A Neural Network Model With Feature Selection for Korean Speech Act Classification

Abstract

A speech act is a linguistic action intended by a speaker. Speech act classification is an essential part of a dialogue understanding system because the speech act of an utterance is closely tied with the user's intention in the utterance. We propose a neural network model for Korean speech act classification. In addition, we propose a method that extracts morphological features from surface utterances and selects effective ones among the morphological features. Using the feature selection method, the proposed neural network can partially increase precision and decrease training time. In the experiment, the proposed neural network showed better results than other models using comparatively high-level linguistic features. Based on the experimental result, we believe that the proposed neural network model is suitable for real field applications because it is easy to expand the neural network model into other domains. Moreover, we found that neural networks can be useful in speech act classification if we can convert surface sentences into vectors with fixed dimensions by using an effective feature selection method.

An Integrated Neural Network Model for Domain Action Determination in Goal-Oriented Dialogues

“…As shown in Table 3, the SADM showed better results than Kim et al [8] at all of the cut-off points. Moreover, the SADM, which used 100 dialogues as a training corpus, had similar precisions to Kim et al [8] by using 500-700 dialogues as a training corpus. The p-value against Kim et al [8] is measured as 0.000002.…”

“…In Table 3, the results of Kim et al [8] are similar to those of the SADM for users' utterances, except that they do not use the concept sequence features as input features. As shown in Table 3, the SADM showed better results than Kim et al [8] at all of the cut-off points.…”

“…Previous studies on the determination of users' domain actions are divided into the following two groups: one is the use of handcrafted rules such as recipes for plan inference and domain specific knowledge [2][3][4], and the other is the use of machine learning techniques based on a large annotated corpus [5][6][7][8][9][10][11]. The rule-based methods have shown good performances in small dialogue domains.…”

“…Many machine learning methods have recently been proposed for overcoming these problems. The previous machine learning methods have mainly dealt with only speech acts [5,6,8,10,11] or have independently dealt with speech acts and concept sequences [1,9]. However, a speech act and concept sequence pair should be simultaneously determined for precisely catching users' intentions because speech acts and concept sequences are tightly associated.…”

“…To resolve this ambiguity, a dialogue system should analyze the context of utterance (9). In this case, checking the previous utterance (i.e., utterance (8)) is necessary for choosing "response & timetable-update-date" as the domain action of utterance (9). In addition, it is difficult to determine the system's domain actions because they depend on dialogue history and domain knowledge.…”