Research statement: Computational characterization of mental states: a natural language processing approach

CARRILLO FACUNDO

Conferencia; Doctoral Consortium IJCAI School 2014; 2014

The massive (and increasingly accelerating) digital availability of thoughtproducts in textual format opens a window to study the brain and mind inradically novel ways. We propose that through the use of natural language toolson text and state-of-the-art mathematical approaches we may assess mentalstates with unprecedented detail and precision.While biological psychiatry and neuroscience seek to understand the brainand the mind, they face the problem of bridging the large gap between the microandmacroscopic levels. Historically, behavioral studies have lacked the rigor andsophistication of neuroscientific research. However, the cognitive sciences fieldhas recently developed an approach to amass enormous amounts of data throughgame-like web applications. Together with the access to large web repositoriesof text, and to practically unlimited computational power, these developmentsare changing the way we can characterize cognition and behavior.We focused in mental health study, most psychiatric evaluations are presentlydone by qualitative analysis of interviews in hospitals or mental institutions.Recently, we have detected stereotyped markers in text that may distinguish betweenpsychosis pathologies (mania and schizophrenia). With the developmentof better analytic methods for detection of these markers, we propose the generationof quantitative (and semi-automatic) techniques that provide the specialistwith additional information taken from written text to help in the diagnosis andtracing of patients.The goal of this plan is the development and use of machine-learning techniquesto study massive-scale digital text corpora associated with cognitive processes,aiming at identifying the mental operations underlying behavioral processes.The goal is two fold: first, to decode the regularities in the corpora to infergenerative rules of human thought, and second, to incorporate them in artificialintelligence models.Written text is the maximum exponent of this new reality. Every day, webpages, books, exams, interviews are being produced and stored in digital format,e.g. on the Internet. Based on these repositories, analytic strategies have been investigatedwhich recover the semantic structure of words. Also, machine-learningtechniques have been developed to analyze texts and obtain hidden informationabout them, such as the topics involved or the mental state of the author [1].One of the most complicated issues is the availability of methods and toolsin languages other than English. In previous research we observed that thisDoctoral Consortium IJCAI School 2014 43 JAIIO - DC IJCAI School 2014 - ISSN 2362-5120 - Página 15issue may sometimes be skipped via machine translation methods, obtainingresults similar to manual translation. Still, the adaptation of natural languageprocessing tools into a general framework with texts from different repositoriesin many languages (e.g. Spanish and Portuguese) remains an important aspectof this research program.Most text analysis tools rely on a training corpus, generally curated by experts.With the massive digital availability of texts in most of the world?s mainlanguages (digital newspapers, blogs, web pages) we propose the creation of aframework that may incorporate selected sources of text from the Internet astraining corpus and generate a semantic space in a multi-language perspective.With these tools, we propose the identification of new markers to contributeto the diagnosis and treatment of psychiatric patients.During the last year we created and tested some methods to characterizemental alteration. For this, we implemented a new method to identified thespeech changes produced by drug intoxication effects [2] and we tested in a realcase.At this time we are working in a new multi-language semantic method basedon Twitter. This method is similar to Google Similarity Distances [3] but withmore resolution time. This features allows to study how particulars events changethe semantic network.Also we are working in other medical application: Automated analysis ofspeech predicts transition to psychosis in high-risk patients