Interpretation of the Intent Detection Problem as Dynamics in a Low-dimensional Space

TL;DR

The paper probes how RNNs solve SNIPS intent detection by viewing their dynamics as a nonlinear discrete-time system restricted to a low-dimensional manifold. Through reverse-engineering, it shows sentences traverse trajectories from an initial state toward fixed-point attractors, with final predictions driven by alignment between readout vectors and intent-specific state-region centroids. Key findings include a variable intrinsic dimensionality $d$ governed by $embed\_dim$ and $hidden\_dim$, robust clustering of final states by intent, and a fixed-point topology featuring attractors and multiple saddle points. These insights offer mechanistic interpretability for sequence models in intent detection and point to potential extensions to Transformer-based architectures and broader SLU tasks.

Abstract

Intent detection is a text classification task whose aim is to recognize and label the semantics behind a users query. It plays a critical role in various business applications. The output of the intent detection module strongly conditions the behavior of the whole system. This sequence analysis task is mainly tackled using deep learning techniques. Despite the widespread use of these techniques, the internal mechanisms used by networks to solve the problem are poorly understood. Recent lines of work have analyzed the computational mechanisms learned by RNNs from a dynamical systems perspective. In this work, we investigate how different RNN architectures solve the SNIPS intent detection problem. Sentences injected into trained networks can be interpreted as trajectories traversing a hidden state space. This space is constrained to a low-dimensional manifold whose dimensionality is related to the embedding and hidden layer sizes. To generate predictions, RNN steers the trajectories towards concrete regions, spatially aligned with the output layer matrix rows directions. Underlying the system dynamics, an unexpected fixed point topology has been identified with a limited number of attractors. Our results provide new insights into the inner workings of networks that solve the intent detection task.

Figures (9)

• Figure 1: a) Folded representation of RNNs emphasizing the idea of recurrence. b) Unfolded RNN with explicit reference to time flow.
• Figure 2: Sequence of hidden states associated to a tokenized input sentence fed into a RNN.
• Figure 3: Variance explained of visited states vs principal components of a RNN trained on the SNIPS dataset. a) GRU(e:16,h:16). b) GRU(e:10,h:10).
• Figure 4: State space dimensionality and accuracy of RNNs trained on the SNIPS dataset for different embed_dim and hidden_dim combinations. (a,b): Vanilla cell. (c,d): GRU cell. (e,f): LSTM cell
• Figure 5: State space top-2 and top-3 PCA projections of RNNs trained on the SNIPS dataset. a) GRU(e:16,h:16). b) GRU(e:16,h:16). c) Vanilla(e:20,h:20). d) LSTM(e:10,h:10). e) GRU(e:10,h:10).