Transcript t-score
ENG 626 CORPUS APPROACHES TO LANGUAGE STUDIES word-based methodology Bambang Kaswanti Purwo [email protected] » the concordance output ▪ a concordance program arranges all instances of a particular search item in the center of the page ▪ the search item = the node ▪ the items to the left and to the right of the node = the span ▪ the span can be specified: a span of four or five items to the left (L) and to the right (R) ▪ four or five items to L and to R – a commonly used range ▪ a span to L “N-1, N-2, etc.” and to R “N+1, N+2, etc.” collocation (Hunston 2005, 68 ff.) • the tendency of words to be biased in the way they co-occur ▪ motivated: there is a logical explanation for collocation e.g. toys – children, more frequently than – women, men ▪ unmotivated: no logical explanation e.g. strong tea vs. powerful car • collocation may be observed informally in any instance of language • more reliable to measure collocation statistically corpus study ▪ a tendency of two words to co-occur ▪ a tendency of one word to attract another measurements of collocation • any program which calculates collocation ▪ takes a node word ▪ counts the instances of all words occurring within a particular span (four words to the L or the R of the node) (see a 4:4 span of the node word gaze) ▪ punctuation marks are ignored ▪ items marked with‘s are counted as a separate word ▪ sentence boundaries are ignored ▪ some of the words co-occurring with gaze are supposed to do so by chance (e.g. wait … gaze, life … gaze) ▪ others might be said to be “meaningful” (e.g. penetrating gaze, my/her/his gaze) ▪ [calculation of large quantities of data] “chance” collocation – insignificant when compared with “meaningful” collocation ▪ the lines include both gaze [n] and gaze [v] ▪ the search could be restricted to ▫ restrict the search to [n] or [v] ▫ to choose only the word-form gaze or lemma GAZE ▪ see all instances of gaze [n] from Bank of English ▫ a list of fifteen most frequent collocates (out of 2,864 lines) ▫ the plural is not found in the corpus ▪ the words at the top – all grammatical words ▪ the high frequency: determiners (the, his, her, my, their) prepositions (from, with, under) ▪ two lexical words in the list: ▫ public in the phrase public gaze ▫ fixed in expressions such as fixed his gaze on, her gaze was fixed on ▫ a list of raw frequencies: how to attach a precise degree of importance to any of the figure? his occurs near the top of the list: is it significant? » how calculate the significance of each co-occurrence? • three most common measures of significance are ▪ Mutual Information (MI) score ▪ t-score ▪ z-score • MI score n t-score depend on two calculations: ▪ of the co-occurring words are found in the designated span? (the Observed) how many instances ▪ might be expected in that span, given the frequency of the cooccurring word in the corpus as whole? (the Expected) ▪ t-score uses a calculation of standard deviation; it takes into account ▫ the probability of co-occurrence of the node and its collocates ▫ the number of tokens in the designated span in all lines ▪ t-score is calculated by subtracting Expected from Observed and dividing the result by the standard deviation ▪ MI-score is the Observed divided by the Expected, converted to a base-2 logarithm ▪ [the “Lookup” package available with Bank of English corpus] ▫ MI-score indicates the strength of a collocation the actual co-occurrence of the two items with their expected co-occurrence if the words in the corpus used were to co-occur in totally random order ▪ the MI-score measures the amount of non-randomness present when two words co-occur MI-score of 3 or higher significant e.g. ballpoint + pen; distinctly + unenthusiastic, hardly + surprising ▪ words such as baleful and unwavering are more generally associated with gaze, although they are not particularly frequent words ▪ if a word occurs rarely, but in most of its few co-occurrences appears in the proximity of another word, the collocation between those words will obtain a high MI-score MI is a measure of how strongly two words seem to associate in a corpus, based on the independent relative frequency of the two words » the case of baleful and gaze: knowing the strength of the collocation not always reliable indication of meaningful association » how certain we can be that the collocation is the result of more than the vagaries of a particular corpus? [vagaries = ‘unexpected changes that you cannot control’] » the calculation of t-score (which takes the amount of evidence into account) can be used ▪ a t-score of 2 or higher significant ▪ high t-score in Bank of English corpus: e.g. things + considered, could + hardly, argument + that, children + toys (see the list of 15 collocates of gaze with highest t-score ▪ the t-score list is different from ▫ the MI-score list (collocates depending on particular parts of the corpus) ▫ the raw frequency list ▪ the + gaze does not have a high t-score the because it is a frequent word, not because of its association with gaze ▪ t-score confirms the collocation between his + gaze not due purely to the high frequency of his but to the lexical preferences of gaze ▪ his + gaze do not have a high MI-score, because his also collocates with so many other things there is a lot of evidence of their co-occurrence (hence they have a high t-score) » importance difference between MI-score n t-score • MI-score is a measure of strength of collocation t-score is a measure of certainty of collocation this is because: ▪ the value of an MI-score not particularly dependent on the size of the corpus ▪ for t-score, corpus size is important (the amount of evidence is being taken into account) • MI-scores can be compared across corpora, even if the corpora are of different sizes • absolute t-scores cannot be compared across corpora (though it is reasonable to compare t-score ranking) • in the Bank of English: corpus of Times twice the size as that of the Economist ▪ decided occurs roughly the same frequency in each ▪ collocation of decided with to is very comparable in each corpus in terms of MI-score but the t-score in Times corpus is much higher the collocation decided to is stronger in Times corpus the Times figures is higher not because the collocation is stronger, but because the corpus is bigger (both in Times and Economist, to is the most frequent word to the right of decided) • looking at the top collocates from the view point of t-score info about the grammatical behavior of a word from the viewpoint of MI-score info about the lexical behavior (particularly, more idiomatic (“fixed”) co-occurrences Hunston Ch. 4 word-based methodology vs. category-based methodology • to answer different sets of questions • a useful synergy between the two methodologies (see Hunston, p. 86, BEGIN vs. START) • word-based methodology frequency and key-word lists collocation • the use of collocational information (Hunston, p. 76) LEAK [v] ▪ associated with the physical meaning ▪ metaphoric sense ▪ prepositions and advs of direction – important give a semantic profile of LEAK cf. CAUSE vs. PROVIDE: CAUSE is used with nouns “sth bad’ ▪ LEAK in the physical sense can have either ▫ the substance as subject (the oil leaked out) ▫ the container as subject (the tank leaked out) ▪ the metaphoric sense behaves in a similar way, with an additional pattern: + to (it’s not the case with the physical sense) This can be seen only from the concordance lines • shoulder examine the words that occur one, two, and three places to the left ▪ his – the most significant word immediately to the left ▪ the most significant word occurring two places to L: over over is used following a range of verbs: LOOK over his shoulder, GLANCE over his shoulder ▪ shoulder often follows a possessive (R or L)