Leveraging Recursive Processing for Neural-Symbolic Affect-Target Associations

Alexander Sutherland, Sven Magg, Stefan Wermter

Abstract

Explaining the outcome of deep learning decisions based on affect is challenging but necessary if we expect social companion robots to interact with users on an emotional level. In this paper, we present a commonsense approach that utilizes an interpretable hybrid neural-symbolic system to associate extracted targets, noun chunks determined to be associated with the expressed emotion, with affective labels from a natural language expression. We leverage a pre-trained neural network that is well adapted to tree and sub-tree processing, the Dependency Tree-LSTM, to learn the affect labels of dynamic targets, determined through symbolic rules, in natural language. We find that making use of the unique properties of the recursive network provides higher accuracy and interpretability when compared to other unstructured and sequential methods for determining target-affect associations in an aspect-based sentiment analysis task.

Introduction

In the field of Affective Computing, most systems focus on "global" sentiment—determining if a whole sentence is positive or negative. However, for a social companion robot, this isn't enough. If a user says, "The food was excellent, but the service was disappointing," a global "neutral" label is unhelpful. The robot needs to know that the food is the target of praise and the service is the target of complaint.

The challenge lies in the complexity of natural language. Standard deep learning models, such as sequential LSTMs, often struggle to map emotions to specific entities (targets) because they process text as a flat sequence. This research addresses the need for interpretability and granularity: creating a system that can reason about how specific words (like "not" or "excellent") influence the sentiment of specific targets within a complex sentence.

Methodology: Neural-Symbolic Hybrid Approach

The proposed system utilizes a neural-symbolic hybrid approach to link emotions to specific objects:

  1. Symbolic Target Identification: Using the SpaCy NLP library, the system generates a dependency parse tree for a sentence, identifying all verbs and noun chunks.
  2. Recursive Neural Processing: A Dependency Tree-LSTM processes the sentence according to its grammatical structure. Unlike a standard LSTM that reads left-to-right, the Tree-LSTM calculates a sentiment score for every node (sub-phrase) in the tree.
  3. Tree Deconstruction: Symbolic rules split the main tree into sub-trees rooted at verbs.
  4. Sentiment Association: The sentiment identified at the root of a verb's sub-tree is "inherited" by the noun chunks (targets) within that sub-tree.
Figure 4: The process of deconstructing a complex dependency tree into verb-rooted sub-trees to isolate target-affect associations.

This approach allows researchers to visualize the "train of thought"—for example, seeing how the word "not" flips the sentiment of "outstanding" before it reaches the target "food."

Evaluation and Key Results

The model was trained on the Stanford Sentiment Treebank (SST), which provides sentiment labels for every sub-phrase in a parse tree. Evaluation was conducted using the SemEval 2016 Aspect-Based Sentiment Analysis (ABSA) restaurant dataset to test real-world domain performance.

The researchers compared the Tree-LSTM approach against a standard Bidirectional LSTM (BLSTM) and a Bag-of-Words (BoW) Logistic Regression classifier.

Figure 6: Performance comparison on the SemEval 2016 Restaurant dataset. The Dependency Tree-LSTM significantly outperforms sequential and unstructured models.

While the Logistic Regression model performed slightly better on the original SST training data (85.5% vs 84.9%), the Tree-LSTM significantly outperformed all others on the SemEval restaurant data, achieving an accuracy of 0.669. This suggests that the Tree-LSTM’s structural bias makes it much more robust when generalizing from general sentiment training to specific target-labeling in new, unseen domains.

Discussion and Future Work

The "aha moment" of this work is that recursive models (Tree-LSTMs) are inherently better at transferring learned features to substrings than sequential models. While a standard LSTM might get "confused" by the distance between a subject and its adjective, the Tree-LSTM sees them as closely related branches.

However, limitations remain. The current symbolic rules rely heavily on verbs, making the system struggle with short, verbless exclamations like "Great food!" Additionally, the symbolic approach is currently very "inclusive," sometimes identifying irrelevant noun chunks as targets. Future work aims to refine target selection and integrate this system into live Human-Robot Interaction (HRI) scenarios, allowing robots to make dialogue decisions based on fine-grained emotional associations.

Summary updated

Abstract

Explaining the outcome of deep learning decisions based on affect is challenging but necessary if we expect social companion robots to interact with users on an emotional level. In this paper, we present a commonsense approach that utilizes an interpretable hybrid neural-symbolic system to associate extracted targets, noun chunks determined to be associated with the expressed emotion, with affective labels from a natural language expression. We leverage a pre-trained neural network that is well adapted to tree and sub-tree processing, the Dependency Tree-LSTM, to learn the affect labels of dynamic targets, determined through symbolic rules, in natural language. We find that making use of the unique properties of the recursive network provides higher accuracy and interpretability when compared to other unstructured and sequential methods for determining target-affect associations in an aspect-based sentiment analysis task.
BibTeX
@inproceedings{sutherland2019leveraging,
  author = {Alexander Sutherland and Sven Magg and Stefan Wermter},
  title = {{Leveraging Recursive Processing for Neural-Symbolic Affect-Target Associations}},
  booktitle = {IJCNN},
  year = {2019},
  url = {https://openreview.net/forum?id=AEZd2SKypO},
}
Figure