Description
We are increasingly relying on the online information spaces to look for answers to a variety of questions and make important decisions. Online decision-making often involves classifying different attributes of information objects and then assimilating them to make a final decision. For example, when predicting whether a news article should be recommended or not, we might classify the article beforehand as relevant or not. However, whether we will find the article relevant can be based on various different criteria, such as its topical relevance, novelty, reliability and readability, and might also depend on the articles read just before. As another example, while judging whether a user would like to be recommended a hotel, we may first classify its different aspects (such as location, cleanness, friendliness of staff, etc.) into positive/negative states. Problems arise when the some of these attributes are “incompatible” in that they cannot be judged simultaneously without considering the order of the judgments, i.e., user’s views on different aspects of the product might be contradictory and judging on one would interfere with the judgment on another. This creates uncertainty in the user's mind towards reaching a final decision, which is “irrational” in term of its violation of various axioms of classical probability theory (e.g., Bayesian rule and Law of Total Probability). Proactive handling of such uncertainty would improve the system’s ability to predict user decisions and behaviours. A common theme behind such decisions is the presence of multiple criteria behind making a judgement, which may conflict with each other leading to uncertainty in the decision which affects subsequent decisions. This uncertainty over a single classification decision may be handled by current models, but current models fail to represent the contextual decisions which have more than one predictions possible or those which are influenced by the context of previous decisions.
In recent years, cognitive models based on the mathematical framework of Quantum Theory (QT) have been proposed to model, predict and provide a conceptual explanation of many of the above-mentioned irrational behaviours, where the uncertainty created in the human cognitive system interferes with the decisions in ways which is similar to that in quantum systems. When applied to model human decisions, QT extends the norms of rationality by generalising classical probability and vector space models (VSM), as complex-valued spaces (called Hilbert space) where the usage of complex numbers carries additional contextual information, which when combined with the rules of probability calculation in QT, helps us model interference between different decisions or decision criteria.
However, so far a systematic empirical investigation on irrational behaviours in online decision making is missing, especially in interactions where it is important to model user behaviours in order to drive more accurate predictions or augment decisions based on QT as a generalisation of classical logic and probability theory. It might be due to a flaw in existing data collection methodologies. For example, common practice while collecting labels for different classes is to reject annotator disagreement as noise and fix one class label for a given object which has been judged by the majority of annotators. It should possible that both labels exist for the object at the same time, and the disagreement between annotators is due to their different contexts which leads to different semantic interpretations of the same information object.
This project aims to address this problem by empirical studies to capture irrational user behaviours in online decision-making. The key is to design protocols for data collection and annotation which take into account dynamic contexts and enable the possibility of irrational decision-making. Take sentiment classification as an example. One way would be to setup different contexts. For example, while evaluating a document, the order of paragraphs or ideas could be changed in a way that changes the sentiment or creates conflicting evidence for different sentiment labels. Another way could be to prime the users differently for a given text or image. For example, showing a highly negative text or image first may make a neutral text or image appear positive in comparison. In addition, we may also create texts or images which have content expressing all possible sentiments with respect to different aspects. The ultiumate goal is to explore how the classical norms of rationality are violated and can the underlying cognitive structure be modelled using quantum mathematics (e.g., using complex-valued Hilbert spaces, projection operators for calculating probabilities, tensor products) or conceptually augmented using features from QT (e.g., superposition, interference, contextuality).
Skills Required:
Good undergraduate degree (2.1 or above) or Master degree in Computing, ideally with experience in natural language processing and information retrieval.
Background Reading:
Uprety, S., Gkoumas, D., Song, D. (2020). A Survey of Quantum Theory Inspired Approaches to Information Retrieval. ACM Computing Surveys. 53(5), Article 98.
Uprety, S., Dehdashti, S., Fell, L., Bruza, P.D., and Song, D. (2019). Modelling Dynamic Interactions Between Relevance Dimensions. ICTIR2019, 35-42.
Uprety, S., Tiwari, P., Dehdashti, S., Fell, L., Song, D., Bruza, P.D., and Melucci, M. (2020). Quantum-like Structure in Multidimensional Relevance Judgements. ECIR2020, 728-742.
Liu, Y., Zhang, Y., Li, Q., Wang, B., Song, D. (2021). What Does Your Smile Mean? Jointly Detecting Multi-Modal Sarcasm and Sentiment Using Quantum Probability. Findings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP2021).
Contact:
Prof. Dawei Song (dawei.song@open.ac.uk)
Explore our qualifications and courses by requesting one of our prospectuses today.