{ "paper_id": "I05-1049", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:24:59.064846Z" }, "title": "Relative Compositionality of Multi-word Expressions: A Study of Verb-Noun (V-N) Collocations", "authors": [ { "first": "Sriram", "middle": [], "last": "Venkatapathy", "suffix": "", "affiliation": { "laboratory": "", "institution": "International Institute of Information Technology -Hyderabad", "location": { "settlement": "Hyderabad", "country": "India" } }, "email": "sriram@research.iiit.ac.in" }, { "first": "Aravind", "middle": [ "K" ], "last": "Joshi", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Pennsylvania", "location": { "settlement": "Philadelphia", "region": "PA", "country": "USA" } }, "email": "joshi@linc.cis.upenn.edu" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Recognition of Multi-word Expressions (MWEs) and their relative compositionality are crucial to Natural Language Processing. Various statistical techniques have been proposed to recognize MWEs. In this paper, we integrate all the existing statistical features and investigate a range of classifiers for their suitability for recognizing the non-compositional Verb-Noun (V-N) collocations. In the task of ranking the V-N collocations based on their relative compositionality, we show that the correlation between the ranks computed by the classifier and human ranking is significantly better than the correlation between ranking of individual features and human ranking. We also show that the properties 'Distributed frequency of object' (as defined in [27]) and 'Nearest Mutual Information' (as adapted from [18]) contribute greatly to the recognition of the non-compositional MWEs of the V-N type and to the ranking of the V-N collocations based on their relative compositionality.", "pdf_parse": { "paper_id": "I05-1049", "_pdf_hash": "", "abstract": [ { "text": "Recognition of Multi-word Expressions (MWEs) and their relative compositionality are crucial to Natural Language Processing. Various statistical techniques have been proposed to recognize MWEs. In this paper, we integrate all the existing statistical features and investigate a range of classifiers for their suitability for recognizing the non-compositional Verb-Noun (V-N) collocations. In the task of ranking the V-N collocations based on their relative compositionality, we show that the correlation between the ranks computed by the classifier and human ranking is significantly better than the correlation between ranking of individual features and human ranking. We also show that the properties 'Distributed frequency of object' (as defined in [27]) and 'Nearest Mutual Information' (as adapted from [18]) contribute greatly to the recognition of the non-compositional MWEs of the V-N type and to the ranking of the V-N collocations based on their relative compositionality.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The main goals of the work presented in this paper are (1) To investigate a range of classifiers for their suitability in recognizing the non-compositional V-N collocations, and (2) To examine the relative compositionality of collocations of V-N type. Measuring the relative compositionality of V-N collocations is extremely helpful in applications such as machine translation where the collocations that are highly non-compositional can be handled in a special way.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Multi-word expressions (MWEs) are those whose structure and meaning cannot be derived from their component words, as they occur independently. Examples include conjunctions like 'as well as' (meaning 'including'), idioms like 'kick the bucket' (meaning 'die'), phrasal verbs like 'find out' (meaning 'search') and compounds like 'village community'. A typical natural language system assumes each word to be a lexical unit, but this assumption does not hold in case of MWEs [6] [12] . They have idiosyncratic interpretations which cross word boundaries and hence are a 'pain in the neck' [23] . They account for a large portion of the language used in day-to-day interactions [25] and so, handling them becomes an important task.", "cite_spans": [ { "start": 474, "end": 477, "text": "[6]", "ref_id": "BIBREF6" }, { "start": 478, "end": 482, "text": "[12]", "ref_id": "BIBREF12" }, { "start": 588, "end": 592, "text": "[23]", "ref_id": "BIBREF23" }, { "start": 676, "end": 680, "text": "[25]", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "A large number of MWEs have a standard syntactic structure but are noncompositional semantically. An example of such a subset is the class of noncompositional verb-noun collocations (V-N collocations). The class of V-N collocations which are non-compositional is important because they are used very frequently. These include verbal idioms [22] , support-verb constructions [1] [2] etc. The expression 'take place' is a MWE whereas 'take a gift' is not a MWE.", "cite_spans": [ { "start": 340, "end": 344, "text": "[22]", "ref_id": null }, { "start": 374, "end": 377, "text": "[1]", "ref_id": "BIBREF0" }, { "start": 378, "end": 381, "text": "[2]", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "It is well known that one cannot really make a binary distinction between compositional and non-compositional MWEs. They do not fall cleanly into mutually exclusive classes, but populate the continuum between the two extremes [4] . So, we rate the MWEs (V-N collocations in this paper) on a scale from 1 to 6 where 6 denotes a completely compositional expression, while 1 denotes a completely opaque expression. But, to address the problem of identification, we still need to do an approximate binary distinction. We call the expressions with a rating of 4 to 6 compositional and the expressions with rating of 1 to 3 as non-compositional. (See Section 4 for further details).", "cite_spans": [ { "start": 226, "end": 229, "text": "[4]", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Various statistical measures have been suggested for identification of MWEs and ranking expressions based on their compositionality. Some of these are Frequency, Mutual Information [9] , Log-Likelihood [10] and Pearson's \u03c7 2 [8] .", "cite_spans": [ { "start": 181, "end": 184, "text": "[9]", "ref_id": "BIBREF9" }, { "start": 202, "end": 206, "text": "[10]", "ref_id": "BIBREF10" }, { "start": 225, "end": 228, "text": "[8]", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Integrating all the statistical measures should provide better evidence for recognizing MWEs and ranking the expressions. We use various Machine Learning Techniques (classifiers) to integrate these statistical features and classify the V-N collocations as MWEs or Non-MWEs. We also use a classifier to rank the V-N collocations according to their compositionality. We then compare these ranks with the ranks provided by the human judge. A similar comparison between the ranks according to Latent-Semantic Analysis (LSA) based features and the ranks of human judges has been done by McCarthy, Keller and Caroll [19] for verb-particle constructions. (See Section 3 for more details). Some preliminary work on recognition of V-N collocations was presented in [28] .", "cite_spans": [ { "start": 610, "end": 614, "text": "[19]", "ref_id": "BIBREF19" }, { "start": 756, "end": 760, "text": "[28]", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In the task of classification, we show that the technique of weighted features in distance-weighted nearest-neighbour algorithm performs slightly better than other machine learning techniques. We also find that the 'distributed frequency of object (as defined by [27] )' and 'nearest mutual information (as adapted from [18] )' are important indicators of the non-compositionality of MWEs. In the task of ranking, we show that the ranks assigned by the classifier correlated much better with the human judgement than the ranks assigned by individual statistical measures.", "cite_spans": [ { "start": 263, "end": 267, "text": "[27]", "ref_id": "BIBREF27" }, { "start": 320, "end": 324, "text": "[18]", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "This paper is organised in the following sections (2) Basic Architecture, Now, to recognise the MWEs, the classifier has to do a binary classification of this vector. So, ideally, the classifier should take the above information and classify 'raise an eyebrow' as an MWE. The classifier can also be used to rank these vectors according to their relative compositionality.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Church and Hanks (1989) proposed a measure of association called Mutual Information [9] . Mutual Information (MI) is the logarithm of the ratio between the probability of the two words occurring together and the product of the probability of each word occurring individually. The higher the MI, the more likely are the words to be associated with each other. The usefulness of the statistical approach suggested by Church and Hanks [9] is evaluated for the extraction of V-N collocations from German text Corpora [7] . Several other measures like Log-Likelihood [10] , Pearson's \u03c7 2 [8] , Z-Score [8] , Cubic Association Ratio (MI3), Log-Log [17] , etc., have been proposed. These measures try to quantify the association of the two words but do not talk about quantifying the non-compositionality of MWEs. Dekang Lin proposes a way to automatically identify the non-compositionality of MWEs [18] . He suggests that a possible way to separate compositional phrases from non-compositional ones is to check the existence and mutual-information values of phrases obtained by replacing one of the words with a similar word. According to Lin, a phrase is probably non-compositional if such substitutions are not found in the collocations database or their mutual information values are significantly different from that of the phrase. Another way of determining the non-compositionality of V-N collocations is by using 'distributed frequency of object'(DFO) in V-N collocations [27] . The basic idea in there is that \"if an object appears only with one verb (or few verbs) in a large corpus we expect that it has an idiomatic nature\" [27] .", "cite_spans": [ { "start": 84, "end": 87, "text": "[9]", "ref_id": "BIBREF9" }, { "start": 432, "end": 435, "text": "[9]", "ref_id": "BIBREF9" }, { "start": 513, "end": 516, "text": "[7]", "ref_id": "BIBREF7" }, { "start": 562, "end": 566, "text": "[10]", "ref_id": "BIBREF10" }, { "start": 583, "end": 586, "text": "[8]", "ref_id": "BIBREF8" }, { "start": 597, "end": 600, "text": "[8]", "ref_id": "BIBREF8" }, { "start": 642, "end": 646, "text": "[17]", "ref_id": "BIBREF17" }, { "start": 892, "end": 896, "text": "[18]", "ref_id": "BIBREF18" }, { "start": 1473, "end": 1477, "text": "[27]", "ref_id": "BIBREF27" }, { "start": 1629, "end": 1633, "text": "[27]", "ref_id": "BIBREF27" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "3" }, { "text": "Schone and Jurafsky [24] applied Latent-Semantic Analysis (LSA) to the analysis of MWEs in the task of MWE discovery, by way of rescoring MWEs extracted from the corpus. An interesting way of quantifying the relative compositionality of a MWE is proposed by Baldwin, Bannard, Tanaka and Widdows [3] . They use latent semantic analysis (LSA) to determine the similarity between an MWE and its constituent words, and claim that higher similarity indicates great decomposability. In terms of compositionality, an expression is likely to be relatively more compositional if it is decomposable. They evaluate their model on English NN compounds and verb-particles, and showed that the model correlated moderately well with the Wordnet based decomposibility theory [3] .", "cite_spans": [ { "start": 20, "end": 24, "text": "[24]", "ref_id": "BIBREF24" }, { "start": 295, "end": 298, "text": "[3]", "ref_id": "BIBREF2" }, { "start": 759, "end": 762, "text": "[3]", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "3" }, { "text": "Evert and Krenn [11] compare some of the existing statistical features for the recognition of MWEs of adjective-noun and preposition-noun-verb types. Galiano, Valdivia, Santiago and Lopez [14] use five statistical measures to classify generic MWEs using the LVQ (Learning Vector Quantization) algorithm. In contrast, we do a more detailed and focussed study of V-N collocations and the ability of various classifiers in recognizing MWEs. We also compare the roles of various features in this task.", "cite_spans": [ { "start": 16, "end": 20, "text": "[11]", "ref_id": "BIBREF11" }, { "start": 188, "end": 192, "text": "[14]", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "3" }, { "text": "McCarthy, Keller and Caroll [19] judge compositionality according to the degree of overlap in the set of most similar words to the verb-particle and head verb. They showed that the correlation between their measures and the human ranking was better than the correlation between the statistical features and the human ranking. We have done similar experiments in this paper where we compare the correlation value of the ranks provided by the classifier with the ranks of the individual features for the V-N collocations. We show that the ranks given by the classifier which integrates all the features provides a significantly better correlation than the individual features.", "cite_spans": [ { "start": 28, "end": 32, "text": "[19]", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "3" }, { "text": "The data used for the experiments is British National Corpus of 81 million words. The corpus is parsed using Bikel's parser [5] and the Verb-Object Collocations are extracted. There are 4,775,697 V-N of which 1.2 million were unique. All the V-N collocations above the frequency of 100 (n=4405) are taken to conduct the experiments so that the evaluation of the system is feasible. These 4405 V-N collocations were searched in Wordnet, American Heritage Dictionary and SAID dictionary (LDC,2003). Around 400 were found in at least one of the dictionaries. Another 400 were extracted from the rest so that the evaluation set has roughly equal number of compositional and non-compositional expressions. These 800 expressions were annotated with a rating from 1 to 6 by using guidelines independently developed by the authors. 1 denotes the expressions which are totally non-compositional while 6 denotes the expressions which are totally compositional. The brief explanation of the various rating are (1) No word in the expression has any relation to the actual meaning of the expression. Example: \"leave a mark\". (2) Can be replaced by a single verb. Example : \"take a look\". (3) Although meanings of both words are involved, at least one of the words is not used in the usual sense. Example : \"break news\". (4) Relatively more compositional than (3). Example : \"prove a point\". (5) Relatively less compositional than (6). Example : \"feel safe\". (6) Completely compositional. Example : \"drink coffee\". For the experiments on classification (Section 7), we call the expressions with ratings of 4 to 6 as compositional and the expressions with rating of 1 to 3 as non-compositional. For the experiments on ranking the expressions based on their relative compositionality, we use all the 6 ratings to represent the relative compositionality of these expressions.", "cite_spans": [ { "start": 124, "end": 127, "text": "[5]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Data Used for the Experiments", "sec_num": "4" }, { "text": "The data was annotated by two fluent speakers of English. For 765 collocations out of 800, both the annotators gave a rating. For the rest, atleast one of the annotators marked the collocations as \"don't know\". Table 1 illustrates the details of the annotations provided by the two judges. From the table we see that annotator1 distributed the rating more uniformly among all the collocations while annotator2 observed that a significant proportion of the collocations were completely compositional. To measure the agreement between the two annotators, we used the Kendall's TAU (\u03c4 ). \u03c4 is the correlation between the rankings 1 of collocations given by the two annotators. W ranges between 0 (little agreement) and 1 (full agreement). W is calculated as below,", "cite_spans": [], "ref_spans": [ { "start": 211, "end": 218, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Agreement Between the Judges", "sec_num": "5" }, { "text": "\u03c4 = iRatings654321 Compositional Non-Compositional(4 to 6)(1 to 3)Annotator1 141 122 127 119 161 95390375Annotator2 303 8879 101 118 76470195", "type_str": "table", "num": null, "html": null }, "TABREF1": { "text": "List of features and their top-3 example collocations", "content": "
FeatureTop-3FeatureTop-3
take placeMutual Information shrug shoulder
Frequencyhave effect[9]bridge gap
have timeplead guilty
Cubic Associationtake placeLog-Logshake head
Measureshake head[17]commit suicide
(Oakes, 1998)play rolefall asleep
Log-Likelihoodtake placePearson's \u03c7 2
", "type_str": "table", "num": null, "html": null }, "TABREF2": { "text": "Average accuracies of MWE recognition using simple nearest-neighbour algorithms and weighted distance nearest neighbour algorithms", "content": "
Simple K-Nearest neighbourWeighted-distance Nearest neighbour
TypeK=1K=2K=3K=1K=2K=3K=All
Annot.1 62.3561.3162.4862.3562.3562.6166.66
Annot.2 57.6454.1060.8957.6457.6460.3763.52
", "type_str": "table", "num": null, "html": null }, "TABREF3": { "text": "Average accuracies of MWE recognition using SVMs (Linear, Polynomial and Radial Basis Function Kernel)", "content": "
Linear Ker. Polynomial Ker.Radial Basis networks
Parameters\u03c3 = 0.5 \u03c3 = 1.0 \u03c3 = 1.5 \u03c3 = 2.0
Annot.165.8965.7567.0666.6666.9367.06
Annot.262.6165.0964.1763.5162.9962.99
", "type_str": "table", "num": null, "html": null }, "TABREF4": { "text": "The correlation values of the ranking of individual features and the ranking of classifier with the ranking of human judgements", "content": "
MI-0.125 Z-Score -0.059
MI30.001 \u03c6-coeff -0.102
Log-Log-0.086 DFO -0.113
Log-Likelihood 0.005NMI -0.167
\u03c7 2-0.056 Class. 0.388
T-Score0.045
", "type_str": "table", "num": null, "html": null } } } }