Limitations in learning an interpreted language with recurrent models
- Speaker
- Denis Paperno (Loria)
- Date
- Room
- Salle 206, bâtiment IMAG
I report work in progress on learning simplified interpreted languages by means of recurrent models. The data is constructed to reflect core properties of natural language as modeled in formal syntax and semantics: recursive syntactic structure and compositionality. Preliminary results suggest that LSTM networks do generalise to compositional interpretation, albeit only in the most favorable learning setting, with a well-paced curriculum, extensive training data, and left-to-right (but not right-to-left) composition.
Bio : While I have experience in different subfields of language science such as field linguistics, language typology, and formal semantics, my current work mainly focuses on computational semantic representations for natural language, including word, phrase, and sentence embeddings. My work has included proposing new models such as the Practical Lexical Function model for syntax-driven vector compositionality or the Boolean Distributional Semantic Model for entailment detection, as well as analyzing existing models and evaluating them on new tasks.
I hold an undergraduate degree in Linguistics from Moscow State University and a PhD from the University of California, Los Angeles. After finishing my thesis, I was a postdoc at Marco Baroni’s COMPOSES group at the University of Trento. Since 2016, I am a researcher (CR CNRS) at the Lorraine laboratory of computer science and its applications (Loria).
