GlobalTrait: Multilingual Personality Recognition

• GlobalTrait is a framework for multilingual personality recognition that aligns trait-specific words across languages despite cultural and lexical variations.
• It employs orthogonal rotational mapping and adversarial training to co-locate words correlated with the same personality traits in a shared embedding space.
• Empirical results show significant F1-score improvements across multiple languages, highlighting its effectiveness in cross-linguistic personality classification.

GlobalTrait is a framework for personality-trait recognition in a multilingual context that explicitly addresses the misalignment between semantic similarity and personality-correlated linguistic signals across languages. Standard cross-lingual embeddings cluster words with similar semantics but fail to capture alignment of personality traits due to cultural and lexical variation. GlobalTrait introduces a trait-specific alignment methodology to ensure that words positively correlated with the same Big Five personality dimension are proximate in a shared embedding space, thereby optimizing cross-linguistic transfer for personality classification and regression tasks (Siddique et al., 2018).

1. Multilingual Personality Recognition: Motivation and Challenges

Automatic personality-trait recognition requires models that can generalize across languages, cultures, and varying lexical signals. However, existing multilingual embeddings (e.g., MUSE) focus exclusively on semantic alignment, not on sociolinguistic or psychological correlates. For instance, the Spanish word “mundo” (“world”) may reflect Openness in English but Extraversion in Spanish contexts. This lexical–cultural disconnect means that naive multilingual embeddings cluster trait-relevant words by language, limiting transferability and impeding fine-grained personality detection.

GlobalTrait proposes a solution by formulating trait-level alignment objectives to co-locate trait-positive words, regardless of language, in the embedding space. This enables robust transfer from high-resource languages (such as English) to low-resource languages where personality-labeled data is scarce.

2. Trait-Specific Rotational Mapping

Let $d=300$ denote the embedding dimensionality. For each source language $\ell$ and for each trait $t \in \{Extr, Agr, Cons, Emot, Openn\}$, GlobalTrait learns an orthogonal mapping

$W_t^{(\ell)} \in O_d(\mathbb{R})$

so that positive-trait words in language $\ell$ are rotated to align closely with their English equivalents.

Given:

• $X_t^{(\ell)} = \{x_1, \dots, x_n\}$: embeddings of the top $N$ positive-trait words in language $\ell$
• $Y_t^{(en)} = \{y_1, \dots, y_m\}$: embeddings of the corresponding positive-trait words in English

the objective is:

$W_t^{(\ell)} \cdot X_t^{(\ell)} \approx Y_t^{(en)}$

The mapping $W_t^{(\ell)}$ is a $d \times d$ orthogonal matrix (i.e., $W_t^{(\ell)T} W_t^{(\ell)} = I$), ensuring that the transformation preserves distances and angles within the embedding space. A distinct mapping is learned for each trait–language pair.

3. Adversarial Training Objectives

Trait alignment training uses an adversarial approach:

• Discriminator $D$ with parameters $\theta_D$: attempts to distinguish between mapped source vectors $W_t x_i$ and English target vectors $y_j$.
• Mapping $W_t$: trained to make $W_t x_i$ indistinguishable from $y_j$, subject to the orthogonality constraint.

Formally:

• Discriminator loss:

$$L_D(\theta_D \mid W_t) = -\frac{1}{n} \sum_{i=1}^n \log P(\text{source}=1 \mid W_t x_i) -\frac{1}{m} \sum_{j=1}^m \log P(\text{source}=0 \mid y_j)$$

• Mapping loss:

$$L_W(W_t \mid \theta_D) = -\frac{1}{n} \sum_{i=1}^n \log P(\text{source}=0 \mid W_t x_i) -\frac{1}{m} \sum_{j=1}^m \log P(\text{source}=1 \mid y_j)$$

The alternation between discriminator and mapping updates follows the min–max adversarial optimization scheme as in MUSE.

4. Identification and Selection of Trait-Correlated Words

Selection of words positively correlated with each trait in each language employs a tf-idf–based approach. For each trait $t$, users are split into high and low trait groups via median-split on self-reported scores. Each user’s aggregate text is represented using tf-idf features. Vocabulary is then ranked by the average tf-idf difference between high-trait and low-trait groups; the top $N$ (e.g., $N=3{,}000$) most positively associated words are selected. These comprise the sets $X_t^{(\ell)}$. For visualization, the top 750 words are used in t-SNE projections.

5. Multichannel CNN Architecture for Personality Classification

Classification leverages both aligned and unaligned embeddings within a two-channel convolutional neural network:

• Channel 1 (dynamic): Standard multilingual embeddings, set as trainable parameters during CNN optimization.
• Channel 2 (static): GlobalTrait-aligned embeddings, frozen during CNN training.

Each channel applies 1D convolutions with window sizes $3$, $4$, and $5$, each with $64$ feature maps. Max-over-time pooling is performed for each filter. The pooled feature maps across all windows and channels are concatenated and passed through a fully connected hidden layer (size 100, tanh activation), terminating in a softmax layer for binary trait classification.

Training specifics:

• Loss function: binary cross-entropy
• Optimizer: Adam with learning rate $10^{-4}$
• Batch size: $10$, epochs: $100$

6. Experimental Protocols and Empirical Results

Experiments utilize:

• PAN-2015 Author Profiling corpus (English, Spanish, Italian, Dutch; tweets + self-reported BFI-10)
• BIT Speaker Personality Corpus (Chinese, ASR-transcribed speech)

Evaluation is conducted via:

• F₁-score for binary classification (median split)
• RMSE for regression on raw trait scores

Baselines:

• Monolingual embeddings + logistic regression (Lgr-mono)
• Multilingual (unaligned) embeddings + logistic regression (Lgr-multi)
• Monolingual embeddings + CNN (CNN-mono)
• CNN with GlobalTrait-aligned embeddings (CNN-GlobalTrait)

Key averaged F₁-score results (CNN-mono vs CNN-GlobalTrait):

Language	CNN-mono	CNN-GlobalTrait	ΔF₁
Spanish	71.1	74.6	+3.5
Italian	68.1	68.2	+0.1
Dutch	68.4	77.5	+9.1
Chinese	58.8	62.3	+3.5

An overall average F-score improvement of +8.4 points (65→73.4) is reported when comparing monolingual vs multilingual CNN+GlobalTrait across non-English languages. Regression RMSE improvements are also observed for Spanish and Italian (Siddique et al., 2018).

7. Embedding Space Geometry and Qualitative Analysis

t-SNE projections of the top trait words demonstrate the effect of GlobalTrait alignment:

• Prior to alignment, trait-positive words from Spanish, Italian, and Dutch form distinct, language-specific clusters, even when semantically or psychologically related.
• Post-alignment, these words aggregate in the same region, indicating trait-level proximity.

A prominent illustration is the Spanish word “mundo” (“world”), which, under the Openness mapping, clusters with English terms like “travel” and “flights”, but under Extraversion, is repositioned near “parties”, “love”, and “life”, evidencing trait-specific differential alignment.

8. Limitations, Extensions, and Applications

Identified limitations:

• Small labeled data per language (PAN-2015: <150 users per language)
• tf-idf–based selection captures only linear word–trait associations

Suggested extensions:

• Incorporation of larger datasets (e.g., myPersonality corpus)
• Integration with sequence models (e.g., Bi-RNN with attention)
• Application to low-resource, cross-script languages (demonstrated on Chinese)

Applications encompass personality-adaptive dialogue systems, HR-screening and marketing, and cross-cultural user profiling. GlobalTrait’s methodology facilitates robust cross-linguistic transfer, aligning psycholinguistically relevant lexical items in the embedding space and enhancing both classification and regression of personality traits (Siddique et al., 2018).