Relation extraction (RE) identifies semantic relations between named entities and is crucial for tasks such as knowledge graph construction and question answering. RE can be leveraged for automatic relationship annotation by employing trained RE models to identify and annotate semantic relationships between ontology-driven entities in unstructured, conversational texts. This paper examines RE in real-world conversational texts collected from an online dementia forum comprising 45,216 sentences, where language is often ambiguous and departs from standard medical terminology. We present a unified RE pipeline based on the domain-specific annotations and conduct a comparative evaluation of two model types: (i) a neural network-based architecture combining a bidirectional long-short-term memory (Bi-LSTM) model with static GloVe: Global Vectors for Word Representation (GloVe) and a conditional random field (CRF) classifier, and (ii) a transformer-based language model, BERT-Large, followed by a CRF layer. The results show that the use of GloVe embedding in the Bi-LSTM model achieves a significantly higher F1 score of 0.81, compared to 0.64 with the BERT-Large model. Despite lacking contextualised embeddings, the GloVe embedding captures relation patterns more effectively in this low-resource, domain-specific setting. These results demonstrate that lightweight models can outperform large pre-trained transformers in challenging real-world scenarios. The output of the RE task is a set of structured annotations, which can be integrated into downstream knowledge bases. Therefore, the investigated RE models have the potential to be applied as automatic annotators of informal texts from the domain of dementia, transforming the text into richly annotated resources by systematically annotating entity relationships without human intervention.