Unsupervised Online Multitask Learning of Behavioral Sentence Embeddings

Shao-Yen Tseng, Brian Baucom, Panayiotis GeorgiouarXiv 2018 · arXiv

This paper, published by researchers at the University of Utah and USC, introduces a clever way to make unsupervised sentence embeddings more useful for specialized tasks like behavioral analysis and emotion recognition. By adding a "noisy" but automated secondary task during training, the authors guide the model toward learning representations that are more sensitive to human affect without requiring expensive manual labeling.

TL;DR

The authors propose an online multitask learning (MTL) framework that enhances unsupervised sentence embeddings by incorporating a simple, rule-based affective classification task during training. This approach significantly improves performance on downstream behavioral and emotional tasks, outperforming several state-of-the-art general-purpose embedding models.

What Problem Does This Paper Address?

Unsupervised representation learning is a powerhouse in NLP because it leverages vast amounts of unlabeled data (like movie subtitles or web text) to learn the "meaning" of sentences. However, these embeddings often suffer from two main issues:

  1. Unpredictability: The resulting vector space can be random or contain redundant information that isn't useful for specific downstream tasks.
  2. Domain Portability: Embeddings trained on generic datasets often fail to capture the nuances required for specialized fields, such as clinical psychology or behavioral science.

Usually, fixing this requires manual annotation—a slow and expensive process. The authors ask: Can we inject domain-specific "hints" into the unsupervised process without needing humans to label millions of sentences?

Methods

The core of the paper is a Bidirectional Sequence-to-Sequence (seq2seq) model. Standard seq2seq models learn by trying to predict the next sentence in a conversation (contextual learning).

The authors modify this by adding an Online Multitask Objective. As the model sees an input sentence, it performs two tasks simultaneously:

  1. Primary Task: Predict the next sentence (unsupervised).
  2. Multitask: Classify the current sentence as having "Positive" or "Negative" affect.
The model architecture showing the parallel unsupervised contextual learning and the online transfer learning heads.

The "magic" is in how they get the labels for the second task. Instead of human labels, they use a simple look-up table of affective words (e.g., "love" is positive, "hate" is negative). While this labeling method is "naive" and prone to errors, it provides a consistent behavioral signal that guides the encoder toward a more discriminative feature space.

Datasets

  • OpenSubtitles: Used for the initial unsupervised training. This massive corpus contains 30 million sentence pairs from movie and TV dialogue, providing a rich source of conversational language.
  • Couples Therapy Corpus: A specialized dataset of 134 real couples discussing marital issues. It includes 33 behavioral codes (like Acceptance, Blame, and Negativity) rated by experts.
  • IEMOCAP: A standard benchmark for emotion recognition containing scripted and improvised dyadic interactions labeled with emotions like happiness, sadness, and anger.

Key Insights and Ideas

  • Weak Supervision is Better Than No Supervision: Even though the online labels are generated by a simple word-list (see Table I below), they act as a "compass" for the neural network. This suggests that for specialized domains, a little bit of noisy, automated guidance is better than purely generic unsupervised learning.
  • Intermediate Layers Matter: The authors found that concatenating the internal states from all layers of the encoder (not just the final one) captured different levels of conceptual abstraction, leading to better behavioral features.
  • Domain Adaptation via MTL: The multitask approach allows the model to adapt to the "behavioral" domain during the pre-training phase, making the resulting embeddings much more potent for final tasks like rating a therapy session.
Examples of the simple word-list used to generate the "noisy" online labels.

Key Results

The proposed Online MTL model consistently outperformed the base unsupervised model and several popular general-purpose embeddings (like InferSent and GenSen) across multiple benchmarks.

  • Behavior Identification: In supervised classification (k-NN), the MTL approach achieved a mean accuracy of 78.77%, a 3.24% absolute improvement over the non-MTL version. It was particularly effective at identifying "Negativity" and "Positivity."
  • Emotion Recognition: On the IEMOCAP dataset, the MTL framework improved the weighted accuracy of unsupervised embeddings by 8.02% (a 14% relative improvement).
  • Consistency: As shown in Figure 2, the MTL-guided model (blue line) maintained higher classification accuracy throughout the training process compared to the purely unsupervised version (red line).
Standard error plot showing that the MTL objective consistently boosts classification accuracy across training iterations.

Limitations and Future Work

The authors acknowledge that their online labeling scheme is extremely simple. They suggest that using more sophisticated automated labeling (perhaps via more complex rule-based systems or sentiment analyzers) could further improve results.

Additionally, there is a potential risk of "over-specialization." If the multitask labels are too domain-specific, the embeddings might lose their general linguistic utility. The researchers leave the exploration of this trade-off—and the search for the optimal weighting between the unsupervised and multitask losses—for future work.

Summary updated

Abstract

Unsupervised learning has been an attractive method for easily deriving meaningful data representations from vast amounts of unlabeled data. These representations, or embeddings, often yield superior results in many tasks, whether used directly or as features in subsequent training stages. However, the quality of the embeddings is highly dependent on the assumed knowledge in the unlabeled data and how the system extracts information without supervision. Domain portability is also very limited in unsupervised learning, often requiring re-training on other in-domain corpora to achieve robustness. In this work we present a multitask paradigm for unsupervised contextual learning of behavioral interactions which addresses unsupervised domain adaption. We introduce an online multitask objective into unsupervised learning and show that sentence embeddings generated through this process increases performance of affective tasks.
BibTeX
@misc{tseng2018unsupervised,
  author = {Shao-Yen Tseng and Brian Baucom and Panayiotis Georgiou},
  title = {{Unsupervised Online Multitask Learning of Behavioral Sentence Embeddings}},
  year = {2018},
  eprint = {1807.06792},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.06792},
}
Figure