Expressing causality in natural language

Causality is not a linguistic concept, but language is the best only tool human beings possess to express causality between events or, more generally, between states of affairs. In this contribution, I would like to show that the fine study of the linguistic means to express causality is of great interest for the study of human cognition. The main purpose of my paper is to illustrate two very interesting facts about linguistic causation.

The first fact concerns the syntactic means natural languages possess to express causality. As an anticipation of one of my claims, I will argue that the general pattern linguistic constructions offer to express causality is not the explicit description of an event and its agent, but the explicit description of the resulting state and its patient.

This observation is very surprising, because an a priori and intuitive description of causality would predict the opposite: causality is inferable from events and agents. The second point is at a first sight very strange, and until now unexplained: when we have to report causal relations between events or states, the linguistic order is the consequence-cause order, and not the cause-consequence order. Any attempt to connect the cognitive properties of causal relations and linguistic reports of causal relations would predict, for reason of processing, that the natural order between events (cause-consequence) would be the linguistically preferred order (Ahn & Noseck 1998). What is surprising is that natural languages all possess at least one causal connective (engl. ‘because’, fr. ‘parce que’, it. ‘perche’, etc.) that imposes the consequence-cause order:

(1) Axel fell because Abi pushed him.

(2) Axel est tombé parce que Abi l’a poussé.

What is surprising is that if we change the order of the events in (1) or (2), the causal reading is still accessible, but the appropriate interpretation is one in which the falling event causes the pushing event (for instance to put the patient out of danger), the push-fall causal reading (called ‘inferential order’) being more difficult to access:

(3) Abi pushed Axel because he fell.

(4) Abi a poussé Axel parce qu’il est tombé.

What is still more puzzling is the following fact: if we suppress the causal connective, as in (5) and (6), only the consequence-cause order in (5) triggers the causal reading, whereas (6) is typical of what linguists call narratives, that is, discourses where the linguistic order parallels the chronological order of events:

(5) Axel fell. Abi pushed him.

(6) Abi pushed Axel. He fell.

We have now all the basic linguistic material allowing discussing the causality puzzle: (i) why is (5) the natural order for conveying a causal report of events and (ii) why does the causal connective (‘because’) impose the consequence-cause order?

I will give two main arguments for the cognitive motivation of the consequence-cause order in causal reports: the first refers to the semantics of causal constructions and the second to the linguistic distribution of connectives. Finally, I will give an experimental argument in favour of the consequence-cause order.

A general semantic pattern for causal constructions

The syntax and semantics of causality has been extensively and very precisely analysed (Kaynes 1975, Levin & Rappaport, Jackendoff 1990, Pustejovsky 1995 among others). Linguists recognize three syntactic means to express causality: (i) causative constructions, (ii) ergative constructions and (iii) inaccusative constructions. The first strategy is used either when no causative verb (like ‘kill’ = cause to die, ‘sink’ = cause to be sunk, ‘bake’ = cause to become baked, etc.) is accessible in the lexicon. Languages differ widely on this point: for instance, French has a very poor causative verbal lexicon, whereas English can easily give a causative meaning to transitive verbs, as in (7):

(7) Bill raced Aga Khan’s horse at the Prix d’Amérique (= made the horse run)

(8) *Bill a couru le cheval de l’Aga Khan au Prix d’Amérique (≠ a fait courir le cheval)

Causative constructions use a predicate operator (‘make’ in English, ‘faire’ in French) which trigger the so-called causative reading, as in (9):

(9) Mary made the children eat the soup.

(10) Marie a fait manger la soupe aux enfants.

What seems to be an universal property of causative construction in natural language is that the use of the causative operator in addition to a causative verb conveys the implicature (= the pragmatically inferred meaning) that the means to cause the event is not an ordinary one.

Compare (11) and (12):

(11) Bill stopped the car.

(12) Bill made the car stop.

The implicature in (12) is that Bill stopped the car by a special means, the handbrake for instance. The second means to express causality in natural language consists in using a causative verb in a transitive construction. This type of construction, called ergative (the ergative case is the mark of the agent in ergative languages, such as Basque), has the special property to entail a resulting state, described by the causative predicate:

(13) Bill stops the car → the car is stopped

(14) Mary opens the door → the door is open

(15) Autumn yellows the leaves → the leaves are yellow

Some of these verbs can have an intransitive use, in which the agent (ergative case in ergative language, nominative case in nominative-accusative languages) is deleted, but the resulting state is still entailed:

(16) The car stops → the car is stopped

(17) The door opens → the door is open

(18) The leaves yellows → the leaves are yellow

If we now try to find the common properties between these three strategies for expressing causality, what we have to target are the basic parameters expressed by these constructions. Causative constructions express the agent and the caused event or state, containing respectively an agent (optionally a theme, i.e. the object of the event) and a patient:

(19) Mary (AGENT) made the children (AGENT) eat (EVENT) the soap (THEME)

(20) The wind (AGENT) made the leaves (PATIENT) fall (EVENT)

In the ergative (transitive) constructions, the agent, the causal event and the patient are expressed, but the event entails the resulting state (the meaning of a causative verb is to entail the resulting state)

(21) John (AGENT) broke (EVENT) the branch (PATIENT)

Finally, the inaccusative (intransitive) construction only expresses the patient (in the subject position) and the event, which also entails the result-state:

(22) The branch (PATIENT) broke (EVENT)

Now, the answer to the previous question is easier to give: what is common to all these constructions is the event predicate and the patient. And the event entails the resulting state, that is, what is meant by the causative verb or the causative construction. So the conjecture I would like to propose is the following: what creates a causative meaning in a sentence is not the agent or the event functions, but the resulting state and the patient ones. In other terms, the explicit description of a patient and of a result state is what makes the construction of a causal meaning possible. If the conjecture is correct, then we have a first argument to explain why causal discourses have the consequence-cause order: the consequence of the causal-event is either an event or a state, whose argument is a patient. The practical implication of the argument is the following: the utterance of (23) and (24), which describe states, are good candidates for being followed by an explanation, whose pragmatic function is to state the causal event:

(23) Axel is sick.

(24) John is dead.

(25) Axel is sick. He ate too much chocolate.

(26) John is dead. He was killed in a road accident.

So, the semantics of causative constructions can be given by the following schema: the resulting state is part of a complex structure containing the event predicate and the agent, none of them being mandatory. Figure 1 represents the general pattern, and Figure 2 the realisation of the pattern for (22):

Figure 1: semantic structure of a causative sentence

Figure 2: semantic structure of ‘the branch broke’

The distribution of causal, inferential and temporal connectives

The second argument I would like to present is linked to the distribution of connectives in French, i.e. causal connective ‘parce que’ (‘because’), inferential connective ‘donc’ (‘so’) and temporal connective ‘et’ (‘and’). One relevant distributional fact is the order of causal relations with connectives. Causal connectives as ‘parce que’ (‘because’) generally introduce the cause. The combination of event and state (as cause or consequence) gives four possible pairs of utterances, given in (27), all interpreted as causal in the consequence-cause order. (28) shows the reverse order (cause-consequence) implying the inferential use of ‘parce que’ (i.e., the cause-consequence order). (29) presents cause-consequence sets of utterance with an inferential connective (‘donc’, ‘so’), (30) consequence-cause sets of the reverse, (31) and (32) the same structure with a temporal connective (‘et’, ‘and’).

(27) Canonical (consequence-cause) series with ‘parce que’: causal readings

a. CAUSE (EVENT, STATE) : Marie est malade parce qu’elle a trop mangé

‘Mary is sick because she ate too much'

b. CAUSE (EVENT, EVENT) : Jean est tombé parce que Marie l’a poussé

‘John fell because Mary pushed him‘

c. CAUSE (STATE, STATE) : Marie ne peut pas boire d’alcool parce qu’elle est mineure

‘Mary cannot drink alcohol because she is a minor‘

d. CAUSE (STATE, EVENT) : Le médecin soigne Axel parce qu’il est malade

‘The doctor is treating Axel because he is sick’

(28) Non-canonical (cause-consequence) series with ‘parce que’: inferential readings

a. Marie a trop mangé, parce qu’elle est malade

‘Mary ate too much, because she is sick’

b. Marie a poussé Jean, parce qu’il est tombé

‘Mary pushed John, because he fell’

c. Marie est mineure, parce qu’elle ne peut pas boire d’alcool

‘Mary is a minor, because she cannot drink alcohol’

d. Axel est malade, parce que le médecin le soigne ‘Axel is sick, because the doctor is treating him’

(29)Canonical (cause-consequence) series with ‘donc’: inferential and causal readings

a. # Marie a trop mangé, donc elle est malade (# means a change in reading)

‘Mary ate too much, so she is sick’

b. # Marie a poussé Jean, donc il est tombé ‘Mary pushed John, so he fell‘

c. Marie est mineure, donc elle ne peut pas boire d’alcool

‘Mary is a minor, so she cannot drink alcohol’

d.Axel est malade, donc le médecin le soigne

‘Axel is sick, so the doctor is treating him’

What happens in (29) with the cause-consequence order of ‘donc’ is that the causal readings are restricted the states as causes, event-causes giving rise to inferential reading.

The test to trigger the inferential reading is the following: if the consequence can be false, then the reading is inferential.

(30)Non-canonical (consequence-cause) serie with ‘donc’: inferential readings

a. Marie est malade, donc elle a trop mangé

‘Mary is sick, so she ate too much’

b. Jean est tombé, donc Marie l’a poussé

‘John fell, so Mary pushed him’

c. Marie ne peut pas boire d’alcool, donc elle est mineure

‘Mary cannot drink alcohol, so she is a minor’

d. Le médecin soigne Axel, donc il est malade

‘The doctor is treating Axel, so he is sick’

(31)Canonical (cause-consequence) series with ‘et’ : causal and inferential readings

a. Marie a trop mangé et elle est malade

‘Mary ate too much and she is sick’

b. Marie a poussé Jean et il est tombé

‘Mary pushed John and he fell’

c. # Marie est mineure et elle ne peut pas boire d’alcool

‘Mary is a minor and she cannot drink alcohol’

d. ? Axel est malade et le médecin le soigne

“Axel is sick and the doctor is treating him”

(32)Non-canonical (consequence-cause) series with “et”: inferential reading impossible

a. ?? Marie est malade, et elle a trop mangé

“Mary is sick, and she ate too much”

b. ?? Jean est tombé, et Marie l’a poussé

‘John fell, and Mary pushed him’

c. ?? Marie ne peut pas boire d’alcool, et elle est mineure

“Mary cannot drink alcohol, and she is a minor”

d. ?? Le médecin soigne Axel, et il est malade

“The doctor treats Axel, and he is sick”

The following table gives a summary of these distributions:

Table 1 : causal and inferential readings of ‘parce que’, ‘donc’, ‘et’

The picture is rather interesting: causal readings cannot be obtained in the cause-consequence with donc when the cause is an event and with et when the cause is a state: parce que is the only connective allowing the causal reading whatever the cause is (a state or an event). But the cost to ensure causal relation is the consequence-cause reading. This order triggers a necessary constraint: the consequence-cause with donc always yields an inferential reading, and no readings are accessible with et. So, the causal readings with the cause-consequence order are obtained either by donc (the cause is a state) or by et (the cause is an event), the consequence of this distribution being that parce que is the optimal candidate for causal readings, whatever the cause is. We have thus a second argument for the consequence-cause order: this order is the only one compatible with any type of aspectual classes (state or event) and corresponds to the canonical order of causal relations with a connective (parce que).

Experimental data on causal discourses

The connective argument is not strong enough: it should be tested with other languages than French to have a greater strength. What I would like to do now is to add an third argument, based on experimental data. I will give a very brief survey of a first interesting conclusion obtained from two experiments. The first experiment was a simple elicitation task: we show a series of event sentences composed of 8 syllables and ask 38 students from University of Lyon 2 to complete the stimuli either by a cause (20 students) or a consequence (18 students). 36 stimuli were offered to each subject. From these 36 initial propositions, 10 pairs of propositions have been selected, 5 pairs of highly associated propositions (more than 50% of given responses) with the consequence-cause and the cause-consequence order, 5 weakly associated (less than 35% of given responses) with both orders. The first condition to be tested was the strength of association, and the second condition the order of utterances (cause-consequence and consequence-cause).Table 2 (Moeschler et al. 2006) gives the series of inputs for the second experiment:

These 20 likely utterances were balanced by 20 unlikely utterances (in the order cause-consequence and consequence-cause), and all of these 40 utterances have been checked by control-utterances. 22 subjects passed the cause-consequence series, 27 subjects the consequence-cause series, and 22 the control-utterances. The design of the experiment was implemented with the E Prime software, and the subject, after having read the prompt, had to read the second proposition and type ‘e’ or ‘p’ for ‘likely’ or ‘unlikely’ . Reading time has been recorded, the analysis dealing with reading time. The results are the following: (i) with high associated pairs of propositions, no significant difference in reading time occurs; (ii) on the contrary, with weak associated pairs of propositions, a significant difference in reading time occurs: the causal reading (consequence-cause) is quicker than the cause-consequence order (164,80 ms against 308,84 ms). This result must be precisely analyzed, and our data have now to be completed by new experiments. At this stage, the precise interpretation is not certain, but data from the second experiment allows us to conclude that the consequence-cause order has some cognitive motivation. In effect, it seems that when no strong conceptual association between event predicates exist, the causal reading is more accessible than the inferential one. This is a crucial point for explaining the consequence-cause order in causal discourses, though it must be checked by other experiments, in particular about causal and inferential connectives. Can we for instance predict that the presence of ‘parce que’ will give better results than its absence, or that ‘parce que’ will trigger a better treatment of causal connection than ‘donc’?

Conclusion

References

Ahn W. & Nosek B. (1998). "Heuristics used in reasoning with multiple causes and effects". Proceedings of the 20th Annual Conference of the Cognitive Science Society, Mahwah (NJ), Erlbaum, 24-29.

Blochowiak et al. (2006). "Le projet causalité: analyses quantitatives et qualitatives d’un pré-test". Cahiers de linguistique française 27, to appear, clf.unige.ch.

Jackendoff R. (1990), Semantic Structures, Cambridge (MA), MIT Press.

Kayne R.S. (1975). French Syntax, Cambridge (MA), The MIT Press.

Levin B. & Rappaport Hovav M. (1995). Unaccusativity. At the Syntax-Lexical Interface, Cambridge (MA), MIT Press.

Moeschler, J. (2003). "Causality, lexicon, and discourse meaning". Rivista di lingusitica 15/2: pp. 277-303.

Moeschler J.Chevallier C., Castelain T., Van der Henst J.B.,Tapiero I.(2006). "Le raisonnement causal. De la pragmatique du discours à la pragmatique expérimentale". Cahiers de linguistique française 27, to appear, clf.unige.ch.

Pustejovky J. (1995). Generative Lexicon, Cambridge, The MIT Press.