μ-TBL lite

Abstract

This short paper describes-and in fact gives the complete source for-a tiny Prolog program implementing a flexible and fairly efficient Transformation-Based Learning (TBL) system:

Surface and Contextual Linguistic Cues in Dialog Act Classification: A Cognitive Science View

“…Because of their flexibility, machine learning approaches allow for multiple different cues. While word sequences, encoded through a bag‐of‐words approach, are dominant (e.g., Grau et al., 2004; Ribeiro et al., 2015; Sridhar et al., 2009; Surendran & Levow, 2006), syntactic cues (Di Eugenio et al., 2010; Lager & Zinovjeva, 1999; Novielli & Strapparava, 2009; Verbree et al., 2006) and semantic cues, through a latent semantic analysis (Di Eugenio et al., 2010; Novielli & Strapparava, 2009), have also been used. Some machine learning studies did not consider context (Ang et al., 2005; Grau et al., 2004; Novielli & Strapparava, 2009), but most others used contextual cues, such as surface information of previous utterances (Ribeiro et al., 2015; Sridhar et al., 2009), cues on the speakers of the utterances (Di Eugenio et al., 2010; Lager & Zinovjeva, 1999; Sridhar et al., 2009), and cues related to the organization of the discourse, through encoding turns with a hierarchical structure, such as subdialogs (Di Eugenio et al., 2010).…”

“…In both these studies, the models were evaluated on the Map Task corpus where the instruction giver and instruction follower might use words differently. In (other) machine learning and deep learning approaches, speaker cues have typically been encoded using a feature that tells the model who the speaker is (Bothe, Weber et al., 2018 ; Cerisara et al., 2018; Di Eugenio et al., 2010) or whether the previous utterance is uttered by the same or different speaker from the current utterance (e.g., Lager & Zinovjeva, 1999; Liu et al., 2017; Yano et al., 2021; Zhao & Kawahara, 2019). Interestingly, many deep learning models, despite encoding contextual cues such as the structure of turns directly into the model, do not take into account who uttered which utterance, and still achieve a good performance.…”

“…Syntactic cues solely capture the structure, and not the contents of an utterance, using, for example, PoS n ‐grams (Verbree et al., 2006) or the length of the utterance (Lager & Zinovjeva, 1999; Li & Wu, 2016; Verbree et al., 2006; Webb et al., 2005). Other syntactic cues have been used as well, such as labels representing elementary constituent trees (Sridhar et al., 2009) or labels representing the general syntactic structure of the whole utterance (Novielli & Strapparava, 2009).…”

“…Especially in earlier studies, the dialog acts of previous utterances were used as contextual cues in the prediction of the current dialog acts. While some studies have used the true, that is, gold‐standard, dialog acts of previous utterances in their model (Lager & Zinovjeva, 1999; Ribeiro et al., 2015; Sridhar et al., 2009; Verbree et al., 2006), other studies have used the dialog acts predicted by the model itself to serve as information for the prediction of the next utterance (e.g., Ji & Bilmes, 2006; Stolcke et al., 2000; Surendran & Levow, 2006; Tran et al., 2017a).…”

“…Because of their flexibility, the range of models that use a machine learning approach is more varied (Table 1, Studies 6–18). Machine learning models have made use of many different learning algorithms, including naive Bayes (Grau et al., 2004), logistic regression (Ang et al., 2005; Sridhar et al., 2009), support vector machines (Ribeiro et al., 2015; Surendran & Levow, 2006), decision trees (Verbree et al., 2006), Bayesian networks (Brychcín & Král, 2017; Ji & Bilmes, 2006), latent semantic analysis (Di Eugenio et al., 2010; Novielli & Strapparava, 2009), and transformation‐based learning (Lager & Zinovjeva, 1999).…”