Exploratory Search

dead ends and rabbit chases on the research journey

What we’re doing today:

Exploratory search (ES)

Visualization in ES

Problem space

ML methods I wanted to learn

Generative LLMs

Opportunities for advancing ES and IS

What is exploratory search (ES)?

ES is:

as a task
ill-formed, poorly structured
repetitive, iterative
long-term with starts and stops
simultaneously broad and narrow
Bates, Kuhlthau, Marchionini, Wilson, White & Roth, Pirolli, Inwersen & Järvellin

Visualization will solve the problem!

The problem space

ES systems (ESS) have not been developed that address bias in citation behaviours

There are few available interfaces that support the ES task

Current information retrieval (IR) applications reinforce‘top of the list’behaviours

ES research has stalled using methods from the 1970’s

ML methods

bag of words methods have advantages, simple, cheap, fast to run, but you loose context/sematics, stop words, counting aggregates, stemming/lemmatization, no proximity

LDA is better, but you still loose context, can compensate with n-grams, but still loss of knowledge of the document, context/semantics, stemming lemmatization, stop words loss, proximity

Word embeddings - much better for context, training needed unless using pretrained set, problems with the training for discipline/field specific issues, small context window, very sparse, but better! now your distance measures become meaningful - cosine, euclidean, normalized euclidean

Recurrent neural networds - much better!, but training data is needed, huge problem is long-distance relationships in text - think long sentences. SLLOW. LSTM (long-short term memory) tried to solve the problem, but it works sequentially - one word at a time. SLOOOOW.

Alll of these get slower the more data you give it.

The transformer advanced LSTM to capture context and dependencies - use of tensors, uses dot product for finding similarity - most important - it gets better with more information! and this is where we are today, transformers are at the heart of LLMs with predition capabilities that get more accurate the more training data and prompt data we throw at it.

Bag of words methods/co-occurrance

Statistical methods - Latent Dirichlet Allocation (LDA)

Word embeddings - Word2vec

Recurrent Neural Networks

Transformers

Image source: (Pitié 2023)

Generative LLMs

This is interesting!

Support natural language query

Query support / prompt recommendations

Semantic, contextual relationships

Synthesis / summarization

Grounded in specific data

The more data, the better!

Are gen LLM’s good for ES?

What if I could do both search and retrieval?

IS perspective

Query formulation/re-formulation

Synthesis / summarization

Learning, comprehension, and mental models

Grounding with bibliometric metadata

Citaitonal justice with semantic, contextuall relationships

Task definition and perception

Affect responses

Thank you!

Interested in this research area?

Poppy Riddle

pnriddle@dal.ca