tucSage: Grammar Rule Induction for Spoken Dialogue Systems via Probabilistic Candidate Selection

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 668–672, Dublin, Ireland, August 23-24, 2014. tucSage: Grammar Rule Induction for Spoken Dialogue Systems via Probabilistic Candidate Selection Arodami Chorianopoulou†, Georgia Athanasopoulou†, Elias Iosif‡ †, Ioannis Klasinas†, Alexandros Potamianos∗ † School of ECE, Technical University of Crete, Chania 73100, Greece ∗ School of ECE, National Technical University of Athens, Zografou 15780, Greece ‡ “Athena” Research Center, Marousi 15125, Greece {achorianopoulou,gathanasopoulou,iklasinas}@isc.tuc.gr iosife@telecom.tuc.gr, apotam@gmail.com Abstract We describe the grammar induction sys- tem for Spoken Dialogue Systems (SDS) submitted to SemEval’14: Task 2. A sta- tistical model is trained with a rich fea- ture set and used for the selection of can- didate rule fragments. Posterior probabil- ities produced by the fragment selection model are fused with estimates of phrase- level similarity based on lexical and con- textual information. Domain and language portability are among the advantages of the proposed system that was experimen- tally validated for three thematically dif- ferent domains in two languages. 1 Introduction A critical task for Spoken Dialogue Systems (SDS) is the understanding of the transcribed user input, that utilizes an underlying domain grammar. An obstacle to the rapid deployment of SDS to new domains and languages is the time-consuming development of grammars that require human ex- pertise. Machine-assisted grammar induction has been an open research area for decades (K. Lari and S. Young, 1990; S. F. Chen, 1995) aiming to lower this barrier. Induction algorithms can be broadly distinguished into resource-based, e.g., (A. Ranta, 2004), and data-driven, e.g., (H. Meng and K.-C. Siu, 2002). The main drawback of the resource-based paradigm is the requirement of pre-existing knowledge bases. This is addressed by the data-driven paradigm that relies (mostly) on plain corpora. SDS grammars are built by uti- lizing low- and high-level rules. Low-level rules This work is licenced under a Creative Commons Attri- bution 4.0 International License. Page numbers and pro- ceedings footer are added by the organizers. License de- tails: http://creativecommons.org/licenses/ by/4.0/ are similar to gazetteers consisting of terminal en- tries, e.g., list of city names. High-level rules can be lexicalized as textual fragments (or chunks), which are semantically defined on top of low- level rules, e.g., ‘depart from <City>’. The data-driven induction of low-level rules is a well-researched area enabled by various technolo- gies including web harvesting for corpora creation (Klasinas et al., 2013), term extraction (K. Frantzi and S. Ananiadou, 1997), word-level similarity computation (Pargellis et al., 2004) and cluster- ing (E. Iosif and A. Potamianos, 2007). High-level rule induction is a less researched area that poses two main challenges: 1) the extraction and selec- tion of salient candidate fragments from a corpus that convey semantics relevant to the domain of in- terests and 2) the organization of such fragments (e.g., via clustering) according to their semantic similarity. Despite the recent interest on phrase (J. Mitchell and M. Lapata, 2010) and sentence simi- larity, each respective problem remains open. Next, our submission1 for the Se- mEval’14: Task2 is briefly described, which constitutes …

Content truncated to 500 words