This study proposes a method for constructing language resources to train Natural Language Understanding (NLU) models for legal advice services. A labelled dataset based on the language resources we propose is essential for developing non-face-to-face legal services that provide information related to legal problems. The language resources were constructed by analysing in a bottom-up form the linguistic patterns of legal expressions, expressions describing background, and expressions requesting information, in online legal advice texts. Moreover, we analysed the hierarchical classification of keywords in existing legal service systems and delimited 4 representative categories of legal topics, with 20 subcategories. Local Grammar Graphs (LGGs), an efficient framework for describing local linguistic phenomena, were adopted to describe the linguistic patterns in this domain. These local patterns, modularized in LGG format, are converted into Finite State Transducers (FSTs) and generate labelled datasets required for training a language model for NLU. To evaluate this approach, we trained with a resulting dataset an NLU classifier of user intents, designed according to the open-source chatbot architecture Rasa. The classifier has a 0.91 f1-score, which shows that the language resources and the method proposed in this study can actually be applied in developing legal advice chatbot systems.