GlobalTrait: Personality Alignment of Multilingual Word Embeddings

Farhad Bin Siddique, Dario Bertero, Pascale Fung

Here is a concise, engaging summary of the paper "GlobalTrait: Personality Alignment of Multilingual Word Embeddings".

TL;DR

The authors introduce GlobalTrait, a novel framework that aligns multilingual word embeddings based on psychological traits rather than just dictionary definitions. By mapping personality-correlated words from low-resource languages into a shared space with English, the model enables highly effective cross-lingual personality detection, boosting binary classification F-scores by an average of 8.4 points.

What Problem Does This Paper Address?

Predicting a user's "Big Five" personality traits (Extraversion, Conscientiousness, Agreeableness, Neuroticism, and Openness) from text is incredibly useful for building empathetic virtual agents, personalized marketing, and HR screening. However, data labeled with personality traits is scarce, particularly for languages other than English.

While we can use multilingual word embeddings (like fastText or MUSE) to map words with similar semantic meanings across languages, the authors discover a major catch: words with the same semantic meaning do not necessarily imply the same personality traits across different cultures. For instance, standard translation alignments fail to capture the nuanced, culturally specific ways people express personality. GlobalTrait addresses this gap by creating a trait-specific alignment that allows personality features learned in high-resource languages (English) to transfer effectively to lower-resource ones.

Methods

The GlobalTrait pipeline operates in three main phases:

  1. Initial Semantic Mapping: The authors use pre-trained fastText embeddings and apply the MUSE framework (an unsupervised GAN-based approach) to align four languages (English, Spanish, Italian, and Dutch) into a shared semantic space.
  2. Trait-Based Alignment (The Core Innovation): Using tf-idf features, they identify the top 3,000 words most positively correlated with each of the Big Five traits per language. They then train five distinct, trait-specific orthogonal mapping matrices ($W$) per source language. This forces words that signal the same personality trait across different languages to cluster closely together in a target space (English).
  3. Two-Channel CNN Classifier: The text is fed into a Convolutional Neural Network with two parallel embedding channels:
  4. Channel 1 (Dynamic): Standard unaligned multilingual embeddings that update during training.
  5. Channel 2 (Static): The newly engineered GlobalTrait aligned embeddings.

The features from both channels are concatenated, max-pooled, and passed to a fully connected layer for final classification or regression.

[Input Text] ---> Channel 1: Unaligned Multilingual Embeddings (Dynamic) ---\
             ---> Channel 2: GlobalTrait Aligned Embeddings (Static)   ---> [Max-Pooling] ---> [FC Layer] ---> [Output]

Datasets

The model was trained and evaluated on two primary datasets:

  • PAN 2015 Author Profiling Dataset: Twitter data labeled via the BFI-10 personality questionnaire. It includes 152 English, 110 Spanish, 38 Italian, and 34 Dutch users.
  • BIT Speaker Personality Corpus: A separate Chinese dataset consisting of 498 speech clips transcribed via Automatic Speech Recognition (ASR). This was used to test how well the model generalizes to non-Latin scripts.

Key Insights and Ideas

  • Semantic $\neq$ Personality Alignment: Visualization using t-SNE revealed that words highly correlated with specific traits (like Extraversion) clustered strictly by language rather than by trait in standard multilingual spaces. Cultural differences mean direct translations don't carry the same psychological weight.
  • One Word, Five Meanings: Because GlobalTrait trains a separate matrix for each of the Big Five traits, a single word can have five different vector representations depending on the trait being analyzed. For example, the Spanish word 'mundo' (world) maps closest to 'travel, flights, fresh' under the Openness trait, but aligns closest to 'parties, love, life' under Extraversion.

Key Results

  • Classification Boost: The CNN-GlobalTrait model achieved the highest average F-scores across all non-English languages, boosting the average F-score from 65.0 (monolingual baseline) to 73.4 (an absolute improvement of +8.4).
  • Language Breakdowns: The model achieved an average F-score of 74.6 in Spanish, 68.2 in Italian, and 77.5 in Dutch.
  • Cross-Script Generalization: When evaluated on the Chinese dataset, CNN-GlobalTrait outperformed both the monolingual baseline and unaligned multilingual models, proving that the alignment holds true even across entirely different writing systems.
  • Regression Performance: In continuous trait score prediction, the model achieved the best performance in predicting the Extraversion trait compared to previous character-based RNN architectures.

Limitations and Future Work

  • Dataset Imbalance: The PAN 2015 dataset is quite small and imbalanced for Italian and Dutch, causing simpler logistic regression baselines to occasionally outperform the CNN on specific traits due to easier convergence.
  • Trait Variance: The model struggled to perform as well on the Openness trait compared to Extraversion, suggesting some traits may be inherently harder to align cross-culturally using word frequencies alone.
  • Future Directions: The authors plan to integrate GlobalTrait embeddings into sequential architectures like Bi-directional RNNs and Hierarchical Attention Networks, and scale up training using massive datasets like the 22-million-status myPersonality Facebook dataset.

Summary updated

Abstract

We propose a multilingual model to recognize Big Five Personality traits from text data in four different languages: English, Spanish, Dutch and Italian. Our analysis shows that words having a similar semantic meaning in different languages do not necessarily correspond to the same personality traits. Therefore, we propose a personality alignment method, GlobalTrait, which has a mapping for each trait from the source language to the target language (English), such that words that correlate positively to each trait are close together in the multilingual vector space. Using these aligned embeddings for training, we can transfer personality related training features from high-resource languages such as English to other low-resource languages, and get better multilingual results, when compared to using simple monolingual and unaligned multilingual embeddings. We achieve an average F-score increase (across all three languages except English) from 65 to 73.4 (+8.4), when comparing our monolingual model to multilingual using CNN with personality aligned embeddings. We also show relatively good performance in the regression tasks, and better classification results when evaluating our model on a separate Chinese dataset.
BibTeX
@inproceedings{siddique2019globaltrait,
  author = {Farhad Bin Siddique and Dario Bertero and Pascale Fung},
  title = {{GlobalTrait: Personality Alignment of Multilingual Word Embeddings}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33017015},
}
Figure