Distributional Semantics and Word Embeddings. Chase Geigle

Transcription

1 Distributional Semantics and Word Embeddings Chase Geigle

2 What is a word? dog 2

3 What is a word? dog canine 2

4 What is a word? dog canine 3 2

5 What is a word? dog 3 canine 399,999 2

6 What is a word? dog canine 399,999 2

7 What is a word? dog canine 399,

8 What is a word? dog canine 399, Sparsity! 2

9 The Scourge of Sparsity In the real world, sparsity hurts us: Low vocabulary coverage in training high OOV rate in applications (poor generalization) one-hot representations huge parameter counts information about one word completely unutilized for another! Thus, a good representation must: reduce parameter space, improve generalization, and somehow transfer or share knowledge between words 3

10 The Distributional Hypothesis You shall know a word by the company it keeps. (J. R. Firth, 1957) Words with high similarity occur in the same contexts as one another. 4

11 The Distributional Hypothesis You shall know a word by the company it keeps. (J. R. Firth, 1957) Words with high similarity occur in the same contexts as one another. A bit of foreshadowing: A word ought to be able to predict its context (word2vec Skip-Gram) A context ought to be able to predict its missing word (word2vec CBOW) 4

12 Brown Clustering

13 Language Models A language model is a 5

14 Language Models A language model is a probability distribution over 5

15 Language Models A language model is a probability distribution over word sequences. 5

16 Language Models A language model is a probability distribution over word sequences. Unigram language model p(w θ): p(w θ) = N p(w i θ) i=1 5

17 Language Models A language model is a probability distribution over word sequences. Unigram language model p(w θ): p(w θ) = N p(w i θ) i=1 Bigram language model p(w i w i 1, θ): p(w θ) = N p(w i w i 1, θ) i=1 5

18 Language Models A language model is a probability distribution over word sequences. Unigram language model p(w θ): p(w θ) = N p(w i θ) i=1 Bigram language model p(w i w i 1, θ): p(w θ) = N p(w i w i 1, θ) i=1 n-gram language model: p(w n w n 1 1, θ) p(w θ) = N i=1 p(w i w i 1 i n+1, θ) 5

19 A Class-Based Language Model 1 Main Idea: cluster words into a fixed number of clusters C and use their cluster assignments as their identity instead (reducing sparsity) If π : V C is a mapping function from a word type to a cluster (or class), we want to find π = arg max p(w π) where N p(w π) = p(c i c i 1 )p(w i c i ). i=1 1 Peter F. Brown et al. Class-based N-gram Models of Natural Language. In: Comput. Linguist (Dec. 1992), pp

20 Finding the best partition π = arg max π One can derive 2 L(π) to be w P(W π) = arg max log P(W π) π L(π) = n w log n w + n ci,c j log n ci n cj ci }{{}},cj {{} (nearly) unigram entropy (nearly) mutual information (fixed w.r.t. π) (varies with π) nc i,c j 2 Sven Martin, Jörg Liermann, and Hermann Ney. Algorithms for Bigram and Trigram Word Clustering. In: Speech Commun (Apr. 1998), pp

21 Finding the best partition L(π) = w n w log n w + n ci,c j log n c i,c j n c }{{} i,c ci n cj j }{{} (nearly) unigram entropy (nearly) mutual information (fixed w.r.t. π) (varies with π) What does maximizing this mean? Recall MI(c, c ) = c i,c j p(c i, c j ) log p(c i, c j ) p(c i )p(c j ) which can be shown to be MI(c, c 1 ) = n ci,c }{{} N j log c i,c j (constant) n ci,cj n ci n cj + log N }{{} (constant) and thus maximizing MI of adjacent classes selects the best π. 8

22 Finding the best partition (cont d) π = arg max π n w log n w w }{{} (nearly) unigram entropy (fixed w.r.t. π) + c i,c j n ci,c j log n c i,c j n ci n cj }{{} (nearly) mutual information (varies with π) Direct maximization is intractable! Thus, agglomerative (bottom-up) clustering is used as a greedy heuristic. The best merge is determined by the lowest loss in average mutual information. 9

23 Agglomerative Clustering 10

24 Do they work? (Yes.) Named entity recognition: Scott Miller, Jethran Guinness, and Alex Zamanian. Name Tagging with Word Clusters and Discriminative Training. In: HLT-NAACL. 2004, pp Lev Ratinov and Dan Roth. Design Challenges and Misconceptions in Named Entity Recognition. In: Proc. CoNLL. 2009, pp Dependency parsing: Terry Koo, Xavier Carreras, and Michael Collins. Simple Semi-supervised Dependency Parsing. In: Proc. ACL: HLT. June 2008, pp Jun Suzuki and Hideki Isozaki. Semi-Supervised Sequential Labeling and Segmentation Using Giga-Word Scale Unlabeled Data. In: Proc. ACL: HLT. June 2008, pp Constituency parsing: Marie Candito and Benoı t Crabbé. Improving Generative Statistical Parsing with Semi-supervised Word Clustering. In: IWPT. 2009, pp Muhua Zhu et al. Fast and Accurate Shift-Reduce Constituent Parsing. In: ACL. Aug. 2013, pp

25 Vector Spaces for Word Representation

26 Toward Vector Spaces Brown clusters are nice, but limiting: Arbitrary cut-off point two words may be assigned different classes arbitrarily because of our chosen cutoff Definition of similarity is limited to only bigrams of classes 3 Completely misses lots of regularity that s present in language 3 There are adaptations of Brown clustering that move past this, but what we ll see today is still even better. 12

27 What do I mean by regularity? woman is to sister as man is to 13

28 What do I mean by regularity? woman is to sister as man is to brother 13

29 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to 13

30 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow 13

31 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to 13

32 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen 13

33 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen fell is to fallen as ate is to 13

34 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen fell is to fallen as ate is to eaten 13

35 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen fell is to fallen as ate is to eaten running is to ran as crying is to 13

36 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen fell is to fallen as ate is to eaten running is to ran as crying is to cried 13

37 What do I mean by regularity? woman is to sister as man is to brother summer is to rain as winter is to snow man is to king as woman is to queen fell is to fallen as ate is to eaten running is to ran as crying is to cried The differences between each pair of words are similar. Can our word representations capture this? (demo) 13

38 14

39 15

40 Neural Word Embeddings

41 Associate a low-dimensional, dense vector w with each word w V so that similar words (in a distributional sense) share a similar vector representation. 15

42 If I only had a vector: The Analogy Problem To solve analogy problems of the form w a is to w b as w c is to what?, we can simply compute a query vector q = w b w a + w c and find the most similar word vector v W to q. If we normalize q to unit-length ˆq = q q and assume each vector in W is also unit-length, this reduces to computing arg max v ˆq v W and returning the associated word v. 16

43 The CBOW and Skip-Gram Models 4 (word2vec) INPUT PROJECTION OUTPUT INPUT PROJECTION OUTPUT w(t-2) w(t-2) w(t-1) SUM w(t-1) w(t) w(t) w(t+1) w(t+1) w(t+2) w(t+2) CBOW Skip-gram 4 Tomas Mikolov et al. Efficient estimation of word representations in vector space. In: ICLR Workshop (2013). 17

44 Skip-Gram, Mathematically Training corpus w 1, w 2,..., w N (with N typically in the billions) from a fixed vocabulary V. Goal is to maximize the average log probability: 1 N N i=1 L k L;k 0 log p(w i+k w i ) Associate with each w V an input vector w R d and an output vector w R d. Model context probabilities as p(c w) = exp(w c) c V exp(w c ). The problem? V is huge! log p(c w) takes time O( V ) to compute! 18

45 Negative Sampling 5 Given a pair (w, c), can we determine if this came from our corpus or not? Model probabilistically as p(d = 1 w, c) = σ(w c) = exp( w c). Goal: maximize p(d = 1 w, c) for pairs (w, c) that occur in the data. Also maximize p(d = 0 w, c N ) for (w, c N ) pairs where c N is sampled randomly from the empirical unigram distribution. 5 Tomas Mikolov et al. Distributed Representations of Words and Phrases and their Compositionality. In: Advances in Neural Information Processing Systems , pp

46 Skip-Gram Negative Sampling Objective Locally, l = log σ(w c) + k E cn p D (c N ) [log σ( w c N )] and thus globally L = (n w,c ) ( log σ(w c) + k E cn P n(c N ) [log σ( w c N )] ) w V c V k: number of negative samples to take (hyperparameter) n w,c : number of times (w, c) was seen in the data E cn P n(c N ) indicates an expectation taken with respect to the noise distribution P n (c N ). 20

47 Implementation Details (Read: Useful Heuristics) 1. What is the noise distribution? P n (c N ) = (n c N /N) 3/4 (which is the empirical unigram distribution raised to the 3/4 power). 2. Frequent words can dominate the loss. Throw away word w i in the training data according to t P(w i ) = 1 where t is some threshold (like 10 5 ). Z n wi Both are essentially unexplained by Mikolov et al. 21

48 What does this actually do? It turns out this is nothing that new! Levy and Goldberg 6 show that SGNS is implicitly factorizing the matrix M i,j = w i c j = PMI(w i, c j ) log k = log p(w i, c j ) p(w i )p(c j ) log k using an objective that weighs deviations in more frequent (w, c) pairs more strongly than less frequent ones. Thus 6 Omer Levy and Yoav Goldberg. Neural Word Embedding as Implicit Matrix Factorization. In: Advances in Neural Information Processing Systems , pp

49 Matrix Factorization Methods for Word Embeddings

50 Why not just SVD? Since SGNS is just factorizing a shifted version of the PMI matrix, why not just perform a rank-k SVD on the PMI matrix directly? Assignment 6! With a little TLC, this method can actually work and does surprisingly well. 23

51 GloVe: Global Vectors for Word Representation 7 How can we capture words related to ice, but not to steam? Prob. or Ratio w k = solid w k = gas w k = water w k = fashion P(w k ice) P(w k steam) P(w k ice) P(w k steam) Probability ratios are most informative: solid is related to ice but not steam gas is related to steam but not ice water and fashion do not discriminate between ice or steam (ratios close to 1) 7 Jeffrey Pennington, Richard Socher, and Christopher D. Manning. GloVe: Global Vectors for Word Representation. In: EMNLP. 2014, pp

52 Building a model: intuition We would like for the vectors w i, w j, and w k to be able to capture the information present in the probability ratio. If we set then we can easily see that w i w k = log P(w k w i ) (w i w j ) w k = log P(w k w i ) log P(w k w j ) = log P(w k w i ) P(w k w j ), ensuring that the information present in the co-occurrence probability ratio is expressed in the vector space. 25

53 GloVe Objective w i w j + b i + b j = log X ij Again: two sets of vectors: target vectors w and context vectors w. X ij = n wi,w j is the number of times w j appears in the context of w i. Problem: all co-occurrences are weighted equally J = V f (X ij )(w i w j + b i + b j log X ij ) 2 i,j=1 Final objective is just weighted least-squares. f (X ij ) serves as a dampener, lessening the weight of the rare co-occurrences. ( ) α x x f (x) = max if x < xmax 1 otherwise where α s default is (you guessed it) 3/4, and x max s default is

54 Relation to word2vec GloVe Objective: V J = f (X ij )(w i i,j=1 w j + b i + b j log X ij ) 2 word2vec (Skip-Gram) Objective (after rewriting): V V J = P(w j w i ) log Q(w j w i ) i=1 X i j=1 where X i = k X ik and P and Q are the empirical co-occurrence and model co-occurrence distributions, respectively. Weighted cross-entropy error! Authors show that replacing cross-entropy with least-squares nearly re-derives GloVe: Ĵ = i,j X i (w i w j log X ij ) 2 27

55 Model Complexity word2vec: O( C ), linear in corpus size GloVe naïve estimate: O( V 2 ), square of vocab size But it actually only depends on the number of non-zero entries in X. If co-occurrences are modeled via a power law, then we have X ij = k (r ij ) α. Modeling C under this assumption, the authors eventually arrive at { O( C ) if α < 1 X = O( C 1/α ) otherwise where the corpora studied in the paper were well modeled with α = 1.25, leading to O( C 0.8 ). In practice, it s faster than word2vec. 28

56 Accuracy vs Vector Size Accuracy [%] Semantic Syntactic Overall Vector Dimension 29

57 Accuracy vs Window Size Accuracy [%] Semantic Syntactic Overall Accuracy [%] Semantic Syntactic Overall Window Size Window Size Left: Symmetric window, Right: Asymmetric window 30

58 Accuracy vs Corpus Choice 85 Semantic Syntactic Overall 80 Accuracy [%] Wiki2010 1B tokens Wiki B tokens Gigaword5 4.3B tokens Gigaword5 + Wiki2014 6B tokens Common Crawl 42B tokens 31

59 Training Time 72 Training Time (hrs) Accuracy [%] GloVe Skip-Gram Iterations (GloVe) Negative Samples (Skip-Gram) 32

60 This evaluation is not quite fair! 32

61 But wait (Referring to word2vec) the code is designed only for a single training epoch specifies a learning schedule specific to one pass through the data, making a modification for multiple passes a non-trivial task. 33

62 But wait (Referring to word2vec) the code is designed only for a single training epoch specifies a learning schedule specific to one pass through the data, making a modification for multiple passes a non-trivial task. This is false, and a bad excuse. How could you modify word2vec to support multiple epochs with zero effort? 33

63 But wait we choose to use the sum W + W as our word vectors. 34

64 But wait we choose to use the sum W + W as our word vectors. word2vec also learns W! Why not do the same for both? 34

65 Embedding performance is very much a function of your hyperparameters! 34

66 Investigating Hyperparameters 8 Levy, Goldberg, and Dagan perform a systematic evaluation of word2vec, SVD, and GloVe where hyperparameters are controlled for and optimized across different methods for a variety of tasks. Interesting results: And: MSR s analogy dataset is the only case where SGNS and GloVe substantially outperform PPMI and SVD. SGNS outperforms GloVe in every task. 8 Omer Levy, Yoav Goldberg, and Ido Dagan. Improving Distributional Similarity with Lessons Learned from Word Embeddings. In: Transactions of the Association for Computational Linguistics 3 (2015), pp

67 But wait (mark II) Both GloVe and word2vec use context window weighting: word2vec: samples a window size l [1, L] for each token before extracting counts GloVe: weighs contexts by their distance from the target word using the harmonic function 1 d What is the impact of using the different context window strategies between methods? Levy, Goldberg, and Dagan did not investigate this, instead using word2vec s method across all of their evaluations. 36

68 Evaluating things properly is non-obvious and often very difficult! 36

69 What You Should Know

70 What You Should Know What is the sparsity problem in NLP? What is the distributional hypothesis? How does Brown clustering work? What is it trying to maximize? What is agglomerative clustering? What is a word embedding? What is the difference between the CBOW and Skip-Gram models? What is negative sampling and why is it useful? What is the learning objective for SGNS? What is the matrix that SGNS is implicitly factorizing? What is the learning objective for GloVe? What are some examples of hyperparameters for word embedding methods? 37

71 Implementations word2vec: The original implementation: Yoav Goldberg s word2vecf modification (multiple epochs, arbitrary context features): Python implementation in gensim: GloVe: The original implementation: MeTA: R implementation text2vec: 38

72 Thanks! 38