From: AAAI-00 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights reserved. The Automatic Interpretation of Nominalizations Maria Lapata Division of Informatics, University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, UK [email protected] Abstract compound satellite observation which may mean observa- This paper discusses the interpretation of nominalizations in tion by satellite or observation of satellites. In order to trans- domain independent wide-coverage text. We present a statis- late satellite observation into Spanish, we have to work out tical model which interprets nominalizations based on the co- whether satellite is the subject or object of the verb ob- occurrence of verb-argument tuples in a large balanced cor- serve. In the first case satellite observation translates as ob- pus. We propose an algorithm which treats the interpretation servaci´on por satelite (observation by satellite), whereas in task as a disambiguation problem and achieves a performance the latter it translates as observaci´on de satelites (observa- of approximately 80% by combining partial parsing, smooth- ing techniques and domain independent taxonomic informa- tion of satellites). tion (e.g., WordNet). A considerable amount of work within NLP focused on the interpretation of two word compounds whose nouns are related via a basic set of semantic relations (e.g., CAUSE re- Introduction lates onion tears, FOR relates pet spray). With the exceptions The automatic interpretation of compound nouns has been of Wu (1993) and Lauer (1995) who have proposed proba- a long-standing unsolved problem within Natural Language bilistic models for the interpretation of compounds, the ma- Processing (NLP). Compound nouns in English have three jority of proposals are symbolic: most algorithms rely on basic properties which pose difficulties for their interpreta- hand-crafted knowledge bases or dictionaries containing de- tion: (a) the compounding process is extremely productive, tailed semantic information for each noun; a sequence of (b) the semantic relationship between the compound head rules exploit this information in order to choose the cor- and its modifier is implicit (this means that it cannot be eas- rect interpretation for a given compound (Leonard 1984; ily recovered from syntactic or morphological analysis), and Vanderwende 1994). Most of the proposals contain no qual- (c) the interpretation can be influenced by a variety of con- itative evaluation. The exceptions are Leonard (1984) who textual and pragmatic factors. reports an accuracy of 76% (although on the training set), To arrive at an interpretation of the compound onion tears Vanderwende (1994) whose algorithm attains an accuracy (e.g., onions CAUSE tears) it is necessary to identify that of 52%, and Lauer (1995) who reports an accuracy of 47%. tears is a noun (and not the third person singular of the verb The low accuracy is indicative of the difficulty of the task tear) and to use semantic information about onions and tears given the variety of contextual and pragmatic factors which (for example the fact that onions cannot be tears or that tears can influence the interpretation of a compound. are not made of onions). Even in the case of a compound In this paper, we focus solely on the interpretation of nom- like government promotion where the head noun is derived inalizations, i.e., compounds whose head noun is a nominal- from the verb promote and the modifier government is its ar- ized verb and whose prenominal modifier is derived from ei- gument, it is necessary to determine whether government is ther the underlying subject or direct object of the verb (Levi the subject or the object. One might argue that the preferred 1978) (see the examples in (2)–(3)). analysis for government promotion is “government that is (2) a. SUBJ child behaviour ⇒ child behaves promoted by someone”. However, this interpretation can be b. OBJ car lover ⇒ love cars easily overridden in context as shown in example (1) taken c. OBJ soccer competition ⇒ compete in soccer from the British National Corpus: here it is the government (3) a. SUBJ|OBJ government promotion that is doing the promotion. b. SUBJ|OBJ satellite observation (1) By the end of the 1920s, government promotion of agricul- The nominalized verb can either take a subject (cf. (2a)), a tural development in Niger was limited, consisting mainly of crop trials and model sheep and ostrich farm. direct object (cf. (2b)) or a prepositional object (cf. (2c)). In some cases, the relation of the modifier and the nominalized The interpretation of compound nouns is important for sev- verb (SUBJ or OBJ) can be predicted either from the subcat- eral NLP tasks, notably machine translation. Consider the egorization properties of the verb (cf. (2a) where child can Copyright c 2000, American Association for Artificial Intelli- only be the subject of the intransitive verb behave) or from gence (www.aaai.org). All rights reserved. the semantics of the of the nominalization suffix of the head noun (cf. (2b) where the agentive suffix -er of the noun lover ney 1996). Cass is a robust chunk parser designed for the indicates that the modifier car is the object of love). In other shallow analysis of noisy text. We used the parser’s built- cases, the relation of the modifier and the head noun is gen- in function to extract verb-subject and verb-object tuples. uinely ambiguous (see (3)). The tuples obtained from the parser’s output are an imper- The interpretation of nominalizations poses a challenge fect source of information about argument relations. For for empirical approaches since the argument relations be- example, the tuples extractor mistakes adjectives for verbs tween a head and its modifier are not readily available (cf. (7a)) and nouns for verbs (cf. (7b)). in the corpus. We present a probabilistic algorithm which (7) a. SUBJ isolated people treats the interpretation task as a disambiguation problem, b. SUBJ behalf whose and show how the severe sparse data problem in this task In order to compile a comprehensive count of verb-argument can be overcome by combining partial parsing, smoothing relations we discarded tuples containing verbs or nouns with techniques, and domain independent taxonomic information a BNC frequency of one. This resulted in 588,333 distinct (e.g., WordNet). We report on the results of five experiments types of verb-subject pairs and 615,328 distinct types of which achieve a combined precision of approximately 80% verb-object pairs. on the British National Corpus (BNC), a 100 million word collection of samples of written and spoken language from The data a wide range of sources designed to represent a wide cross- It is beyond the scope of the present study to develop an section of current British English, both spoken and written algorithm which automatically detects nominalizations in a (Burnard 1995). corpus. In the experiments described in the subsequent sec- tions compounds with deverbal heads were obtained as fol- The model lows: Given a nominalization, our goal is to develop a procedure 1. Two word compound nouns were extracted from the BNC by to infer whether the modifier stands in a subject or object using a heuristic which looks for consecutive pairs of nouns relation to the head noun. In other words, we need to as- which are neither preceded nor succeeded by a noun (Lauer sign probabilities to the two different relations (SUBJ, OBJ). 1995). For each relation rel we calculate the simple expression 2. A dictionary of deverbal nouns was created using: (a) Nomlex P(rel|n1;n2) given in (4) below. (Macleod et al. 1998), a dictionary of nominalizations con- ( ; ; ) taining 827 lexical entries and (b) Celex (Burnage 1990), a ( | ; )= f n1 rel n2 (4) P rel n1 n2 ( ; ) general morphological dictionary, which contains 5,111 nom- f n1 n2 inalizations; Since we have a choice between two outcomes we will use 3. Candidate nominalizations were obtained from the com- a likelihood ratio to compare the two relation probabilities pounds acquired from the BNC by selecting noun-noun se- (Mosteller & Wallace 1964). In particular we will compute quences whose head (i.e., rightmost noun) was one of the de- the log of the ratio of the probability P(OBJ|n1;n2) to the verbal nouns contained either in Celex or Nomlex. The pro- cedure resulted in 172,797 potential nominalizations. probability P(SUBJ|n1;n2). We will call this log-likelihood ratio the argument relation (RA)score. From these candidate nominalizations a random sample of ( | ; ) ( ; ; )= P OBJ n1 n2 2,000 tokens was selected. The sample was manually in- (5) RA rel n1 n2 log2 P(SUBJ|n1;n2) spected and compounds with modifiers whose relation to the head noun was other than subject or object were discarded. Notice, however, that we cannot read off f (n1;rel;n2) di- rectly from the corpus. What we can obtain from a corpus Nominalizations whose heads were derived from verbs tak- (through parsing) is the number of times a noun is the object ing prepositional objects (cf. (2c)) were also discarded. Af- or the subject of a given verb. By making the simplifying ter manual inspection the sample contained 796 nominal- assumption that the relation between the nominalized head izations. From these, 596 tokens were used as training data and its modifier noun is the same as the relation between the for the experiments described in the following sections. The latter and the verb from which the head is derived, (4) can remaining 200 nominalizations were used as test data and be rewritten as follows: also to evaluate whether humans can reliably disambiguate f (v ;rel;n ) the argument relation between the nominalized head and its (6) P(rel|n ;n ) ≈ n2 1 1 2 ∑ ( ; ; ) modifier.