Share
Embed
Transcript
18 slides
-
[論文紹介] Don’t count, predict! A systema7c comparison
of context-‐coun7ng vs. context-‐predic7ng seman7c vector Marco Baroni, Georgiana Dinu, German Kruszewski (Center for Mind/Brain Sciences(University of Trento, Italy)) ACL 2014 Yoshifumi Seki 2014.09.02 @Gunosy研究会 -
概要 • word2vecなどのようなcontext-‐predic7ng modelが流行しているが,それらとよく知られ ている頻度ベースのベクトルモデルを体系的 に比較することは行われていない • 様々なlexical
seman7cs taskをいろんなパラ メータを用いて評価した • 結果としてcontext-‐predic7ng modelがcount-‐ based modelより優れた結果を示した -
distribu7onal seman7c models(DSMs) • Using vector that keep track
of context • for decades, raw co-‐occurrence counts don’t work that well • archive higher performance, when various transforma7on are applied to the raw vector. – ex. by reweigh7ng the count for context informa7veness and smoothing them with dimensionality reduc7on techniques. – this vector op7miza7on process is generally unsupervised -
new genera7on of DSMs • The last few years have
seen the development of a new genera7on of DSMs – the vector directly es7ma7on problem • the weights in a word vector are set to maximize the probability of the contexts in the corpus -
new genera7on DSMs is aYrac7ve! • replaces the essen7ally heuris7c
stacking of vector transforms • no manual annota7on cost • some of the relevant methods can efficiently scale up to process very large amounts of input data -
Dataset • a corpus of about 2.8 billion tokens
– ukWaC – English Wikipedia – Bri7sh Na7onal Corpus • the top 300k most frequent words in the corpus as target and context elements -
Count models • using DISSECT tookit – hYp://clic.cimec.unitn.it/composes/toolkit/
• count vectors from symmetric context windows of two and five words • two weigh7ng scheme – pointwise mutual informa7on(PMI) – Local mutual Informa7on • full and compressed vectors – Singular Value Decomposi7on – Non-‐nega7ve Matrix Factoriza7on – ranging 200 to 500 in steps of 100 • In total, 36 count model were evaluated. -
Predict models • using word2vec toolkit • context windows
of 2 and 5 words • vector dimensionality 200 to 500 range in steps of 100 • k: number of nega7ve samples – 5 and 10 • t: words that occur with higher frequency than t are aggressively subsampled – without subsampled – t = exp(-‐5) • we evaluate 48 predict models -
Out-‐of-‐the-‐box models • Distribu7onal Memory(dm) – Using “linguis7cally rich”
data – Baroni and Lenci (2010) – hYp://clic.cimec.unitn.it/dm/ • Collobert and Weston vectors(cw) – 100 dimensional vectors trained for two months on the wikipedia – The vector were trained to op7mize the task of choosing the right word over a random alterna7ve in middle of an 11 word context window – Collobert et al. (2011) – hYp://ronan.collobert.com/senna/ -
Evalua7on Materials • Seman7c relatedness • Synonym detec7on
• Concept categoriza7on • Selec7on preferences • Analogy -
Seman7c relatedness • Rate the degree of seman7c similarity or
relatedness between two words on numerical scale • Rubenstein and Goodenough(1965)(rg) – Consists of 65 noun pairs – state of the art: Hassan and Mihalcea(2011) • Exploits wikipedia linking structure and word sense disambigua7on technique • WordSim353(ws) – Finkelstein et al.(2002) – Consists of 353 pairs – State of the art: Halawi et al.(2014) • Predict models using WordNet • Agirre et al.(2009) split ws set into similarity(wss) and relatedness(wsr) • MEN(men) – 1000 word pairs – Bruni et al.(2014) -
Synonym detec7on • TOEFL set • 80 mul7ple-‐choices ques7on
that pair a target term with 4 synonym candidate • Bullinaria and Levy(2012) archive 100% accuracy -
Concept categoriza7on • The task is to group nominal concepts
into natural categories. • Using CLUTO toolkit – hYp://glaros.dtc.umn.edu/gkhome/views/cluto • Almuhared-‐Poesio benchmark(ap) – Almuhared(2006) – 492 concepts into 21 category • The ESLLI 2008 Distribu7onal Seman7c Workshop shared-‐task set(esslli) – 44 concepts into 6 category -
Selec7on preferences • Verb-‐noun pairs that where rated by subject
for the typicality of the noun as a subject or object of the verb • Ulrike Pado(2007) (up) – 211 pairs • Macrae – 100 noun-‐verb pairs -
Analogy • Specifically to test predict model • 9K
seman7cs and 10.5K syntac7c analogy ques7on – Example • Brother-‐sister, Grandson-‐? – Ans: granddaughter • work-‐works, spreak-‐? – Ans: speaks • En7re dataset (an) – Syntac7c subset(ansyn) – Seman7c subset(ansem) -
Conclusion • 全体として見た時にCount modelよりpredict modelのほうがよい – 精度が全体的に高い – パラメータ設定・データセットの違いに対してロバ
ストである • Seman7cs, synonym両方に強い