Matthew Graci, Robert Thorstad, Theodore Waters

0 downloads 0 Views 621KB Size Report
Matthew Graci, Robert Thorstad, Theodore Waters, Robyn Fivush, and Phil Wolff. For more information, contact [email protected]. Input: Trauma Narratives.
Deriving semantic coherence of personal memory using neural networks Matthew Graci, Robert Thorstad, Theodore Waters, Robyn Fivush, and Phil Wolff For more information, contact [email protected]

Automated coding of Narrative Coherence:

• More coherent personal narratives are generally related to higher wellbeing (Waters & Fivush, 2015) and less depressive symptoms (Baeger & McAdams, 1999), whereas less coherent narratives can be an indicator of clinical disorders (e.g. borderline personality disorder) (Adler et al., 2012)

Main idea: coherent narratives have similar adjacent sentences 1) Word in the vocabulary represented as vectors using neural network (word2vec; Mikolov, Chen, Corrado, & Dean, 2013) with the constraint that similar words are represented by similar vectors

Relation between Hand-coded and Automated Coherence

• However, mixed findings suggest coherence is multifaceted (Adler et al., 2016)

Relation between Hand-coded and 2nd order Automated Coherence R² = 0.1098*

10

R² = 0.0567*

10

9

9

8

8

Hand-coded coherence

• Coherent narratives provide a window into one’s understanding of an event and one’s emotional response to it, in addition to how the sequence of events unfolded

Results

Hand-coded Coherence

Introduction

7 6 5 4 3

7 6 5 4 3 2

2

1

1

0 0

0.6 0.6

• Preliminary machine learning approaches successfully utilized underlying semantic components of language usage to predict levels of psychological disorder (e.g. onset of schizophrenia) (Bedi et al. 2015)

0.7

0.75

0.8

0.85

0.9

0.65

0.7

0.95

0.75

0.8

0.85

0.9

0.95

2nd order Automated Coherence

Automated Coherence

Relation between Positive Self-view Wellbeing and 2nd Order Automated Coherence

Illustration of word2vec 2

Input: Text

Output: Word Vectors

R² = 0.0806* 1.5

1

Positive Self-view Wellbeing

• Neural networks can explore underlying semantics of language usage, which can provide insights into components of coherence in 3 ways: Space (analysis in higher dimensional space allowing for new insights); Scale (studying more narratives); and Speed (quicker analysis)

0.65

Objectives • To create an automated coherence measure

0.5

0

-0.5

-1

-1.5

• To establish convergent validity with a dataset with measures of psychological health and hand-coded coherence • Explore ways to improve the tool and its potential to help answer open questions in the literature

Method Participants

103 Emory undergraduates (56 female; Mage = 18.87) wrote about events in their lives Narrative elicitation

“As you write about the event you have in mind please describe, in detail, what happened, where you were, who was involved, what you did, and what you were thinking and feeling during the event. Also, try to convey what impact this single unique event has had on you, and why it is an important event in your life. Try to be specific and provide as much detail as you can. ”

-2

-2.5 0.6

1) Positive self-view Wellbeing Rosenberg Self Esteem Satisfaction with Life Scale Environmental Mastery Autonomy Self-acceptance

2) Purposeful living Wellbeing Personal Growth Purpose in Life

3) Positive relations Wellbeing Social Acceptance Social Actualization

Social Integration Social Contributions Social Coherence

Positive Relations

Hand-Coded Narrative Coherence: Composite of 3 subscales: • Theme - expresses a salient topic, wherein the topic is elaborated and resolved in some manner • Context - narrator places the personal experience in a specific location and timeline • Chronology - narrator places the actions and events of the experience in a discernable timeline

0.7

0.75

0.8

0.85

0.9

0.95

2nd Order Automated Coherence

2) Represent dataset as vectors: Using formula, , where we averaged N word vectors in a sentence from index k=1 to N to compute a sentence vector Illustration of word2vec on Dataset

Input: Trauma Narratives

Output: Word Vectors

Discussion • There is more to the words we express than what the words we say • The statistics of semantics that neural networks capture an inform narrative theory • Automated measure related to handed-coded measure of coherence • Looking at the semantics on language helps support the idea of disorganization vs rigidity in coherence • Optimal level of semantic similarity • On lower end is disorganization • On upper end is rigidity

Measures Psychological Health Measures: Principal Component Analysis was performed on health measures

0.65

3) Compute Coherence:

Future Directions

Using formula, { cos(Li, Li+1) } , where coherence is computed as the average similarity of all adjacent sentences S1

S2

S2

S3

S3

S4

S1

S3

S2

S4

S3

S5

• Replicate findings • Improve tool • Expand window size of tool • Explore syntactic component of coherence • Vectorize sentences • Investigate emotional coherence Acknowledgments: Special thanks to the Family Narratives Lab and the Mind and Language lab for the feedback and ideas, as the Emory College Collaborative Research Catalyst Award for funding