Date of Award
Doctor of Philosophy (PhD)
School of Computing
Dr Peter Hastings
Dr Noriko Tomuro
Dr. Jonathan Gemmell
Dr. Barbara Di Eugenio
Dr. Daniela Raicu
While there has been an increasing focus on higher-level thinking skills arising from the Common Core Standards, many high-school and middle-school students struggle to combine and integrate information from multiple sources when writing essays. Writing is an important learning skill, and there is increasing evidence that writing about a topic develops a deeper understanding in the student. However, grading essays is time consuming for teachers, resulting in an increasing focus on shallower forms of assessment that are easier to automate, such as multiple-choice tests. Existing essay grading software has attempted to ease this burden but relies on shallow lexico-syntactic features and is unable to understand the structure or validity of a student’s arguments or explanations. Without the ability to understand a student’s reasoning processes, it is impossible to write automated formative assessment systems to assist students with improving their thinking skills through essay writing.
In order to understand the arguments put forth in an explanatory essay in the science domain, we need a method of representing the causal structure of a piece of explanatory text. Psychologists use a representation called a causal model to represent a student's understanding of an explanatory text. This consists of a number of core concepts, and a set of causal relations linking them into one or more causal chains, forming a causal model. In this thesis I present a novel system for automatically constructing causal models from student scientific essays using Natural Language Processing (NLP) techniques.
The problem was decomposed into 4 sub-problems - assigning essay concepts to words, detecting causal-relations between these concepts, resolving coreferences within each essay, and using the structure of the whole essay to reconstruct a causal model. Solutions to each of these sub-problems build upon the predictions from the solutions to earlier problems, forming a sequential pipeline of models. Designing a system in this way allows later models to correct for false positive predictions from downstream models. However, this also has the disadvantage that errors made in earlier models can propagate through the system, negatively impacting the upstream models, and limiting their accuracy. Producing robust solutions for the initial 2 sub problems, detecting concepts, and parsing causal relations between them, was critical in building a robust system.
A number of sequence labeling models were trained to classify the concepts associated with each word, with the most effective approach being a bidirectional recurrent neural network (RNN), a deep learning model commonly applied to word labeling problems. This is because the RNN used pre-trained word embeddings to better generalize to rarer words, and was able to use information from both ends of each sentence to infer a word's concept. The concepts predicted by this model were then used to develop causal relation parsing models for detecting causal connections between these concepts. A shift-reduce dependency parsing model was trained using the SEARN algorithm and out-performed a number of other approaches by better utilizing the structure of the problem and directly optimizing the error metric used.
Two pre-trained coreference resolution systems were used to resolve coreferences within the essays. However a word tagging model trained to predict anaphors combined with a heuristic for determining the antecedent out-performed these two systems. Finally, a model was developed for parsing a causal model from an entire essay, utilizing the solutions to the three previous problems. A beam search algorithm was used to produce multiple parses for each sentence, which in turn were combined to generate multiple candidate causal models for each student essay. A reranking algorithm was then used to select the optimal causal model from all of the generated candidates.
An important contribution of this work is that it represents a system for parsing a complete causal model of a scientific essay from a student's written answer. Existing systems have been developed to parse individual causal relations, but no existing system attempts to parse a sequence of linked causal relations forming a causal model from an explanatory scientific essay. It is hoped that this work can lead to the development of more robust essay grading software and formative assessment tools, and can be extended to build solutions for extracting causality from text in other domains. In addition, I also present 2 novel approaches for optimizing the micro-F1 score within the design of two of the algorithms studied: the dependency parser and the reranking algorithm. The dependency parser uses a custom cost function to estimate the impact of parsing mistakes on the overall micro-F1 score, while the reranking algorithm allows the micro-F1 score to be optimized by tuning the beam search parameter to balance recall and precision.
Hughes, Simon Mark, "Automatic inference of causal reasoning chains from student essays" (2019). College of Computing and Digital Media Dissertations. 19.