Fig. 1. An example of a long-distance dependency problem in spoken language understanding. In this case, a word token “dec.” with local feature set (dotted line box) is ambiguous for determining the correct label(DEPART.MONTH or RETURN.MONTH).

Fig. 2. Approaches for long-distance dependency problem: (a) hidden vector state (He and Young, 2005); (b) previous slot’s context (Wang et al., 2005b); (c) tree-path and head word (Gildea and Jurafsky, 2002) and (d) trigger pair.

Fig. 3. Example of trigger features for air travel data.

Fig. 4. A distribution of hierarchical classes in air travel corpus.

Fig. 5. Feature templates.

Fig. 6. Learning curves for the air travel data set, averaged over 5-fold CV’s. Trigger model performs equivalently to the local model with two times more data: (a) text and (b) ASR.

Fig. 7. A precision-recall curve that summarizes the experimental results on the air-travel data set.

Fig. 8. Top-20 trigger samples extracted by the various selection methods.

Fig. 9. Learning curves for the FI vs MI based trigger selection methods, averaged over 5-fold CV’s. Each error bar represents a standard error of the results for 5-fold CV’s.

Fig. 10. Effect of gain threshold values.

Examples of trigger features

Result of local and non-local models (Text)

F₁ scores are averaged over 5-fold CV’s with standard errors.

Result of local and non-local models (ASR)

F₁ scores are averaged over 5-fold CV’s with standard errors.

Comparison of performance on air travel data (1178 test utterances used in He and Young (2005))

The trigger features outperform others. (The first two lines of this table are adapted from He and Young (2005).)

Result of the trigger selection methods

Comparison of variations in the induction algorithm

Columns are induction and total training time (h:m:s), number of trigger and local features, and F₁-score on test data.

Statistics of two spoken dialog data

SLU results of two spoken dialog data