We present Local Graph-based Dictionary Expansion (LGDE), a method for data-driven discovery of the semantic neighborhood of words using tools from manifold learning and network science. At the heart of LGDE lies the creation of a word similarity graph from the geometry of word embeddings followed by local community detection based on graph diffusion. The diffusion in the local graph manifold allows the exploration of the complex nonlinear geometry of word embeddings to capture word similarities based on paths of semantic association, over and above direct pairwise similarities. Exploiting such semantic neighborhoods enables the expansion of dictionaries of pre-selected keywords, an important step for tasks in information retrieval, such as database queries and online data collection. We validate LGDE on two user-generated English-language corpora and show that LGDE enriches the list of keywords with improved performance relative to methods based on direct word similarities or co-occurrences. We further demonstrate our method through a real-world use case from communication science, where LGDE is evaluated quantitatively on the expansion of a conspiracy-related dictionary from online data collected and analyzed by domain experts. Our empirical results and expert user assessment indicate that LGDE expands the seed dictionary with more useful keywords due to the manifold-learning-based similarity network.

Clinical data often exists in different forms across the lifetime of a patient’s interaction with the healthcare system - structured, unstructured or semi-structured data in the form of laboratory readings, clinical notes, diagnostic codes, imaging and audio data of various kinds, and other observational data. Formulating a representation model that aggregates information from these heterogeneous sources may allow us to jointly model on data with more predictive signal than noise and help inform our model with useful constraints learned from better data. Multimodal fusion approaches help produce representations combined from heterogeneous modalities, which can be used for clinical prediction tasks. Representations produced through different fusion techniques require different training strategies. We investigate the advantage of adding narrative clinical text to structured modalities to classification tasks in the clinical domain. We show that while there is a competitive advantage in combined representations of clinical data, the approach can be helped by training guidance customized to each modality. We show empirical results across binary/multiclass settings, single/multitask settings and unified/multimodal learning rate settings for early and late information fusion of clinical data.