Post Doc: Semantic Parsing Into Amr

**About the role**:
However, these results are still insufficient to consider using semantic analysis on languages other than English (when the quality measure is below 82%). Even in English, the good scores mask qualitative flaws that remain prohibitive for real use. Moreover, whatever technology we have been able to implement, we have always found that the slowness of certain tasks, such as question and answer, is incompatible with the real-time requirements.

**Scientific goals**

The objective of the post doc is twofold:

- To study existing AMR parsing tools and the technologies and architectures they have been implement in (Seq2Seq, Transformers, Perceivers, etc) in order to establish a policy for improving this parsing.
- To improve an existing tool **or** create a new, more efficient tool to:

- Exceed the state of the art in terms of Smatch/f-measure, especially for languages other than English.
- Reduce the size of the resources needed (learning time, RAM/GPU size).
- Increase the speed of inference (sentences/seconds).
- Eventually build a multi-task model that integrates other tasks for which we already have models (NER/NEL, coreference resolution) with AMR analysis.

This work will therefore focus mainly on the algorithmic aspects but will also require work on the data.

**The main obstacles to be solved are**
- Obtaining an identical or close semantic graph for the same sentence expressed in a different way (whatever the syntax of the analysed sentence).
- Obtaining an identical or close semantic graph for the same sentence expressed in different languages.
- Limiting the size of the models despite the possible need to embed a language model, possibly multilingual.

**About you**:
Technical and scientific competences required:

- PhD in computer science including experience in Deep Learning
- Knowledge in NLP, notably in parsing
- Knowledge in (formal) semantics
- Knowledge of Python and Pytorch
- Knowledge in Julia would be appreciated

**Additional information**:
The expected results are
- A state of the art on AMR parsing (T0 + 2 months)
- A work plan for improving the state of the art (T0+3)
- A first monolingual tool for English (T0+6)
- A first multilingual tool (T0+12)

It is possible to submit the publication of your work to the NLP communities/conferences.

There is the possibility of an extension of the postdoc, e.g. in order to build a second version of your multilingual tool.

**Department**:
Orange Innovation brings together the research and innovation activities and expertise of the Group's entities and countries. We work every day to ensure that Orange is recognized as an innovative operator by its customers and we create value for the Group and the Brand in each of our projects. With 740 researchers, thousands of marketers, developers, designers and data analysts, it is the expertise of our 6000 employees that fuels this ambition every day.
Orange Innovation anticipates technological breakthroughs and supports the Group's countries and entities in making the best technological choices to meet the needs of our consumer and business customers.

Within the Innovation Division, whose ambition is to take Orange’s innovation further and to reinforce its technological leadership, the team in which the post-doc will work is in charge of research and development in the field of Natural Language Processing (NLP). The research work covers a wide range of NLP tasks (semantic analysis, information extraction, document structuring, knowledge management, etc.).

**Contract**:
Post Doc