Ranking and Selecting Multi-Hop Knowledge Paths to Better Predict Human Needs

Debjit Paul, Anette Frank

Introduction

Standard sentiment analysis excels at identifying what an individual feels, yet it frequently fails to explain the underlying why. In narrative contexts, understanding a character's emotional response requires "Naive Psychology"—the ability to infer internal motivations and needs based on situational cues. For instance, if a character breaks a leg, a human intuitively links the negative sentiment to the need for "health." However, computational models often struggle with this because there is no direct linguistic link between "leg" and "health." Bridging this gap requires leveraging implicit commonsense knowledge through multi-hop reasoning.

Bridging the Gap with Multi-Hop Knowledge

To address this, researchers from Heidelberg University developed a knowledge-enhanced neural architecture. The core innovation lies in extracting and ranking multi-hop reasoning paths from ConceptNet to connect textual concepts with psychological needs.

Figure 3: Example of a narrative story and the multi-hop paths connecting concepts like 'gold medal' to the human need 'status'.

As shown in the figure above, single-step connections are often insufficient. Multi-hop paths allow the model to traverse intermediate concepts (e.g., gold medal → medal → jewelry → status) to find the logical link. To manage the noise inherent in large knowledge graphs, the authors applied graph centrality measures—such as Closeness Centrality and Personalized PageRank—to select the most relevant paths.

Methods and Architecture

The proposed model consists of three primary components:

  1. Text Encoder: A Bi-LSTM with self-attention using ELMo embeddings processes the story and its preceding context.
  2. Knowledge Encoder: A separate Bi-LSTM encodes the selected multi-hop paths.
  3. Gated Integration: A gated attention mechanism dynamically weights the contribution of text versus external knowledge for the final prediction.

The study utilized the Modeling Naive Psychology of Characters in Simple Commonsense Stories (MNPCSCS) dataset, which labels character motivations using Maslow’s Hierarchy of Needs and Reiss’s Basic Motives.

Key Results

The integration of ranked knowledge paths led to a new state-of-the-art performance on the MNPCSCS dataset.

Figure 5: Performance comparison (Precision, Recall, F1) of the proposed model (BM+K) against various baselines on Reiss and Maslow categories.

As detailed in the results table, the model achieved a significant improvement over traditional baselines. For Maslow’s categories, the base model (BM) alone provided a 21.02 percentage point improvement in F1 score over previous models like REN and NPN. The full model (BM + K with ELMo) reached an F1 of 39.44 for the fine-grained Reiss categories, demonstrating the power of combining contextual embeddings with structured commonsense reasoning.

Insights and Interpretability

Beyond raw performance, the model offers improved interpretability. By applying attention over specific knowledge paths, the system can "explain" its reasoning. For example, to predict the need for "rest," the model might highlight a path such as exhausting → fatigue → rest. Human validation confirmed that the paths selected by the algorithm were relevant in 34% of cases, a notable achievement in automated reasoning.

Conclusion and Future Work

While the model sets a new benchmark, challenges remain. Subjectivity in psychological labeling and the inherent sparsity of structured knowledge bases like ConceptNet suggest that future iterations could benefit from incorporating unstructured text corpora. Nevertheless, this work marks a significant step toward machines that can truly understand the "why" behind human emotion.

Summary updated

BibTeX
@inproceedings{paul2019ranking,
  author = {Debjit Paul and Anette Frank},
  title = {{Ranking and Selecting Multi-Hop Knowledge Paths to Better Predict Human Needs}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/n19-1368},
}
Figure