. to provide structural parts for Boeing's 747 jetliners. COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

13 also RB (AM-DIS*AM-DIS) changing VBG (V*V) hands NNS (A2*A2)
Number of predicates: 26 East NNP A0 Rock NNP I-A0 also RB (AM-DIS*AM-DIS) changing VBG (V*V) hands NNS (A2*A2) at IN AM-MNR CD I-AM-MNR yen NN I-AM-MNR , , I-AM-MNR up RB I-AM-MNR from IN I-AM-MNR CD I-AM-MNR COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

14 Observation I observed the result produced is not the same as the inputs data. I am working on the source of the problems. Reason why different output is coming out as the result of the input file? I will try to resolve the problem before the next presentation, (the problem might be coming from the eclipse or the simpleTagger class itself.) COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.

15 Questions and Suggestions
Thank You COPYRIGHT © 2009, Cognitive Informatics LAB, ALL RIGHTS RESERVED.


Download ppt "Progress report on Semantic Role Labeling"

Similar presentations


Ads by Google