Predicting Power Relations between Participants in Written Dialog from a Single Thread
Prabhakaran, Vinodkumar and Rambow, Owen

Article Structure

Abstract

We introduce the problem of predicting who has power over whom in pairs of people based on a single written dialog.

Introduction

Computationally analyzing the social context in which language is used has gathered great interest within the NLP community recently.

Motivation

Early NLP-based approaches such as Bramsen et a1.

Data

In this work, we use the version of Enron email corpus by Yeh and Harnly (2006) which captures the thread structure of email exchanges.

Problem Formulation

Let t denote an email thread and Mt denote the set of all messages in t. Also, let Pt be the set of all participants in t, i.e., the union of senders and recipients (T0 and CC) of all messages in Mt.

Structural Analysis

In this section we analyze various features that capture the structure of interaction between the pairs of participants in a thread.

Predicting Direction of Power

We build an SVM-based supervised learning system that can predict HP (p 1 , p2) to be either superior or subordinate based on the interaction within a thread If for any pair of participants (191,192) 6 RIPPt.

Conclusion

We introduced the problem of predicting who has power over whom based on a single thread of written interactions.

Topics

feature set

Appears in 6 sentences as: feature set (4) feature sets (2)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. This best-performing system uses our new feature set .
    Page 1, “Introduction”
  2. THRPR: This feature set includes two meta-
    Page 3, “Structural Analysis”
  3. data based feature sets — positional and verbosity.
    Page 3, “Structural Analysis”
  4. We use another feature set LEX to capture word ngrams, POS (part of speech) ngrams and mixed ngrams.
    Page 5, “Predicting Direction of Power”
  5. We also performed an ablation study to understand the importance of different slices of our feature sets .
    Page 5, “Predicting Direction of Power”
  6. Using this feature set , we obtain an accuracy of 73.0% on a blind test.
    Page 5, “Conclusion”

See all papers in Proc. ACL 2014 that mention feature set.

See all papers in Proc. ACL that mention feature set.

Back to top.

manual annotations

Appears in 6 sentences as: manual annotation (2) manual annotations (4)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. Another limitation of (Prabhakaran and Rambow, 2013) is that we used manual annotations for many of our features such as dialog acts and overt displays of power.
    Page 2, “Motivation”
  2. Relying on manual annotations for features limited our analysis to a small subset of the Enron corpus, which has only 18 instances of hierarchical power.
    Page 2, “Motivation”
  3. Like (Prabhakaran and Rambow, 2013), we use features to capture the dialog structure, but we use automatic taggers to generate them and assume no manual annotation at all at training or test time.
    Page 2, “Motivation”
  4. We excluded a small subset of 419 threads that was used for previous manual annotation efforts, part of which was also used to train the DA and GDP taggers (Section 5) that generate features for our system.
    Page 2, “Data”
  5. DIAPR: In (Prabhakaran and Rambow, 2013), we used dialog features derived from manual annotations — dialog acts (DA) and overt displays of power (ODP) — to model the structure of interactions within the message content.
    Page 3, “Structural Analysis”
  6. In this work, we obtain DA and GDP tags on the entire corpus using automatic taggers trained on those manual annotations .
    Page 3, “Structural Analysis”

See all papers in Proc. ACL 2014 that mention manual annotations.

See all papers in Proc. ACL that mention manual annotations.

Back to top.

structural feature

Appears in 5 sentences as: structural feature (2) structural features (2) structural features’ (1)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. We propose a new set of structural features .
    Page 1, “Abstract”
  2. In order to mitigate this issue, we use an indicator feature for each structural feature to denote whether or not it is valid.
    Page 5, “Predicting Direction of Power”
  3. The performance of the system using each structural feature class on its own is very low.
    Page 5, “Predicting Direction of Power”
  4. Perplexingly, adding all structural features to LEX reduces the accuracy by around 2.2 percentage points.
    Page 5, “Predicting Direction of Power”
  5. Removing the indicator feature denoting the structural features’ validity also reduces the performance of the system.
    Page 5, “Predicting Direction of Power”

See all papers in Proc. ACL 2014 that mention structural feature.

See all papers in Proc. ACL that mention structural feature.

Back to top.

bigrams

Appears in 4 sentences as: Bigrams (1) bigrams (3)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. Baseline (Always Superior) 52.54 Baseline (Word Unigrams + Bigrams ) 68.56 THRNCW 55.90 THRPR 54.30 DIAPR 54.05 THRPR + THRNew 61.49 DIAPR + THRPR + THRNew 62.47 LEX 70.74 LEX + DIAPR + THRPR 67.44 LEX + DIAPR + THRPR + THRNew 68.56 BEST (= LEX + THRNeW) 73.03 BEST (Using p1 features only) 72.08 BEST (Using IMt features only) 72.11 BEST (Using Mt only) 71.27 BEST (No Indicator Variables) 72.44
    Page 5, “Predicting Direction of Power”
  2. We found the best setting to be using both unigrams and bigrams for all three types of ngrams, by tuning in our dev set.
    Page 5, “Predicting Direction of Power”
  3. We also use a stronger baseline using word unigrams and bigrams as features, which obtained an accuracy of 68.6%.
    Page 5, “Predicting Direction of Power”
  4. The word unigrams and bigrams baseline obtains an accuracy of 68.3%.
    Page 5, “Predicting Direction of Power”

See all papers in Proc. ACL 2014 that mention bigrams.

See all papers in Proc. ACL that mention bigrams.

Back to top.

unigrams

Appears in 4 sentences as: Unigrams (1) unigrams (3)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. Baseline (Always Superior) 52.54 Baseline (Word Unigrams + Bigrams) 68.56 THRNCW 55.90 THRPR 54.30 DIAPR 54.05 THRPR + THRNew 61.49 DIAPR + THRPR + THRNew 62.47 LEX 70.74 LEX + DIAPR + THRPR 67.44 LEX + DIAPR + THRPR + THRNew 68.56 BEST (= LEX + THRNeW) 73.03 BEST (Using p1 features only) 72.08 BEST (Using IMt features only) 72.11 BEST (Using Mt only) 71.27 BEST (No Indicator Variables) 72.44
    Page 5, “Predicting Direction of Power”
  2. We found the best setting to be using both unigrams and bigrams for all three types of ngrams, by tuning in our dev set.
    Page 5, “Predicting Direction of Power”
  3. We also use a stronger baseline using word unigrams and bigrams as features, which obtained an accuracy of 68.6%.
    Page 5, “Predicting Direction of Power”
  4. The word unigrams and bigrams baseline obtains an accuracy of 68.3%.
    Page 5, “Predicting Direction of Power”

See all papers in Proc. ACL 2014 that mention unigrams.

See all papers in Proc. ACL that mention unigrams.

Back to top.

SVM

Appears in 3 sentences as: SVM (3)
In Predicting Power Relations between Participants in Written Dialog from a Single Thread
  1. Handling of undefined values for features in SVM is not straightforward.
    Page 4, “Predicting Direction of Power”
  2. Most SVM implementations assume the value of 0 by default in such cases, conflating them
    Page 4, “Predicting Direction of Power”
  3. Since we use a quadratic kernel, we expect the SVM to pick up the interaction between each feature and its indicator feature.
    Page 5, “Predicting Direction of Power”

See all papers in Proc. ACL 2014 that mention SVM.

See all papers in Proc. ACL that mention SVM.

Back to top.

PDF available from www.aclweb.org.
Send us an email