Michael Zock

Situation + background

I started my research career 1989 with the French National Research Centre (CNRS) at LIMSI (now LISN), an A.I. lab close to Paris.

2006, I moved to Southern France (Marseille) to join the NLP group of LIS (Aix-Marseille Université). Currently I am emeritus research director at CNRS and Honorary Professor of the Research Institute of Information and Language Processing (University of Wolverhampton, UK).

Having learned languages under various circumstances, and having observed language acquisition in different contexts, I was intrigued by the fact that most people succeed so well in a natural setting, while they often miserably fail in school. To get a better understanding of the causes of this problem, I started to look at language learning from a psycholinguistic and computational linguistic (simulation) perspective.

My Ph.D. dissertation was devoted to speaking, a skill that must be learned. Yet, learning to speak and producing language in real-time are two daunting tasks. One must not only acquire a huge amount of knowledge (vocabulary, grammar), but also be able to accomplish rapidly and quasi-simultaneously the following tasks: message planning, formulation, and articulation. What makes this process even more demanding is the fact that the needed operations must be carried out under severe time and space constraints. Speaking is fast and short-term memory is limited.

Teaching the skills of communication (speaking and writing) is one of the missions of school. Alas, schools don't always succeed very well, and there are various reasons for this. Schools tend to emphasize rules and mechanical repetition, leaving aside the need to integrate rules into a process allowing the stepwise transformation of a conceptual input (message) into its corresponding form (sentence). In sum, they do not show how to convert 'declarative knowledge' (principles, rules) into 'procedural knowledge' (the way of using rules). Hence students may know a lot about language without necessarily knowing the language. Not having acquired the necessary operations (process), they don't know how to convert some input (message) into the corresponding output (sentence). It is also noteworthy that what has been learned (product) does not always correspond to what has been taught, an intuition that was confirmed by the eye track data collected during my Ph.D. ("From knowing-what to knowing-how: strategies in language production", Paris, 1980). The data also confirmed that certain methods are inadequate, making learning unnecessarily hard, if not impossible. Yet schools are meant to support learning rather than to prevent us from doing so.

Realizing all these facts, I felt that we needed to change our perspective, and, studying language production from a psychological point of view and trying to emulate the process by computer seemed to be a step in the right direction. Concerning speaking and writing here are some of the questions I keep asking myself:

1. When do we process what?
2. What are the inputs and outputs?
3. What are the specific operations and constraints?
4. What should an authoring workbench, i.e., eco-system for writers, look like?

Motivations + approach.

Speaking and writing are resource-intensive processes whose success depends not only on knowledge but also on our momentary ability or skill to access and use it (synthesis). Yet these three conditions are not always met. Hence, our success in producing language lies somewhere in between two extremes: full access to the needed resources, or more or less limited access, yielding sub-optimal performance revealed by gaps, errors, disfluencies, etc.

This being so it makes sense to create authoring aids, i.e. assistive technologies (1, 2, 3, 4, 5), both for supporting the mother tongue or a foreign language. One may even wonder if this is not (also) one of the missions of computational linguistics. There are good reasons to believe that one should widen the scope and take better into account the human factor (needs and constraints of the human language processor), to build then the adequate tools (Interactive NLP, Human-Centered Computational Language Processing). However, to do so and to get these tools used in real-world (desktop, classroom ), true interdisciplinary work is needed, not only discretely, for not saying 'shamefully', in the backyard, but also more deliberately and visible, at the centre-stage. For additional information concerning the mindset of my approach, see 'Interactive Natural Language Generation' (INLG) here below.

Research summary

My research interests lie in communication, cognitive science and language production or language generation by and large. Starting from user needs and empirical findings (psycholinguistics, neurosciences) I try to build tools helping people to acquire the skill of speaking or writing in a foreign language and in their mother tongue. My current research deals with the following four topics:

1. Message-planning: creation of an interface (i.e. linguistically motivated ontology augmented with a graph generator), to support conceptual authoring (message composition);


2. Authoring Aids to support Writing at the various levels. For example, word access (linguistic level, see here below), or, outline planning (conceptual level). Authors have to perceive the possible links between their ideas and choose among them to produce well-organized thoughts, i.e. coherent discourse in line with their goals. Writing is thinking which also implies linking of hitherto unconnected thought;
   


3. Lexical access: words being a major gateway to the mind, we must learn how to store, use and access them. My goal is to help authors overcome the tip-of-the-tongue problem. To this end I rely on certain features of the human brain (distribution of knowledge, which may be accessible only in terms of fragments) and the mental lexicon (lexical items being linked via association; functional equivalence of 'navigation' and 'spreading activation'). Of course, words in books, computers, or the human brain are not the same, be it for their storage, (organization) access or representation (holistic vs. decomposed). Still, it does make sense to take inspiration from the mental lexicon to see whether, functionally speaking, we can achieve something equivalent. This means in our case that access or retrieval consists of navigating in a huge semantic network, i.e. within a map where all words are linked via associations. Search space can be kept small because users are generally able to activate at least some of the target's (lemon) direct neighbors (yellow, acid fruit). Often they are even able to name the relationship between the two (synonym, hypernym, ...), and if none of this holds, one can still try to organize the output, and present the direct neighbors of the input (the word coming to the user's mind when looking for a target) in the form of a categorial tree.


4. Acquisition of basic speaking skills: help students to become quickly fluent in a foreign language (both western and oriental) by learning the basic vocabulary and syntactic structures (learn words in context). The scope is the survival level, and the method used is to build an open, possibly self-extending, multilingual phrasebook augmented with a parametrizable exercise generator.

Here is a summary of my research expressed in a few words, or simply by a wordle.

Organisation of some workshops

1° Electronic dictionaries and the mental lexicon

1.1 CogALex (Cognitive Aspects of the Lexicon), workshop series co-located with COLING :

2024 (Torino, Italy), workshop;
2022 (Taipei, Taiwan), workshop; 2020 (Barcelona, Spain), workshop, shared task;
2016 (Osaka, Japan), workshop, shared task; 2014 (Dublin, Ireland), workshop, shared task;
2012 (Mumbai, India), 2010 (Bejing, China); 2008 (Manchester, UK), and
2004, a forerunner (Geneva, Switzerland).

1.2 ICCS (International Conference on Cognitive Science), Beijing, 2010.