Parsing enhancing of the conversational module for a service robot

The goal of the project is to improve the syntactic parser of the Golem service robot¹⁾²⁾. The parsing module takes its imput from the Automatic Speech Recognition (ASR), which produces a text line to be parsed. The output of the parser is a SitLog command which triggers an action in the robot.

1. Generate 100 sentences with the Robocup command generator and evaluate the accuracy of the generated command manually.
2. TODO
   • Gather 1000 significant current linguistic expressions and corresponding SitLog commands from Golem for the Gold Standart test
   • Implement the random command generator for the Dynamic test
   • Implement both test.

The goal of this module is to train a model with linguistic and Sitlog commands in order to generate Golem's speech act rules in prolog.

We will implement the deep learning model used in <ref>Amin Sleimi and Claire Gardent. Generating Paraphrases from DBPedia using Deep Learning. WebNLG 2016, The Second International Workshop on NLG and the Semantic Web, September 6 2016, Edinburgh (Scotland).</ref> in order to generate speech act Prolog rules for Golem.

1. Create an evaluation testbed for the parser based on (Doostdar et al., 20907)³⁾ but mostly on the Robocup command parser generator.
2. Train a machine learning model to generate DGC⁴⁾, mostly based on Claires article.

1. Transform the DGC-based (Definite Clause Grammar) parser into a CCG (Combinatorial Categorial Grammar)⁵⁾⁶⁾ based parser
2. Evaluate Google's Parsey MacParseface⁷⁾ with the service robot oriented parsing evaluation testbed.

1. Integrate Parsey MacParceface

• Luis Pineda (IIMAS/UNAM)
• Ivette Vélez (IIMAS/UNAM)
• Jorge García Flores (LIPN/UP13)

• May 23, 2017: Ivette and Jorge