Presentation is loading. Please wait.

Presentation is loading. Please wait.

Semantic Role Labelling Using Chunk Sequences Ulrike Baldewein Katrin Erk Sebastian Padó Saarland University Saarbrücken Detlef Prescher Amsterdam University.

Similar presentations


Presentation on theme: "Semantic Role Labelling Using Chunk Sequences Ulrike Baldewein Katrin Erk Sebastian Padó Saarland University Saarbrücken Detlef Prescher Amsterdam University."— Presentation transcript:

1 Semantic Role Labelling Using Chunk Sequences Ulrike Baldewein Katrin Erk Sebastian Padó Saarland University Saarbrücken Detlef Prescher Amsterdam University Amsterdam

2 1. Representation for Classification Usual choice: Classify Constituents  Intuition: one argument  one constituent  Not available in this task Classify words?  Too fine-grained Classify chunks?  Data anylsis: Not always the right level 34% of arguments: more than one chunk 13% of arguments: do not respect chunk boundaries

3 Chunk sequences as classification instances Sequences of chunks and chunk parts  Adaptive level of structure  „Potential constituents“ [ NP Britain´s] [ NP manufacturing industry] [ VP is transforming] [ NP itself] [ VP to boost] [ NP exports] ARG0 = NP_NP V = VP[VBG] ARG1 = NP ARG2 = VP_NP

4 Frequency-based chunk sequence filtering Filter 1: Use only sequence types which realise arguments in training set 1089 types  Zipf distributed NP (23,000) S (5,000) NP_PP_NP_PP_NP_VP _PP_NP_NP (1) Filter 2: Use only frequent sequence type (f(s)>10) Examine material between sequence and target  Divider sequences  Also Zipf distributed Empty divider (14,000) NP (10,000)  Similar to „Path“ Filter 3: Use only seq.s with frequent divider (f(d)<10) Filter 4: Use only seq.s co- occurring frequently with some divider (f(s,d)<5)

5 Results of filtering Leaves 87 sequence types (was 1089)  43,777 tokens in devel set (about 1 seq / word)  8,698 are proper arguments (about 20%) Bad news: representation loses 16% of proper arguments

6 2. Features „Shallow features“:  Simple properties „Higher-level features“:  Syntactic properties (mostly heuristic) „Divider features“:  Shallow and higher-level properties of dividers

7 EM-based clustering Measure fit between objects y 1 (pred:arg) and y 2 (sequence)  Example: How well does NP fit give:A1? y 1 and y 2 are independent and generated by cluster  p(y 1,y 2 ) =  c p(c) p(y 1 |c) p(y 2 |c)  EM derives clusters from training data  Intention: Generalisation within clusters Features: e.g. „most likely argument slot for this sequence for this predicate“

8 3. Procedure Filter sequences from training set Compute features for sequence tokens and their dividers (training + development + test set) Estimate Maximum Entropy model on training set Classify sequences from devel / test set Recover semantic parses

9 Two-step classification procedure Classifier 1: Argument recognition  Binary decision about argumenthood  All argument classes conflated into ARG Classifier 2: Argument labelling  Consider only sequences assigned ARG by step 1

10 4. Classification result: Sequence chart Themanwiththebeardsleeps A0 (70%), A1 (20%)NOLABEL (70%), AM-MOD(25%) A0 (60%), NOLABEL (40%) A0 (65%), A1 (25%) Need to find optimal „semantic parse“ of argument labels

11 Semantic parse recovery Find most probable semantic parse p = (l 1,l 2,...) Step 1: Beam search:  Simple probability model with independence assumption: P bs (l 1,l 2,...) =  i P c (l i ) Step 2: Reestimation  Global considerations: [A0 A0]  Use counts from training set: P(l 1,l 2,...) = P bs (l 1,l 2,...) * P tr (l 1,l 2,...)

12 5. Results (Development Set) PrecisionRecallF-score Upper Bound10083.390.9 Step 1 (ARG only)77.360.166.1 Final64.941.650.7 Upper Bound: given by lost chunk sequences But filtering is necessary  Only sequence frequency filtering (filter 1 and 2): Good news: 9% arguments are lost (now 16%) Bad news: 127,000 sequences (now 44,000)  Argument recognition much more difficult  F-score with same features only 0.38

13 Results Two steps have different profiles  Arg identification: shallow and divider features important  Arg labelling: shallow and higher-level features important Clustering features unsuccessful: Increase precision at cost of recall  Feature „most probable label for sequence“  Successful in SENSEVAL-3 model Largest problem is recall PrecisionRecallF-score Upper Bound10083.390.9 Step 1 (ARG only)77.360.166.1 Final64.941.650.7

14 What I talked about... and more Chunk sequences for SRL  Adaptive representation with „higher-level“ features  Recall problem (Filtering loses proper arguments)  EM-based features promising, but currently not helpful Since submission  Maxent vs. memory-based learner: virtually same result Left to do  Detailed error analysis  More intelligent filtering  Better features


Download ppt "Semantic Role Labelling Using Chunk Sequences Ulrike Baldewein Katrin Erk Sebastian Padó Saarland University Saarbrücken Detlef Prescher Amsterdam University."

Similar presentations


Ads by Google