[Paper] Deep Learning-Based Document Modeling for Personality Detection from Text

• Paper
• Implementation: Personality-Detection
• Data Set
  • James Pennebaker and Laura King's stream-of-consciousness essay dataset
  • NRC Word-Emotion Association Lexicon

Practical Application of Personality Detection

• Product and Service Recommendation (People with similar personalities might have similar favors)
• Mental Health Diagnosis
• Forensics: Reduce the circle of suspects
• Human Resource: One's suitablitlty for certain jobs

Personality Theory Used in This Paper

Big Five Personality Trait

Basic Idea of the Method

1. Feed sentences from essays to convolution filter → Sentence model in the form of n-gram feature vectors
2. Aggregate the vectors of a document's sentences and combine them with Masiresse features to represent the document
3. Classification: Feed the document vectors into a fully connected neural network

Overview of the Method

1. Preprocessing

• Sentence Splitting
• Data Cleaning
• Unification (e.g. lowercase)

2. Document-level feature extraction

Mairesse baseline feature set (e.g. word count, average sentence length)

3. Filtering

Sentences without personliaty clues are dropped
(Based on NRC Word-Emotion Association Lexicon)

4. Word-level feature extraction

• word2vec
• Variable number of fixed-length word feature vectors → Variable number of sentences → Document

5. Classification

Deep CNN (Conolutional Nerual Network)

• Input
  • Words: Fixed-length feature vector using word2vec
  • Sentences: Variable number of word vectors
• Process
  • Word Vector is reduced to a fixed length vector of each sentence
  • Document: Variable number of such fixed-length sentence vector
  • Document vector is then reduced to a fixed-length document vector
  • This Document vector is then concatenated with document-level features
• Predict
  • Yes / No (5 different personality traits are trained separately)

Network Architecture in Detail

Main Steps (7 Layers)

Word Vectorization

• Layer 1: Input
  • \(R ^{D \times S\times W \times E}\)
  • Use Google's pretrained word2vec
  • In implementation, all the documents contain the same number of sentences.
    Shorter documents are padded shorter sentences with dummy words.

Sentence Vectorization

• Layer 2: Convolution
  • 3 convolutional filters: unigram, bigram, trigram
• Layer 3: Max Polling

Document Vectorization

• Layer 4: 1-max pooling

Classification: (Yes/No)

• Layer 5: Linear with Sigmoid activation
• Layer 6, 7
  • 2 Neuron (yes/no) Softmax Output (ReLU and tanh perform worse)
  • fully connected layer of size 200

Training

Objective Function: Negative Log Likelihood