Computational Approaches to Language and Health

Computational methods help understand human word segmentation in a new language. Research with non-Māori-speaking New Zealanders (NMS) exposed to Te Reo Māori shows their ability to divide words resembles fluent speakers. This skill is often linked to identifying statistically recurrent patterns in language input. However, comparing human performance to a machine learning model, Morfessor, which relies on statistical recurrence, suggests a broader process. Both NMS and Morfessor accurately segment words formed by joining parts, such as in Kaitiakitanga in Māori (compounding Kai + Tiaki + Tanga) or affixing, as with causative prefix Faka- in Tongan (simple affixation). NMS also perform well with words involving structural templates, like reduplication (as in Kūʻai vs. Kūʻaiʻai in Hawaiian) and allomorphy, as in causative suffixes -ti and -ri in Gilbertese.

This indicates human language learning extends beyond tracking how often forms appear together; it appears sensitive to higher-level structural templates. For example, on words formed by total reduplication, NMS showed high performance (0.95 precision, 0.97 recall). Morfessor's performance was lower (0.35 precision, 0.36 recall). This indicates human learners may induce abstract templates for word formation. The study suggests NMS may detect features like vowel length in word segmentation, a cue not explicitly used by the statistical model. Comparing performance on real Māori words to constructed words with the same statistical properties supports the presence of cues beyond statistical recurrence in natural language. Morfessor segmented pseudo-Māori better (mean precision 0.84, recall 0.96) than real Māori (precision 0.80, recall 0.87). This illustrates the model's dependence on statistical principles and the existence of other cues in natural language.  

Moving from language structure and learning to application, a major experiment at Université de Toulouse was conducted, where computational approaches analyzed speech in medical settings, specifically to assess speech intelligibility in patients with head and neck cancers. Treatments for these cancers can affect communication, leading to reduced speech intelligibility and impacting a patient's quality of life. Traditional assessment uses subjective perceptual evaluations by clinicians, which can vary. Automatic methods could provide more objective and dependable predictions.

An automatic approach uses speaker embeddings, which convert speech segments into fixed-dimensional vectors capturing speaker traits. These embeddings are useful in evaluating pathological speech. A system using speaker embeddings and a neural network predicted speech intelligibility and disorder severity in a multi-task framework. The system was trained and validated on data from head and neck cancer patients, and then tested on tasks including text reading, phrases, pseudo-words, and spontaneous speech.

Results show good correlations and low errors for predicting intelligibility and severity. The system performed well on spontaneous speech, with a Spearman correlation of 0.828 and an RMSE of 1.468 using x-vectors. This is important as spontaneous speech better reflects a patient's communication in real clinical settings. Automatic evaluation of spontaneous speech intelligibility addresses a need in the literature and supports more relevant and dependable clinical predictions. The system was trained on text reading, and its performance on other tasks and new patients suggests some robustness. Further training on diverse tasks, languages, and conditions could improve performance.

Beyond individual language processing and clinical speech analysis, computational approaches can identify disparities in health outcomes within populations such as Asian American, Native Hawaiian, and Pacific Islander (AANHPI) individuals in the US, particularly regarding cancer. The AANHPI community in the United States is diverse, with many distinct ethnic groups having different cancer risks and outcomes. Grouping data can hide these differences so there is a need to divide information into smaller subgroups to find specific health issues.

Disparities exist across cancer types within AANHPI populations. Stomach cancer has high incidence and mortality in Korean, Samoan, Japanese, Chinese, and Vietnamese groups. Liver cancer is more common in Vietnamese, Cambodian, Laotian, Chinese, and Korean individuals, partly due to higher chronic hepatitis B rates. Nasopharyngeal cancer shows higher incidence in Laotian, Chinese, Vietnamese, Filipino, Cambodian, and Native Hawaiian populations. For breast cancer, Native Hawaiian and Pacific Islander groups, specifically Tongan, Chamorro, and Samoan individuals, have higher mortality rates. Cervical cancer also shows elevated rates in Native Hawaiian and Pacific Islander populations. Computational analysis of detailed health data can identify these specific disparities. This guides focused actions in screening and addressing risk factors, which our translation team help communicate.  

Furthermore, computational methods help analyze and improve processes in cancer research. MIT work explores multi-document conditional reasoning, a task needing models to understand conditions across several documents and reason about best outcomes. While not directly on cancer treatment, this natural language processing area is similar to treatment rules or clinical trial eligibility. A dataset, MDCR (Multi-Document Conditional Reasoning), tests models on this.  

The MDCR dataset includes documents with eligibility conditions and user scenarios. It asks questions requiring reasoning over multiple documents. These questions involve determining eligibility for one or all outcomes, or finding the highest number of outcomes possible. Difficulty comes from understanding relationships between conditions across documents (conflicting, equivalent, inclusive) and reasoning with conditions not explicitly mentioned. Testing recent language models on MDCR shows this task is hard for current systems. Models have difficulty with longer contexts from multiple documents and the complexity of optimization questions. For example, average short answer accuracy for optimization questions (Q3) was 46.1% across models, compared to 66.0% for simple eligibility questions (Q1). This suggests accurately extracting conditions, understanding if they are met based on a scenario, and logically reasoning with relationships between documents are areas for improvement in computational language understanding. Applying this reasoning to medical documents could help automate checking patient eligibility for treatments or trials based on complex rules.

Computational and statistical approaches to language thus provide varied perspectives. These range from modeling human word learning by analyzing patterns and templates, to using machine learning on speech data for objective health assessment, and using complex reasoning models to work with multi-document information. Analyzing language structure, processing sounds, and reasoning over connected text offers tools for questions in linguistics and extends these findings to areas like health and disease.

The point where language, computation, and health meet presents many areas for future investigation. As computational models improve in understanding human language details and complex information structures, their uses in public health, clinical work, and biomedical research will likely grow. Further investigation at this convergence will yield significant findings.