Bilateral Contracts and Grants with Industry
Bilateral Contracts and Grants with Industry

Section: Research Program

Automatic Context-augmented Linguistic Analysis

This first research strand is centred around NLP technologies and some of their applications in Artificial Intelligence (AI). Core NLP tasks such as part-of-speech tagging, syntactic and semantic parsing is improved by integrating new approaches, such as (deep) neural networks, whenever relevant, while preserving and taking advantage of our expertise on symbolic and statistical system: hybridisation not only couples symbolic and statistical approaches, but neural approaches as well. AI applications are twofold, notwithstanding the impact of language variation (see the next strand): (i) information and knowledge extraction, whatever the type of input text (from financial documents to ancient, historical texts and from Twitter data to Wikipedia) and (ii) chatbots and natural language generation. In many cases, our work on these AI applications is carried out in collaboration with industrial partners (for which cf. Section 7.1). The specificities and issues caused by language variation (a text in Old French, a contemporary financial document and tweets with a non-canonical spelling cannot be processed in the same way) are addressed in the next research strand.

Context-augmented processing of natural language at all levels: morphology, syntax, semantics

Our expertise in NLP is the outcome of more than 10 years in developing new models of analysis and accurate techniques for the full processing of any kind of language input since the early days of the Atoll project-team and the rise of linguistically informed data-driven models as put forward within the Alpage project-team.

Traditionally, a full natural language process (NLP) chain is organised as a pipeline where each stage of analysis represents a traditional linguistic field (in a structuralism view) from morphological analysis to purely semantic representations. The problem is that this architecture is vulnerable to error propagation and very domain sensitive: each of these stage must be compatible at the lexical and structure levels they provide. We arguably built the best performing NLP chain for French [63], [97] and one of the best for robust multilingual parsing as shown by our results in various shared tasks over the years [93], [90], [96], [21]. So we pursue our efforts on each of our components we developed: tokenisers (e.g. SxPipe), part-of-speech taggers (e.g. MElt), constituency parsers and dependency parsers (e.g. FRMG, DyALog-SR) as well as our recent neural semantic graph parsers [90].

In particular, we continue to explore the hybridisation of symbolic and statistical approaches, and extend it to neural approaches, as initiated in the context of our participation to the CoNLL 2017 multilingual parsing shared task (We ranked 3 for UPOS tagging and 6 for dependency parsing out of 33 participants.) and to Extrinsic Parsing Evaluation Shared Task (Semantic graph parsing, evaluated on biomedical data, speech and opinion. We ranked 1 in a joint effort with the Stanford NLP team).

Fundamentally, we want to build tools that are less sensitive to variation, more easily configurable, and self-adapting. Our short-term goals is to explore techniques such as multi-task learning (cf. already  [95]) to propose a joint model of tokenisation, normalisation, morphological analysis and syntactic analysis. We also explore adversarial learning, considering the drastic variation we face in parsing user-generated content and processing historical texts, both seen as noisy input that needs to be handled at training and decoding time.

While those points are fundamental, therefore necessary, if we want to build the next generation of NLP tools, we need to push the envelop even further by tackling the biggest current challenge in NLP: handling the context within which a speech act is taking place.

There is indeed a strong tendency in NLP to assume that each sentence is independent from its siblings sentences as well as its context of enunciation, with the obvious objective to simplify models and reduce the complexity of predictions. While this practice is already questionable when processing full-length edited documents, it becomes clearly problematic when dealing with short sentences that are noisy, full of ellipses and external references, as commonly found in User-Generated Content (UGC).

A more expressive and context-aware structural representation of a linguistic production is required to accurately model UGC. Let us consider for instance the case for Syntax-based Machine Translation of social media content, as is carried out by the ALMAnaCH-led ANR project Parsiti (PI: DS). A Facebook post may be part of a discussion thread, which may include links to external content. Such information is required for a complete representation of the post's context, and in turn its accurate machine translation. Even for the presumably simpler task of POS tagging of dialogue sequences, the addition of context-based features (namely information about the speaker and dialogue moves) was beneficial [72]. In the case of UGC, working across sentence boundaries was explored for instance, with limited success, by [62] for document-wise parsing and by [82] for POS tagging.

Taking the context into account requires new inference methods able to share information between sentences as well as new learning methods capable of finding out which information is to be made available, and where. Integrating contextual information at all steps of an NLP pipeline is among the main research questions addressed in this research strand. In the short term, we focus on morphological and syntactic disambiguation within close-world scenarios, as found in video games and domain-specific UGC. In the long term, we investigate the integration of linguistically motivated semantic information into joint learning models.

From a more general perspective, contexts may take many forms and require imagination to discern them, get useful data sets, and find ways to exploit them. A context may be a question associated with an answer, a rating associated with a comment (as provided by many web services), a thread of discussions (e-mails, social media, digital assistants, chatbots—on which see below–), but also meta data about some situation (such as discussions between gamers in relation with the state of the game) or multiple points of views (pictures and captions, movies and subtitles). Even if the relationship between a language production and its context is imprecise and indirect, it is still a valuable source of information, notwithstanding the need for less supervised machine learning techniques (cf. the use of LSTM neural networks by Google to automatically suggest replies to emails).

Information and knowledge extraction

The use of local contexts as discussed above is a new and promising approach. However, a more traditional notion of global context or world knowledge remains an open question and still raises difficult issues. Indeed, many aspects of language such as ambiguities and ellipsis can only be handled using world knowledge. Linked Open Data (LODs) such as DBpedia, WordNet, BabelNet, or Framebase provide such knowledge and we plan to exploit them.

However, each specialised domain (economy, law, medicine…) exhibits its own set of concepts with associated terms. This is also true of communities (e.g. on social media), and it is even possible to find communities discussing the same topics (e.g. immigration) with very distinct vocabularies. Global LODs weakly related to language may be too general and not sufficient for a specific language variant. Following and extending previous work in ALPAGE, we put an emphasis on information acquisition from corpora, including error mining techniques in parsed corpora (to detect specific usages of a word that are missing in existing resources), terminology extraction, and word clustering.

Word clustering is of specific importance. It relies on the distributional hypothesis initially formulated by Harris, which states that words occurring in similar contexts tend to be semantically close. The latest developments of these ideas (with word2vec or GloVe) have led to the embedding of words (through vectors) in low-dimensional semantic spaces. In particular, words that are typical of several communities (see above) can be embedded in a same semantic space in order to establish mappings between them. It is also possible in such spaces to study static configurations and vector shifts with respect to variables such as time, using topological theories (such as pretopology), for instance to explore shifts in meaning over time (cf. the ANR project Profiterole concerning ancient French texts) or between communities (cf. the ANR project SoSweet). It is also worth mentioning on-going work (in computational semantics) whose goal is to combine word embeddings to embed expressions, sentences, paragraphs or even documents into semantic spaces, e.g. to explore the similarity of documents at various time periods.

Besides general knowledge about a domain, it is important to detect and keep trace of more specific pieces of information when processing a document and maintaining a context, especially about (recurring) Named Entities (persons, organisations, locations...) —something that is the focus of future work in collaboration with Patrice Lopez on named entity detection in scientific texts. Through the co-supervision of a PhD funded by the LabEx EFL (see below), we are also involved in pronominal coreference resolution (finding the referent of pronouns). Finally, we plan to continue working on deeper syntactic representations (as initiated with the Deep Sequoia Treebank), thus paving the way towards deeper semantic representations. Such information is instrumental when looking for more precise and complete information about who does what, to whom, when and where in a document. These lines of research are motivated by the need to extract useful contextual information, but it is also worth noting their strong potential in industrial applications.

Chatbots and text generation

Chabots have existed for years (Eliza, Loebner prize). However, they are now becoming the focus of many concrete industrial developments, with the emergence of operational conversational agents and digital assistants (such as Siri). The current approaches mostly rely on the design of scenarios associated with very partial analysis of the requests to fill expected slots and to generate canned answers.

The next generations of such systems will rely on a deeper understanding of the requests, being able to adapt to the specificities of the users, and providing less formatted answers. We believe that chatbots are an interesting and challenging playground to deploy our expertise on knowledge acquisition (to identify concepts and formulations), information extraction based on deeper syntactic representations, context-sensitive analysis (using the thread of exchanges and profile information but also external data sources), and robustness (depending on the possible users' styles).

However, this domain of application also requires working on text generation, starting with simple canned answers and progressively moving to more sophisticated and diverse ones. This work is directly related to another line of research regarding computer-aided text simplification, for which see section 3.3.4.