Defining Polysemous Words

Peter Norvig

Computer Science Division
University of California, Berkeley
Berkeley, CA 94720

This paper attempts to show what the definition of a word should look like, in order to be of use to AI programs such as analyzers and generators. We concentrate on take, a highly polysemous word which poses many vexing questions. Why is it that we can say take a look or take a whiff, but not take a stare or take a smell? Why can a photographer take a picture, while an artist cannot take a painting? It is clear that there must be several distinct senses of take, because the implications are different for different usages. For example, when I take a book (or when I am given a book) it is felicitous to say I got a book or I have a book. However, if I take it in stride, it is not the case that I got it in stride or I have it in stride. Faced with these facts about language, our goal is to come up with definitions for the various senses of a word such that:

The Homonomy Approach

The simplest solution to the problem of multiple senses is to treat them as homonyms, that is, as completely distinct words that just happen to have the same spelling. This is the approach taken by most AI programs, and it seems correct in many cases. A classic example is the word bank. According to the American Heritage dictionary, bank has three main senses: bank1 (river bank), bank2 (money bank) and bank3 (memory bank). While all of these may have derived historically from the same Germanic word, to the speaker of current-day English they seem to be unrelated.

Unfortunately, three senses are not enough to completely characterize bank. The American Heritage lexicographers also give several senses within each of the three entries. For example, bank1 can be a river bank, a snow bank, a cloud bank, an artificial embankment, or the cushion of a billiard table. While the intent of the lexicographers is clearly to say that these senses are related, they give no indication of how they are related, and thus they might as well be homonyms. For example, one sense of bank2 is ``money-lending institution,'' and another is ``building where money-lending occurs.'' The same dual senses show up in words like university, hospital and senate. Dictionaries and AI programs that fail to explicitly recognize this regularity are missing an important generalization.

The Abstractionist Approach

The abstractionist approach demands exactly one definition for each word. If a given word seems to have several senses, then the definition must be abstract enough to encompass all possible senses. For example, Bendix gives the following definitions for a set of verbs:

A takes B from CC has B, and then A causes A to have B
C gives A BC causes A to have B
A gets BA changes to having B
C gets A Bsome D has B, and then C causes A to have B
A has B there is a relation between A and B

Bendix devotes an entire chapter to describing the kinds of relations that characterize has. He does not attempt to account for any metaphorical senses of the verbs, but as we shall see shortly, there are many common-place, fairly literal uses that fall outside the given definitions.

The problem with abstractionism is that it is very difficult to get the right definition without over- or undergeneralizing. In the list below, (1a) and (3a) are acceptable uses of take, while (2a) is not. To account for this with abstract definitions would seemingly require fairly specific limitations on the object of take.

Bendix's alternative is to define take in terms of have, in such a way that take(x) is acceptable only when have(x) is. But this approach could not fit these data, regardless of how clever the definition of have is. It could work for (1) and (2), where the (a) sentences are acceptable when the corresponding (b) sentences are. However, in (3) we find (3a) is acceptable, while (3b) is not. (While have a picture is grammatical, it does not mean ``have photographed a picture.'') These data, as well as other cases presented below argue that there cannot be a single abstract definition for take.

(1a)I took a look(1b)I had a look
(2a)*I took a stare(2b)*I had/have a stare
(3a) I took a picture(3b)*I have a picture

The Polysemy Approach

In the polysemy approach, it is accepted that words can have a variety of related senses. Linguists and lexicographers have long recognized polysemy, but there have been few serious attempts to show just how polysemous senses are related.

In the remainder of this paper, I will start to give an analysis of the word, take. I start by giving an abstract sense, which is meant to encompass many but not all uses, and a prototypical sense, which is meant as the most basic sense, one which would be learned first as a child, and would be most likely elicited from subjects as the naive definition of take.

Takea: abstract take
Roles:Agentactive actor or causer of an action
Patientobject acted upon or affected by the agent
Instrumentused to transport the patient
Recipientreceives the patient
Donorinitially has the patient
Originlocation where patient started out
Endpointlocation where patient ends up (destination)
Focuscan be on any of the roles, steps, or results
Steps:S1A attends to P, distinguishing it from the background
S2A wants R to have P
S3A moves I to P
S4A secures P with I
S5A moves P with I from O to E
Results:R1R has P
R2D no longer has P
R3P is at E
R4P is no longer at O
R5I is at E
R6R can now employ P for some use

Take0: prototypical take, grab
Restrictions: A = R; A is human; P = easily manipulated physical object; I = A's arm and hand; O = near A; E = at A's body; D is non-existent; focus is on A and P
Examples: the baby took the toy from its mother

The next step is to show that other senses are necessary, and to show how they differ from the prototypical sense. Consider the following expressions:

(4a)The boy took a beating from the teacher.
(4b)The boy got a beating from the teacher.
(4c)The boy was given a beating by the teacher.
(5a)The teacher took a whip to the boy.
(5b)*The teacher got the boy a whip.
(5c)*The teacher gave a whip to the boy.

First, concentrate on (4a) and (5a). They both express a similar situation, where the teacher was the agent of a beating action, and the boy was the person beaten. Superficially, they look like they may be the same sense of take. However, (4a) can be paraphrased as (4b,c), while (5a) cannot be paraphrased by (5b,c). This is evidence that they may be different senses. The evidence suggests the following rule: If two expressions are to be considered as instances of the same sense, they must have the same mapping between grammatical and semantic cases (after adjusting for passives and the like).

The next step in the analysis is to decide what the possible collection of cases for take are. The five most important ones seem to be take from, take to, take with, take in and just plain take (towards oneself). Another important distinction is between senses of possession and of motion. Steps S2 and S4 of take emphasize the former, while S3 and S5 emphasize the later. In the analysis below, senses 2-4 are primarily motion verbs, while 5-10 are primarily verbs of possession. Sense 1 has both connotations. When the sentence John took the book off the desk follows John wanted to read, it stresses possession. When it follows The desk was messy, it stresses motion.

To resolve the problem with (4-5), we can categorize take a beating as sense 5, which is a possession verb with corresponding senses of get and give. On the other hand, take a whip to is categorized as sense 3, which emphasizes motion, and has no corresponding senses for get and give. It is further complicated in that this sub-action of moving the whip to the proximity of the boy is taken as a metonymic reference to the entire action of beating.

I will now go on to list the main senses of take.

Take1: neutral take
Restrictions: D is non-existent; A = R
Examples: I took the book off the desk; take my wife, please; take a point P in the complex plane (see also 6); take a look/glimpse/glance/whiff; take a *listen/*smell/*stare

This sense has the same collection of cases as take0, but relaxes the restrictions on agent and patient; they no longer are required to be a human and an easily manipulated physical object. However, we can also apply metaphorical extensions. Using the metaphors THE MIND IS A CONTAINER and IDEAS ARE OBJECTS, as discussed by Lakoff and Johnson, we can postulate the metaphor UNDERSTANDING IS HAVING. In this metaphor, an idea plays the role of the patient, and the agent's consciousness stands metonymically for the agent. This metaphor shows up in the take my wife and take a point examples above, and in sense 6 as well. It shows up in phrases like I don't get it and this is a hard concept to grasp, so it has widespread applicability.

Another metaphor at use here is THE EYES ARE LIMBS. This is discussed by Lakoff and Johnson, and from a different perspective by Gruber. Because looking is a verb of motion where the eyes are understood as if they were arms, take a look is acceptable. The senses of smell and hearing do not have the intentional and directional properties of sight, and thus are not conceptualized as limbs. Thus take a smell and take a listen are ruled out. In take a whiff we add just the intentional and directional properties that are missing in take a smell. Finally, take a stare seems to negate motion because it connotes a fixed, unmoving gaze.

Notice that the metaphors act orthogonally to the senses I am defining here. Once the basic senses are defined, the hope is that metaphoric usages can be mapped to an existing sense, without having to pre-define every possible metaphoric usage.

Take2: take from
Restrictions: D must exist; A = R; focus on R2
Examples: John took the book from Mary; Mary took the dangerous toy away from the baby; take this problem off my hands; take my wife, please; the war took 50,000 lives

Take3: take to
Restrictions: A ≠ R; E is non-obvious or far from O; hence S3 will be more extensive (more in focus)
Examples: the messenger took the book to Mary; take a whip to the boy; take pen to paper

Sense 3 is atypical in that it is the only one where both agent and recipient are emphasized, yet they are distinct. However, it is primarily a motion verb, so the whole question of possession is de-emphasized.

Take4: take with
Restrictions: A = I; A moves with P in S3; de-emphasize R, possession
Examples: take this paper home with you; take the dog for a walk; John took Mary to the movies; I was taken for a ride; he took me on that deal

Take to the movies seems to be the conjunction of two senses: 4 and 8 (to engage in an activity). Thus, if John pays, he can be said to have taken Mary, even if they arrive separately. For some speakers, these connotations are lost in the passive; Mary was taken to the movies by John may mean that he dropped her off there (sense 3). The last two examples of take4 make use of the metaphor that the trickery is a path down which the agent leads the victim, the patient.

Take5: take in, accept, receive,
Restrictions: A ≠ R; focus on R; A non-existent or de-emphasized
Examples: take a phone call; take a punch; he can't take a joke; don't take it the wrong way; take a beating; take advice; take a pitch; take job applications; take a bet; take in a play; take [in] the view; take [in] boarders; take [on] students; take a poll/survey; *take a questionnaire; take a picture; *take a painting; take his name and number; take a reading on the dial; *take an interpretation of the data

The particle `in' is used when the recipient is metaphorically viewed as a container. There seems to be a sub-sense of `passively recording an event.' Photographing and reading a dial are passive enough to qualify under this sense, while painting and interpreting require a sense with a more active agent.

Take6: pick out, select
Restrictions: A = R; focus on S1; thus D de-emphasized
Examples: take a card, any card; I think I'll take the blue one; take a wife; take a point P in the complex plane (see also 1)

We stated above that a metaphorical use of sense 1 was to have into mind, with the emphasis on bringing into mind. Here we have a similar metaphor, but the emphasis is on choosing, rather than on having. Thus, if take a point P means consider a point, then it is sense 1, but if it means choose a point out of many possibilities, then it is sense 6.

Take7: capture, gain control
Restrictions: A = R; focus on S4 and R1
Examples: take prisoners; take a quail; take his limit of trout; take the offensive; the marines took the hill; the Giants took a doubleheader; take first place; take the lead; take a trick; *take a point; *take a touchdown

Its not surprising that tricks can be taken, since S3 and S5 are involved in gathering up the cards. Scoring a point or touchdown involves no such motion, so take is not used to describe them.

Take8: make use of, engage in
Restrictions: focus on R6; de-emphasize R
Examples: take a bus; take cover; take University Ave.; take a left; take a whip to the child; take pen in hand; take a bath/shower; take an exam/course/*homework assignment; take a nap/rest/break/*dream; take a drive/walk/trip/vacation/*run/*jog; take a shit/piss/*fart; take care; take pity on; *take sympathy; take tea; take a drink/breath/aspirin/*apple (see also 1)

Note the difference between take a poll/survey (sense 5) and take an exam/course (sense 8). In the first the questioner is the agent, in the second, the answerer.

Take9: rightfully obtain
Restrictions: A = R; S4 is an action that entitles A to receive P
Examples: take a cottage for the summer; take the Times; take a leave of absence; the agent's take was 10%; *take a salary; *take a house (meaning buy)

Take10: require
Restrictions: A = R; focus on R1, which is necessary for some purpose
Examples: it takes money to make money; it takes time; the car takes five passengers; this verb takes an animate subject

Take10 is a tenseless sense. We don't say *this morning it took money to make money three times. A good example of how closely different senses can be related is seen in he takes sugar in his coffee, which is sense 10, versus he took two lumps of sugar from the dish, which is sense 2, while the coffee took two lumps before it was too sweet, which is sense 5.


We have examined over 100 uses of take, and categorized them into 10 senses. Along the way, the following points were made:


  1. Bendix, Edward Herman. Componential Analysis of General Vocabulary: The Semantic Structure of a Set of Verbs in English, Hindi, and Japanese., International Journal of American Linguistics, Vol. 32, No. 2, 1966.
  2. Gruber, Jeffrey S. Look and See, Language, Vol. 43, Pp. 937-947, 1967.
  3. Lakoff, George and Johnson, Mark. Metaphors We Live By University of Chicago Press, Chicago, 1980.