An Intra and Inter-Topic Evaluation and Cleansing Method

National audienceTopic modeling is a growing research field and novel ways of interpreting and evaluating results are necessary. We propose a method for evaluating and improving the performance of topic models generating algorithms relying on WordNet data. We first propose a measure for determining a topic model's fitness factoring in its broadness and redundancy. Then, for each individual topic, the amount of relevant information it provides, along with its most important words and related concepts are determined by defining a cohesion function based on the topic's projection on WordNet concepts. The model as a whole is improved by eliminating each topic's outliers with respect to the ontology projection. We define a inter topic ontology based distance and we further use it to investigate the impact of removing redundant topics from a model with regard to the overlap between topics' ontological projections. Clustering similar topics into conceptually cohesive groups is tried as an alternative to pruning less relevant topics. Results show that evaluating and improving statistical models with WordNet is a promising research track that leads to more coherent topic models