2.1 The process of compounding

The Dictionary of Grammatical Terms in Linguistics defines the process of compounding as follows (Trask Reference Trask1993):

Since the process of compounding constructs new words, these words can in turn combine with other words to form longer compounds, and this process can be repeated indefinitely, e.g., orange juice, orange juice company, orange juice company homepage, orange juice company homepage logo, orange juice company homepage logo update, etc. Compounds of length longer than two are less frequent in general, and are more typical for technical and scientific texts.

In English, the process of compounding can combine together words belonging to various parts of speech, e.g., adjective+adjective (e.g., dark-green, light-blue), adjective+adverb (e.g., leftmost), adjective+noun (e.g., hot dog, shortlist, white collar, highlife), adjective+preposition (e.g., forthwith), adjective+verb (e.g., highlight, broadcast, quick-freeze, dry-clean), noun+adjective (e.g., trigger-happy, army strong, bulletproof, dog tired, English-specific, brand-new), noun+noun (e.g., silkworm, honey bee, bee honey, stem cell), noun+preposition (e.g., love-in, timeout, breakup), noun+verb (e.g., finger-point, taperecord), preposition+adjective (e.g., overeager, over-ripe), preposition+noun (e.g., underwater, indoor), preposition+preposition (e.g., within, without, into, onto), preposition+verb (e.g., overestimate, withdraw, upgrade, withhold), verb+adverb (e.g., tumbledown), verb+noun (e.g., pickpocket, cutthroat, know-nothing), verb+preposition (e.g., countdown, stand-by, cut-off, castaway), and verb+verb (e.g., freeze-dry). More complex structures are also possible, e.g., state-of-the-art, part-of-speech or over-the-counter eye drop. Note that not all of these compound types are productive in modern English.

The part of speech of an English compound is typically that of its last word, e.g., hot dog is a noun since its second word is the noun dog, while broadcast is a verb because it ends with the verb cast.Footnote ² Some classes of compounds do not follow this rule though, e.g., noun+preposition and verb+preposition yield complex nouns instead of complex prepositions, e.g., timeout and countdown.

The most typical compounds in English are those composed of nouns only, known as noun compounds, and the most frequent among them are those of type noun+noun (because they are the shortest). Yet, not every sequence of nouns is considered to be a noun compound.

There is little agreement in the research community on how to define the notion of noun compound.Footnote ³ Different authors use different definitions, focusing on different aspects, and often use different terms in order to emphasize particular distinctions. Lauer (Reference Lauer1995) provides the following list of closely related notions used in the literature: compound nominal, nominal compound, compound noun, complex nominal, nominalization, noun sequence, compound, noun compound, noun–noun compound, noun+noun compound, noun premodifier. While some of these terms are broader and some are narrower in scope, most of them refer to objects that are syntactically analyzable as nouns (Chomsky and Halle Reference Chomsky and Halle1968; Jackendoff Reference Jackendoff1975).

One popular definition is that of Downing (Reference Downing1977), who defines the notion of a nominal compound as a sequence of nouns which function as a single noun, e.g., orange juice, company tax policy, law enforcement officer, colon cancer tumor suppressor protein, etc. This definition is both simple and relatively unambiguous, which makes it the preferred choice for linguists working on syntax as well as for computational linguists. The primary difficulty with using it as a decision criterion for extracting noun–noun compounds in the absence of a grammar is the arbitrariness of the adjective/noun distinction for some premodifiers, e.g., are adult and male nouns or adjectives in adult male rat? Or, if dental decay and tooth decay are synonymous, should we then think of tooth as an adjective in that context? See also Huddleston and Pullum (Reference Huddleston and Pullum2002) page 556 for a relevant discussion.

In view of these issues, some researchers have proposed a broader definition that threats some adjectival and noun modifiers alike. For example, Levi (Reference Levi1978) advocates the concept of complex nominals, which groups three partially overlapping classes: nominal compounds (e.g., doghouse, deficiency disease, apple cake), nominalizations (e.g., American attack, presidential refusal, dream analysis), and nonpredicate NPs Footnote ⁴ (e.g., electric shock, electrical engineering, musical criticism). She argues that these subclasses share important syntactic and semantic properties. For example, the nonpredicate NP linguistic difficulties is arguably synonymous with the nominal compound language difficulties.

Many alternative definitions exist, based on various criteria, including orthographic, phonological, morphological, syntactic and semantic, none of which can clearly distinguish between noun compounds and other nominal phrases, but each pointing to some important properties shared by most compounds. Below we will explore some of these criteria, focusing on noun compounds; still, we will occasionally discuss compounds whose elements belong to other parts of speech.

2.2 Types of compounds

The two most important types of noun compounds are endocentric and exocentric. The Lexicon of Linguistics Footnote ⁷ defines them as follows:

There is a third category of compounds, known as copulative or coordinative (dvandva in Sanskrit). They form an entity that is the sum of the two nouns that form the compound and is also distinct from either of them. Copulative compounds are rare in English; examples include mostly hyphenated proper nouns, e.g., Austria–Hungary and Bosnia–Herzegovina, but also some common nouns, e.g., gerund-participle. However, they are very common in Indic languages (e.g., Sanskrit, Hindi, Urdu, Marathi, and Tamil), as well as in Chinese and Japanese.

Copulative compounds are often confused with a fourth category of compounds known as appositional, where each of the nouns denotes independently a different aspect of the entity that the compound represents, e.g., coach-player, sofa-bed, writer-director, learner-driver, and programmer analyst. For example, coach-player is somebody who is both a coach and a player.

Noun compounds can also be formed by reduplication. There are various kinds of reduplication in English such as exact (e.g., bye-bye), ablaut (e.g., chit-chat) rhyming (e.g., walkie-talkie, hokey-pokey), sh-reduplication (e.g., baby-shmaby), contrastive (e.g., I’ll make the tuna salad, and you make the salad-salad.), etc. Despite this variety, reduplication is not very common nor it is very productive in English, except probably for the last two categories. However, it is very common in some other languages, e.g., in Malay it is used, among many other things, for forming plural, e.g., ayam-ayam Footnote ⁸ ‘chickens’ is the plural of ayam ‘chicken’.

Finally, there are portmanteaux compounds, which are composed by blending the sounds of two or more words, while also combining their meanings. For example, brunch is formed from breakfast + lunch, while Eurasia blends together Europe and Asia. A more recent example is Merkozy, which blends Merkel and Sarkozy. Portmanteaux are frequent in Russian, where they used to be the preferred choice for Soviet-style communist and propaganda terms, e.g., komsomol is made of kommunisticheskiy soyuz molodyozhi, meaning ‘Young Communist League’, or sovkhoz, from sovetskoe khozyaistvo, meaning ‘Soviet farm’.

2.3 Properties of compounds

3.1 Adjacency model

According to the bracketing representation introduced above, given a three-word noun compound w ₁w ₂w ₃, the task is to decide whether w ₂ is more closely associated with w ₁ or with w ₃. Therefore, it looks natural to compare the strength of association between the first two and the last two words, which yields an adjacency model (Lauer Reference Lauer1995):

• • if Assoc(w ₁, w _w) < Assoc(w _w, w _w), predict right bracketing;

• • if Assoc(w _w, w _w) = Assoc(w _w, w _w), make no prediction;

• • if Assoc(w _w, w _w) > Assoc(w _w, w _w), predict left bracketing.

3.2 Dependency model

Lauer (Reference Lauer1995) proposed an alternative, syntactically motivated dependency model. It is illustrated in Figure 2, where arcs point from the heads to their modifiers.Footnote ⁹

Fig. 2.

Left versus right bracketing: dependency structures.

In this representation, both the left and the right dependency structures contain a link w ₂ ← w ₃, but differ because of w ₁ ← w ₂ and w ₁ ← w ₃, respectively.Footnote ¹⁰ Therefore, the dependency model focuses not on w ₂ but on w ₁:

• • if Assoc(w _w, w _w) < Assoc(w _w, w _w), predict right bracketing;

• • if Assoc(w _w, w _w) = Assoc(w _w, w _w), make no prediction;

• • if Assoc(w _w, w _w) > Assoc(w _w, w _w), predict left bracketing.

3.3 Adjacency versus dependency

Figure 3 shows the dependency trees corresponding to the dependency structures from Figure 2. Note the structural difference from the constituency trees in Figure 1. First, the constituency trees contain words in the leaves only, while the dependency trees have words in the internal nodes as well. Second, the constituency trees are ordered binary trees: each internal node has exactly two ordered descendants, one left and one right, while there is no such ordering for the dependency trees.

Fig. 3.

Left versus right bracketing: dependency trees.

Consider also examples (3) and (4), which are both right-bracketed, but the order of the first two words is switched (we consider adult and male as nouns here).

1. (3) [adult [male rat]] (right bracketing)

2. (4) [male [adult rat]] (right bracketing)

Despite (3) and (4) being different, the corresponding dependency structures are equivalent as Figures 4 and 5 show: there are no dependency arcs between adult and male, and therefore changing their linear order does not alter the dependency structure.

Fig. 4.

Dependency structures for adult male rat and for male adult rat.

Fig. 5.

Shared dependency tree for adult male rat and for male adult rat.

The adjacency and the dependency models target different kinds of right-bracketed structures. Consider for example home health care: health care is a compound, which in turn is modified by home as a whole, which can easily be seen in the concatenated spelling home healthcare, where the last two words are written without separating space.

This is different from adult male rat, where the order of the modifiers adult and male could be switched, which suggests that adult and male independently modify rat. We can conclude that, given a three-word noun compound w ₁w ₂w ₃, there are two reasons it may be right-bracketed, [w ₁[w ₂w ₃]]:

1. (a) w ₂w ₃ is a compound, which is modified by w ₁;

2. (b) w ₁ and w ₂ independently modify w ₃.

Let us now look closely at the adjacency and the dependency models:

• • adjacency model: compare Assoc(w ₁, w ₂) and Assoc(w _w, w _w);

• • dependency model: compare Assoc(w ₁, w ₂) and Assoc(w _w, w _w).

We can see that the adjacency model checks (a), i.e., whether w ₂w ₃ is a compound (a test for lexicalization), while the dependency model checks (b), i.e., whether w ₁ modifies w ₃ (a test for syntactic modification).

Note that there is only a modificational choice in case of left bracketing. If w ₁ modifies w ₂, then w ₁w ₂ is a noun compound, which now acts as a single noun to modify w ₃. Consider for example law enforcement agent: law is a modifier of enforcement, together they form a noun compound law enforcement, which in turn, as a whole, modifies agent.

3.4 Frequency-based association scores

The simplest association score measures the strength of association as a bigram frequency in some corpus:

(1) \begin{equation} Assoc(w_i,w_j) = \#(w_i,w_j) \end{equation}

Another association score is based on conditional probability:

(2) \begin{equation} Assoc(w_i,w_j) = \mathrm{Pr}(w_i \leftarrow w_j|w_j) \end{equation}

(3) \begin{equation} \mathrm{Pr}(w_i \leftarrow w_j|w_j) = \frac{\#(w_i,w_j)}{\#(w_j)} \end{equation}

Other popular association measures, based on similar unigram and bigram statistics, include pointwise mutual information and Chi square. The n-gram statistics are typically collected from a large text corpus or from the Web.

For example, Lauer (Reference Lauer1995) collected n-gram statistics from Grolier's Encyclopedia,Footnote ¹¹ which contained about eight million words (in 1995). To overcome data sparseness issues, he estimated the probabilities over conceptual categories t_i in Roget's Thesaurus Footnote ¹² rather than for individual words.

For evaluation, he created a dataset of 244 three-word noun compounds extracted from Grolier's Encyclopedia; this is the benchmark testset for the task of noun compound bracketing. Most compounds in this dataset are left-bracketed: 66.80%. Lauer achieved 68.90% and 77.50% accuracy with the adjacency and the dependency models, respectively. He further reported 80.70% with a better tuned system.

More recently, Lapata and Keller (Reference Lapata and Keller2004, Reference Lapata and Keller2005) estimated n-gram counts from the 100M-word BNC and from AltaVista, using page hit counts as a proxy for n-gram frequencies. They reported 68.03% accuracy with BNC and 78.68% accuracy with AltaVista.

Girju et al. (Reference Girju, Moldovan, Tatu and Antohe2005) used fifteen semantic features derived from WordNet (Fellbaum Reference Fellbaum1998), achieving 83.10% accuracy. Unlike the above approaches, their approach is supervised and requires that each noun be annotated with the correct WordNet sense, which is an unrealistic assumption.

Nakov and Hearst (Reference Nakov and Hearst2005a) used Web-derived surface features and paraphrases,Footnote ¹³ achieving 89.34% accuracy on the Lauer dataset, and 95.35% accuracy on their own biomedical dataset.Footnote ¹⁴ Their most important features include the following:

• • Dashes: e.g., finding on the Web the dashed form law-enforcement officer suggests a left bracketing interpretation for law enforcement officer.

• • Genitives: e.g., finding the genitive form math problem's solution indicates a left bracketing for math problem solution.

• • Internal capitalization: e.g., finding on the Web Plasmodium vivax Malaria suggests left bracketing for plasmodium vivax malaria.

• • Internal inflection variety: e.g., bone minerals density favors left bracketing for bone mineral density.

• • Internal slashes: e.g., leukemia/lymphoma cell suggests right bracketing for leukemia lymphoma cell.

• • Parentheses: e.g., finding on the Web growth factor (beta) indicates left bracketing for growth factor beta.

• • Abbreviations: e.g., finding tumor necrosis factor (NF) signals right bracketing for tumor necrosis factor.

• • Concatenations: e.g., finding on the Web healthcare reform favors left bracketing for health care reform.

• • Word reordering: e.g., finding male adult rat indicates right bracketing for adult male rat.

• • Paraphrases: They can indicate the bracketing structure by showing which two words would be kept together, e.g., cells in the bone marrow suggests left bracketing for bone marrow cell since it keeps the two left words together.

4.1 Representation

Most work on noun compound semantics has focused on two-word noun compounds, or noun–noun compounds, which are the most common type. In terms of representation, in computational linguistics, the semantics of a noun compound has been typically expressed by an abstract relation like CAUSE (e.g., malaria mosquito), SOURCE (e.g., olive oil), or PURPOSE (e.g., migraine drug), drawn from a small fixed inventory. Such inventories were initially proposed by linguists, e.g., based on corpus studies, but they have become especially popular in computational linguistics.

Warren (Reference Warren1978) proposed a linguistic theory based on a study of the Brown corpus (Kucera and Francis Reference Kucera and Francis1967). In this theory, noun compounds are characterized by abstract semantic relations organized into a four-level hierarchy, where the top level is occupied by six coarse-grained relations, Possession, Location, Purpose, Activity-Actor, Resemblance, and Constitute, which are further subdivided into finer-grained relations. For example, Constitute is divided into three level-2 relations: Source-Result, Result-Source, or Copula. Copula in turn is subdivided into another three level-3 relations: Adjective-Like_Modifier, Subsumptive, and Attributive. Finally, Attributive is divided into two level-4 relations Animate_Head (e.g., girl friend) and Inanimate_Head (e.g., house boat).

In the alternative linguistic theory of Levi (Reference Levi1978), noun compounds (and complex nominals in general, see Section 2.1) can be derived by the following two processes:

1. (1) Predicate deletion.

   It can delete the 12 abstract recoverably deletable predicates (RDPs) shown in Table 1, e.g., pie made of apples → apple pie. In the resulting nominals, the modifier is typically the object of the predicate; when it is the subject, the predicate is marked with the index 2.

2. (2) Predicate nominalization. It produces nominals whose head is a nominalized verb, and whose modifier is derived from either the subject or the object of the underlying predicate, e.g., the President refused General MacArthur's request→ presidential refusal. Multi-modifier nominalizations retaining both the subject and the object as modifiers are possible as well. Therefore, there are three types of nominalizations depending on the modifier, which are combined with the following four types of nominalizations the head can represent: act, product, agent, and patient. See Table 2 for examples.

Table 1.

Levi's recoverably deletable predicates (RDPs). Column 3 shows the modifier's function in the corresponding paraphrasing relative clause: when the modifier is the subject of that clause, the RDP is marked with the index 2

Table 2.

Levi's nominalization types with examples

These linguistic theories have influenced the efforts to create similar inventories for computational linguistic purposes. For example, Ó Séaghdha (Reference Ó Séaghdha2007) revised the inventory of Levi (Reference Levi1978) based on six pragmatic criteria:

• • the inventory of relations should have good coverage;

• • relations should be disjunct, and should describe a coherent concept;

• • the class distribution should not be overly skewed or sparse;

• • the concepts underlying the relations should generalize to other linguistic phenomena;

• • the guidelines should make the annotation process as simple as possible;

• • the categories should provide useful semantic information.

This yielded an inventory of six semantic relations, which are further sub- divided into subcategories: BE (identity, substance-form, similarity), HAVE (possession, condition-experiencer, property-object, part-whole, groupmember), IN (spatially located object, spatially locatedevent, temporarily located object, temporarily located event), ACTOR (participant-event, participant-participant), INST (participant-event, participant-participant), ABOUT (topic-object, topic-collection, focus-mental activity, commodity-charge). E.g., tax law is topic-object, crime investigation is focusmental activity, and they both are also ABOUT.

The approach of Warren (Reference Warren1978) to organize fine-grained semantic relations into a multi-level hierarchy has been followed by Nastase and Szpakowicz (Reference Nastase and Szpakowicz2003) who designed an inventory of 30 fine-grained relations, grouped into five coarse-grained super-relations (the corresponding fine-grained relations are shown in parentheses): CAUSALITY (cause, effect, detraction, purpose), PARTICIPANT (agent, beneficiary, instrument, object_property, object, part, possessor, property, product, source, whole, stative), QUALITY (container, content, equative, material, measure, topic, type), SPATIAL (direction, location_at, location_from, location), and TEMPORALITY (frequency, time_at, time_through). For example, exam anxiety is classified as effect and therefore also as CAUSALITY, while blue book is property and therefore also PARTICIPANT.

This inventory of Nastase and Szpakowicz (Reference Nastase and Szpakowicz2003) is an extension of the flat, nonhierarchical inventory of 20 relations created by Barker and Szpakowicz (Reference Barker and Szpakowicz1998). It is also a superset of the 19-relation inventory of Kim and Baldwin (Reference Kim and Baldwin2006).

There have also been some independently developed inventories, which are not directly or indirectly derived from those of Levi (Reference Levi1978) and Warren (Reference Warren1978).

One such example is the inventory of Vanderwende (Reference Vanderwende1994), which classifies noun–noun compounds according to questions (mostly Wh-questions), which in turn correspond to 13 semantic relations: Subject (Who/what?), Object (Whom/what?), Locative (Where?), Time (When?), Possessive (Whose?), Whole-Part (What is it part of?), Part-Whole (What are its parts?), Equative (What kind of?), Instrument (How?), Purpose (What for?), Material (Made of what?), Causes (What does it cause?), and Caused-by (What causes it?). For example, alligator shoes is Material and answers the question Made of what?.

Another example is the inventory of Girju et al. (Reference Girju, Moldovan, Tatu and Antohe2005), which consists of 21 abstract relations: POSSESSION, ATTRIBUTE-HOLDER, AGENT, TEMPORAL, PART-WHOLE, IS-A, CAUSE, MAKE/PRODUCE, INSTRUMENT, LOCATION/SPACE, PURPOSE, SOURCE, TOPIC, MANNER, MEANS, THEME, ACCOMPANIMENT, EXPERIENCER, RECIPIENT, MEASURE, and RESULT. It is a subset of the 35-relation inventory of Moldovan et al. (Reference Moldovan, Badulescu, Tatu, Antohe and Girju2004), originally designed for the semantic interpretation of noun phrases in general.

Finally, there has been some work on trying to compare and consolidate existing inventories. Tratz and Hovy (Reference Tratz and Hovy2010) compared the inventories of Levi (Reference Levi1978), Warren (Reference Warren1978), Vanderwende (Reference Vanderwende1994), Barker and Szpakowicz (Reference Barker and Szpakowicz1998), Nastase and Szpakowicz (Reference Nastase and Szpakowicz2003), and Girju et al. (Reference Girju, Moldovan, Tatu and Antohe2005) and found them to cover a largely overlapping semantic space but to partition this space differently. Thus, they proposed a new two-level inventory of 43 fine-grained relations, grouped into ten coarse-grained super-relations; Tratz and Hovy obtained this inventory through iterative crowd-sourcing by maximizing inter-annotator agreement.

While the above abstract relation inventories were designed with broad-coverage text analysis in mind, there have been also proposals tailored to a specific domain. For example, Rosario, Hearst and Fillmore (Reference Rosario, Hearst and Fillmore2002) defined 38 abstract relations to characterize the semantics of biomedical noun–noun compounds, which includes domain-specific relations such as Produce_Genetically, Person_Afflicted, Attribute_of_Clinical_Study, and Defect. For example, hormone deficiency is Defect, and polyomavirus genome is Produce_Genetically.

Another interesting proposal is that of Lauer (Reference Lauer1995), who defined the problem of noun compound interpretation as predicting which among the following eight prepositions best paraphrases the target noun compound: of, for, in, at, on, from, with, and about. For example, olive oil is oil from olives, and war story is story about war. Lauer's approach is attractive since it is simple and allows direct extraction of prepositions from a large text corpus or from the Web (Lapata and Keller Reference Lapata and Keller2004, Reference Lapata and Keller2005). Moreover, prepositions are directly usable as paraphrases in NLP applications. The downside is that prepositions are polysemous: for example, in, on, and at can refer to both LOCATION and TIME, e.g., mouse in the field and flight in the morning express different relations, while flight in the morning, flight at night, and flight on Saturday express the same relation. Similarly, from can correspond to several of Levi's relations (see Table 1), e.g., tears from gas is CAUSE₁, death from drugs is CAUSE₂, cake from apples is HAVE₁, peel from lemon is HAVE₂, silk from worms is MAKE₁, ball from snow is MAKE₂, mouse from the fields is IN, and oil from olives is FROM. Note also that some noun–noun compounds cannot be paraphrased with prepositions, e.g., coach trainer; in contrast, empirical work suggests that most noun–noun compounds do fall into one of Levi's RDPs.

The interpretation of noun compound in terms of abstract semantic relations drawn from a small fixed inventory is also potentially problematic for a number of reasons. First, it is unclear which relation inventory is best to use in general and for a particular task. Second, being both abstract and limited, such relations capture only part of the semantics, e.g., classifying malaria mosquito as CAUSE obscures the fact that mosquitos do not directly cause malaria, but just transmit it. Third, in many cases, multiple relations are possible, e.g., in Levi's theory, sand dune is interpretable as both HAVE and BE. Still, despite these issues, the use of abstract relation inventories remains the mainstream approach in computational linguistics.

In theoretical linguistics, however, the feasibility of the idea has been subject to much criticism. For example, after having performed a series of psychological experiments, Downing (Reference Downing1977) concluded that no fixed and limited set of relations could adequately characterize noun compound semantics, a position held earlier by Jespersen (Reference Jespersen1942). To see this, consider a noun compound like plate length, which can be interpreted in some context as what your hair is when it drags in your food. Obviously, such fine-grained novel interpretations cannot be covered by a finite fixed inventory of abstract semantic relations. This has led some computational linguists such as Hobbs et al. (Reference Hobbs, Stickel, Appelt and Martin1993) to assume that the relationship between the nouns in a noun–noun compound can be anything.

Some researchers have proposed a partial solution based on finer-grained, and even infinite, inventories, e.g., Finin (Reference Finin1980) used specific verbs such as dissolved in for salt water. While this still cannot handle examples such as plate length, it offers much richer representation.

Inspired by Finin (Reference Finin1980), Nakov and Hearst (Reference Nakov, Hearst, Euzenat and Domingue2006) proposed that noun compound semantics is best expressible using paraphrases involving verbs and/or prepositions; this work was further extended in Nakov (Reference Nakov2008b). For example, bronze statue is a statue that is made of, is composed of, consists of, contains, is of, is, is handcrafted from, is dipped in, looks like bronze.

Nakov and Hearst (Reference Nakov, Hearst, Euzenat and Domingue2006) further proposed that selecting one such paraphrase is not enough to express the semantics of a noun compound and that multiple paraphrases are needed for a fine-grained representation. Finally, they observed that not all paraphrases are equally good (e.g., is made of is arguably better than looks like or is dipped in for Make), and thus proposed that the semantics of a noun compound should be expressed as a distribution over multiple possible paraphrases, e.g., malaria mosquito can be carry (23), spread (16), cause (12), transmit (9), etc. These verbs are fine-grained, directly usable as paraphrases, and using multiple of them for a noun compound approximates its semantics better.

It is easy to see that not only the semantics of individual noun compounds but also that of abstract relations such as Make can be represented in the same way, as a distribution over paraphrasing verbs and prepositions; this is an idea explored in Nakov and Hearst (Reference Nakov and Hearst2008). Note, however, that some noun compounds are paraphrasable by more specific verbs that do not necessarily support the target abstract relation. For example, malaria mosquito, which expresses Cause, can be paraphrased using verbs like carry, which do not imply direct causation. Similarly, while both wrinkle treatment and migraine treatment express Treatment-For-Disease, fine-grained differences can be revealed using verbs, e.g., smooth can paraphrase the former, but not the latter.

Finally, note that while verbs and prepositions are the most frequent ways to paraphrase a noun–noun compound, there are many other ways to express the explicit relationship between the two nouns, e.g., for onion tears, we could have: tears from onions, tears due to cutting onion, tears induced when cutting onions, tears that onions induce, tears that come from chopping onions, tears that sometimes flow when onions are chopped, tears that raw onions give you, etc. The exploration of such free paraphrases of noun–noun compounds was the focus of SemEval-2013 Task 4 (Hendrickx et al. Reference Hendrickx, Kozareva, Nakov, Ó Séaghdha, Szpakowicz and Veale2013).Footnote ¹⁵

4.2 Semantic interpretation

The semantic interpretation of noun compounds is complicated by their heterogeneous nature. Thus, it is often addressed using nonparametric instance-based classifiers like the k-nearest neighbor (kNN), which effectively reduce it to measuring the relational similarity between a testing and each of the training examples. The latter is studied in detail by Turney (Reference Turney2006b), who distinguishes between attributional similarity or correspondence between attributes, and relational similarity or correspondence between relations. Attributional similarity is interested in the similarity between two words, A and B. In contrast, relational similarity focuses on the relationship between two pairs of words, i.e., it asks how similar the relations A:B and C:D are. Measuring relational similarity is often done indirectly, and is modeled as a function of two instances of attributional similarity: (1) between A and C, and (2) between B and D.

Going back to semantic relations,Footnote ¹⁶ there is a similar split between two general lines of research. The first one derives the noun compound semantics from the semantics of the nouns it is made of (Rosario and Hearst Reference Rosario, Hearst, Lee and Harman2001; Rosario et al. Reference Rosario, Hearst and Fillmore2002; Girju, Badulescu and Moldovan Reference Girju, Badulescu and Moldovan2003; Moldovan et al. Reference Moldovan, Badulescu, Tatu, Antohe and Girju2004; Kim and Baldwin Reference Kim and Baldwin2005; Girju Reference Girju2006; Nastase et al. Reference Nastase, Sayyad-Shirabad, Sokolova and Szpakowicz2006; Ó Séaghdha and Copestake Reference Ó Séaghdha and Copestake2008; Tratz and Hovy Reference Tratz and Hovy2010), e.g., by generalizing them over a lexical hierarchy. This works for relations like Part-Whole, which are more permanent and context-independent, e.g., door-car. The second line of research models the relationship between the nouns directly (Vanderwende Reference Vanderwende1994; Lauer Reference Lauer1995; Lapata Reference Lapata2002; Turney and Littman Reference Turney and Littman2005; Kim and Baldwin Reference Kim and Baldwin2006; Nakov and Hearst Reference Nakov, Hearst, Euzenat and Domingue2006, Reference Nakov and Hearst2008; Turney Reference Turney2006a, Reference Turney2006b; Butnariu and Veale Reference Butnariu and Veale2008), e.g., using suitable patterns that can connect them in a sentence. This is useful for context-dependent relations like Cause-Effect, which are dynamic and often episodic, e.g., exam anxiety. These two lines are rarely combined; two notable exceptions are (Ó Séaghdha and Copestake Reference Ó Séaghdha and Copestake2009) and (Nakov and Kozareva Reference Nakov and Kozareva2011).

5.1 Semantic entailment

We have seen that for a right-headed endocentric compound n ₁n ₂, the relationship between the two nouns can be summarized as ‘n ₁n ₂ is a type/kind of n ₂’, e.g., for lung cancer we have that ‘lung cancer is a kind/type of cancer’. We can interpret this paraphrase as a semantic entailment: that the compound n ₁n ₂ semantically entails n ₂ in the sense that if we substitute n ₂ in a statement containing n ₁n ₂, we will obtain a new statement that is entailed from the original one.

Of course, this does not hold for all noun–noun compounds, e.g., it fails for exocentric compounds (birdbrain is not a kind/type of brain; it is a kind of person), or for left-headed ones (vitamin D is not a kind/type of D; it is a kind of vitamin).

For a three-word right-headed endocentric noun compound n ₁n ₂n ₃, it can be expected that the assertion ‘n ₁n ₂n ₃ is a type/kind of n ₃’ should be true, e.g., ‘lung cancer drug is a kind/type of drug’. If the compound is left-bracketed (n ₁n ₂)n ₃, it can be further expected that the assertion ‘n ₁n ₂n ₃ is a type/kind of n ₂n ₃’ would hold as well, e.g., ‘lung cancerdrug is a kind/type of cancer drug’. If the compound is right-bracketed n ₁(n ₂n ₃), then it can be expected that the following two assertions would be true ‘n ₁n ₂n ₃ is a type/kind of n ₂n ₃’ and ‘n ₁n ₂n ₃ is a type/kind of n ₁n ₃’, e.g., ‘oxydant wrinkletreatment is a kind/type of wrinkle treatment’ and ‘oxydant wrinkle treatment is a kind/type of oxydant treatment’.

In other words, an endocentric right-headed nonlexicalized three-word noun compound n ₁n ₂n ₃ is expected to entail n ₃ and n ₂n ₃, but not n ₁n ₃, if left-bracketed, and all three n ₃, n ₂n ₃, and n ₁n ₃, if right-bracketed. Therefore, noun compound bracketing could be used to predict semantic entailment. Unfortunately, it is hard for a computer program to distinguish left-headed versus right-headed compound, or left-bracketed versus right-bracketed compound; it is even harder to distinguish between a lexicalized and a nonlexicalized compound, the boundary between which is not clear-cut. The matter is further complicated due to idiomacity, lexicalization, transparency, world knowledge, nominalization, left headedness, and metonymy.

Table 3 shows sample three-word noun compounds, the corresponding bracketing and whether n ₁n ₃ and n ₂n ₃ are entailed or not. The entailments that are not predicted by the bracketing are shown in bold; we can easily explain some of them:

1. (1) Idiomacity. Idiomacity can interfere with semantic entailment by blocking the extraction of the head from a pair of bracketed words. This is the case with [butter ball] guy, where the idiomacity of butter ball (‘fat person’) does not allow the extraction of ball (i.e., here butter ball is not a kind of ball), and thus prevents the entailment of ball guy. The same applies to [lab lit] book, where lab lit, a literary genre, is idiomatic and blocks the extraction of lit (i.e., lab lit is not a kind of lit), and ultimately the entailment of lit book.

2. (2) Lexicalization. Another explanation for [butter ball] guy and [lab lit] book is lexicalization since idiomatic compounds are necessarily lexicalized. Lexicalization also explains the case of vanilla [ice cream], where ice cream does not allow the extraction of cream, thus blocking the entailment of vanilla cream.

3. (3) Transparency. Lexicalization cannot explain all examples in the table, most of which include subparts that are at least partially lexicalized. For example, [army ant] behavior includes the lexicalized compound army ant; still, this does not prevent the extraction of the head ant and thus the entailment of ant behavior. This extraction is possible since army ant is transparent. The same argument applies to compounds that contain even more highly lexicalized subparts, e.g., novel can be extracted from I novel, a kind of Japanese novel, so that novel writer can be entailed from [I novel] writer.

4. (4) World knowledge. The entailment of lung doctor from lung cancer doctor cannot be explained by the relatively low degree of lexicalization of lung cancer alone. The noun compound is left bracketed and this entailment is not predicted by the bracketing. Here, the semantics of the individual nouns and world knowledge come into play: doctor is likely to be modified by organs.

5. (5) Nominalization. Two problematic noun compounds are lung cancer patient and lung cancer survivor. While patient and survivor play similar role in the context of these compounds, lung patient is entailed, but lung survivor is not. Here survivor is a nominalization of the verb survive, which can easily take cancer, and therefore also lung cancer, but not lung, as a direct object. A similar argument applies to [Alien invasion] book and [planet Earth] survival.

6. (6) Left headedness. The actual entailments are completely reversed compared to what the bracketing predicts for vitamin D deficiency, which contains the left-headed noun compound vitamin D as a modifier.

7. (7) Metonymy. Due to metonymy, in the case of US army forces, all possible subsequences are entailed: not only US forces, army forces, and forces, but also US army, army and, in some contexts, even just US.

Table 3.

Semantic entailment and noun compound bracketing. The entailments that are not predicted by the bracketing are shown in bold

5.2 Application to textual entailment

In this section, we present some examples involving noun compounds from the development dataset of the Second Pascal Recognizing Textual Entailment (RTE2) Challenge.Footnote ¹⁸ We have chosen to illustrate the importance of understanding the syntax and semantics of noun compounds on this particular task since it requires generic semantic inference that could potentially help many natural language processing applications, including Question Answering (QA), Information Retrieval (IR), Information Extraction (IE), and (multi-)document summarization (SUM).

Given two textual fragments, a text T and a hypothesis H, the goal is to recognize whether the meaning of H is entailed (can be inferred) from T:

Below, we give examples illustrating different kinds of phrase variability involving noun compounds that a successful inference engine for RTE might need to solve. Of course, recognizing that two phrases are variations of each other does not solve the RTE problem automatically (e.g., below we have both positive and negative examples with respect to RTE), but it could help the overall decision by allowing an RTE engine to focus on the remaining words rather than on the surface variability between noun compounds and their paraphrases/subparts.