for a query • Use the Indri search engine to index all returned snippets – Krovetz stemming, discard stopwords (Lemur stopword list) • Use several different approaches to form a query – Next slide • Rank snippets using Indri’s language modeling algorithm – Default parameters
Transforma7ons inves7gated • iden%ty: The query itself • sdm: The sequenAal dependency model (Metzler & CroS) • expansion with word vectors: Add addiAonal terms to the query
• n = 300 for our system • Vectors preserve semanAc and syntacAc similariAes – vectors for cat and dog are similar • Used vectors trained on a Google news corpus provided by Mikolov at al Add terms to query based on how close their vectors are to original query terms
close their vectors are to original query terms • Obtain the average word vector v for query q – Average the word vectors vi for each query term qi ϵ q • Expansion 1: Select k expansion terms – The k terms that are closest to v • Expansion 2: Select k expansion terms for each query term – The k terms that are closest to qi – k= 3 (set by a parameter sweep over FW 13 data) • Distances measured using Euclidean distance with a threshold of 0.7
• SDM was slightly beger than a plain query – Bigrams and windows don’t help much in snippets • Expansion was slightly beger than not expanding – The two expansion methods performed about equally