Project: Learning to Interact with Humans by Lifelong Interaction with Humans
| Acronym | LIHLITH (Reference Number: ANR-17-CHR2-0001) |
| Duration | 01/01/2018 - 31/12/2020 |
| Project Topic | The LIHLITH project is a fundamental pilot research project which introduces a new lifelong learning framework for the interaction of humans and machines on specific domains. A Lifelong Learning system learns different tasks sequentially, over time, getting better at solving future related tasks based on past experience. LIHLITH will focus on human-computer dialogue, where each dialogue experience is used by the system to learn to better interact, based on the success (or failure) of previous interactions. The key insight is that the dialogue will be designed to produce a reward, allowing the chatbot system to know whether the interaction was successful or not. The reward will be used to train the domain and dialogue management modules of the chatbot, improving the performance, and reducing the development cost, both on a single target domain but specially when moving to new domains. The research will be evaluated on publicly available benchmarks to allow comparison with other approaches in the state of the art. When possible, systems will participate in international comparative/competitive evaluations such as WOCHAT or SemEval. LIHLITH project will also develop and deliver evaluation protocols and benchmarks to allow public comparison and reproducibility based on crowdsourcing. The industrial partner will transfer the research into technology, applying the lessons learnt to the development of chatbots for customer support . LIHLITH will rely on recent advance in multiple research disciplines, including, natural language processing, knowledge induction, reinforcement learning, deep learning, and lifelong learning . |
|
Project Results (after finalisation) |
The main outcome of the project will be research advancing the state of the art in lifelong learning, dialogue systems and knowledge inference for question answering, as well as research in replicable evaluation of dialogue systems that improve as they interact with humans. We will now detail outcomes organized around the three main research areas. Research Area 1 , Lifelong Learning for Dialogue . Outcome 1.1: Method to produce dialogue manager modules, which will not only integrate the context of the interaction (previous utterances, tasks) but also the new knowledge acquired during the dialogues (implicit and explicit user feedbacks) using lifelong learning. These methods will improve over the state of the art in dialogue systems, as measured following the evaluation protocols and benchmarks produced in the project (see below). Research Area 2 : Lifelong Learning for Knowledge Induction and Question Answering . Outcome 2.1: Method to improve low-dimensional representations of terms, concepts and relations (learned based on the textual processing of large document collections which are linked to general ontologies adapted to the domain). These methods will improve results over the state of the art, as measured in downstream application in the dialogue system. In addition, the embeddings will be evaluated in question answering and similarity tasks [AGIR16]. Outcome 2.2: Method to improve question answering systems based on semantic textual similarity systems using lifelong learning. The improvement over the state of the art will be measured directly in questions answering, and indirectly in the dialogue system above. Research Area 3 : Evaluation of dialogue . Outcome 3.1: a benchmark on interactive question-dialogue, including background texts and ontologies, training dialogues annotated with user feedback and the necessary material to reproduce test results, including crowdsourcing templates, and a well-defined API for dialogue models. The benchmark will be designed to assess lifelong learning capabilities of dialogue systems, and will be used in a shared task for lifelong learning of dialogue systems in SemEval 2019. Outcome 3.2: an industrial chatbot with dialogue managing modules demonstrating the ability to learn from interactions. |
| Network | CHIST-ERA III |
| Call | CHIST-ERA Call 2017 |
Project partner
| Number | Name | Role | Country |
|---|---|---|---|
| 1 | University of the Basque Country | Coordinator | Spain |
| 2 | National University of Distance Education | Partner | Spain |
| 3 | Computer Science Laboratory for Mechanics and Engineering Sciences | Partner | France |
| 4 | SYNAPSE DEVELOPMENT | Partner | France |
| 5 | ZHAW | Partner | Switzerland |