🧑‍🔬 Mohammad Javad Hosseini

Mohammad Javad Hosseini works on how computers understand and reason over natural language text. The papers here address two related questions. The first is how a system can follow the quantities described in a short story well enough to answer arithmetic questions about it. The second is why large language models accept inferences that a text does not actually support. The methods range from learned classifiers over linguistic features to controlled probing experiments on existing models.

The most cited of these papers, Learning to Solve Arithmetic Word Problems with Verb Categorization (2014), introduced ARIS, a solver for addition and subtraction problems that needs no equation templates. It keeps track of how much of each entity sits in each container, such as an owner or a place, and reads every sentence as a change to that state. Its central learned step sorts each verb into one of seven categories, among them observation, transfer between two containers, construction and destruction.

The newer work turns to language models and pretraining. Sources of Hallucination by Large Language Models on Inference Tasks (2023) treats mistaken entailment judgements as a kind of hallucination. It studies LLaMA, GPT-3.5 and PaLM by altering natural language inference datasets in targeted ways and observing how the models' answers move. The alterations include scrambling the premise predicate, substituting type labels for entities, or inserting other entities of the same type.

Hosseini is currently affiliated with Purdue University West Lafayette and Stanford University. Earlier bylines include Sharif University of Technology, the University of Washington, the University of Edinburgh, The Alan Turing Institute and Google. Co-authors on the papers here include Hannaneh Hajishirzi, Oren Etzioni, Nate Kushman, Mark Steedman, Nick McKenna and Mark Johnson.

Mohammad Javad Hosseini has 2 papers in our library, in 2014–2023, most of them on Language models & pretraining; often with Hannaneh Hajishirzi, Mark Johnson, Mark Steedman.