The internship will take place in the framework of the PARSEME-FR project, which involves several NLP teams in France. The aim is to boost applications in Natural Language Processing (NLP), by focusing on one of their major challenges: multiword expressions (MWEs).
MWEs are groups of words which exhibit unpredicted properties (Baldwin & Kim, 2010). Most prominently, their meaning does not straightforwardly derive from the meanings of their components. For instance, faire ‘make/do’ and valoir ‘be worth sth’ are verbs, while faire-valoir ‘a stooge, a person who is used by somebody to do things that are unpleasant or dishonest’. Similarly, the meaning of casser sa pipe (literally to break one’s pipe) ‘to die’ cannot be straightforwardly deduced from the meanings of the individual components. Additionally, MWEs exhibit unpredicted morpho-syntactic and lexical constraints. For instance, replacing the verb in lancer un appel (lit. to throw a call) ‘to issue a call’ by a synonym yields an invalid expression *jeter un appel ‘to throw a call’. Doing alike in casser sa pipe ‘to die’ imposes a literal reading of the resulting expression: briser sa pipe ‘to break one’s pipe’.
One of the main aims of MWE-oriented NLP research is to model such expressions so as to optimize their automatic processing (for instance, to avoid their literal translation in machine translation systems). Two major MWE-related NLP tasks include MWE discovery and MWE identification. In the former, the input consists in large quantities of raw texts and the output is a list of potential MWEs. In the latter, and identifier takes a text on input and automatically annotates (points at) the occurrences of MWEs in it. MWE identification is a pre-requisite for downstream applications such as machine translation (which may want to treat MWEs with dedicated procedures).
Automatic identification of MWEs in 19 languages was addressed by the PARSEME shared task1 (Ramisch et al., 20182018), in which the BdTln team participated with the VarIDE system (Pasquer et al., 2018a). The results of the shared task show that identifying unseen MWEs (i.e. those MWEs which do not occur in the training data) is particularly challenging. Thus, identification should, ideally, exploit not only annotated corpora but also MWE lexicons and MWE discovery methods.
This internship is dedicated to discovering how MWE discovery could benefit from the previously seen data, rather than be performed from scratch. The hypothesis to be tested is that new (unseen) MWEs of certain types can be discovered due to their semantic similarity with known (previously seen) MWEs. For instance, knowing that haute température ‘high temperature’ is a known MWE, and replacing its components by synonyms and antonyms, we obtain semantically close but previously unseen MWEs such as température élevée ‘high temperature’ or basse/moyenne température ‘low/middle temperature’ (Savary & Jacquemin, 2003). Thus, new MWEs might be discovered by examining lexical substitution performed within known MWEs. The challenge is to discover to which degree lexical substitution yields valid MWEs, rather than spurious MWE candidates. For instance, a known MWE prendre un bain ‘to take a bath’ allows us to discover prendre une douche ‘to take a shower’. But the use of other semantically related but more distant words leads to invalid or literal expressions, such as prendre une baignoire ‘to take a bathtub’ or prendre un WC ‘to take a WC’.
To perform lexical substitution, a model of semantic similarity of words and expressions is needed. Previous work exploited semantic resources such as WordNet. In this internship, we focus on the domain of distributional semantics, which is based on the hypothesis that words having a similar meaning occur in similar contexts. Recent developments in distributional semantics include the construction of word embeddings, i.e. mappings from words or expressions to low-dimensional vectors of real numbers, which are expected to represent co-occurrence contexts of these words/expressions in a compact way. Thus, an embedding of a word/expression can be considered an abstract representation of its meaning.
The objectives of this internship are to exploit word embeddings for discovery of new MWEs based on their semantic proximity to the previously seen MWEs, contained in a lexicon or in an annotated corpus (resources of both types belong to the outcomes of the PARSEME-FR project). The discovery should lead to (semi-)automatic enrichment of these initial resources. Two stages are to be considered:
Possible extensions of the objectives:
Send your CV and a cover letter to:
——————————