I. INTRODUCCIÓN
1.3. MARCO TEÓRICO (TEORÍAS RELACIONADAS AL TEMA)
1.3.4. Estudio de Tiempos
2.1 Word Association Test
Rather early on, it was noted that words in the human mind are linked. The Ameri- can clinical psychologists G. Kent and A. J. Rosanoff, perceived the diagnostic useful- ness of an analysis of the links between words. In 1910, the duo created and con - ducted a test of the free association of words. They conducted research on 1000 peo- ple of varied educational backgrounds and professions, asking their research subjects to give the first thought that came into their minds as a result from a stimulus-words. Those research was supplied with 100 word-stimuli, (principally nouns and adjec- tives). The Kent-Rosanoff list of words was translated into several languages, in which this experiment was repeated, thereby enabling comparative research to be car- ried out. Word association research was continued by Palermo, Jenkins (1964), Post- man, Keppel (1970), Kiss, Armstrong, Milroy, Piper (1973), Moss, Older (1996), Nel- son, McEvoy, Schreiber (1998), and the repeatability of results allowed the number of research subjects to be reduced, while at the same time increasing the number of word-stimuli to be employed, for example 500 research subjects and 200 words (Pa- lermo, Jenkins, 1964), or 100 research subjects and 8400 words (Kiss, Armstrong, Milroy, Piper, 1973). Research on the free association of words has also been con- ducted in Poland (Kurcz, 1967) and the results makes a basis for the experiment de- scribed below.
Computational linguistics also became involved in research on the free association of words, though at times these experiments didn’t employ the rigors used by psychol- ogists when conducting experiments, for example, those that permitted the possibility of providing several responses to an individual stimulus-word. (Schulte in Walde S., Borgwaldt S., Jauch R., 2012), or those that used word pairs as a stimulus (Rapp, 2008).
There exist some algorithms, which generate an association list on the basis of text corpora. But automatically generated associations were rather reluctantly compared with the results of psycho-linguistic experiments. The situation is changing, Rapp’s results (2002) were really encouraging.
Finally, association norms start serving for different tasks, as for example informa- tion extraction (Borge-Holthoefer, Arenas, 2009) or dictionary expansion (Sinopal- nikova, Smrz, 2004), (Budanitsky, Hirst, 2006).
2.2 The Author’s Experiment
Some 540 students of the Department of Management and Social Communication at the Jagiellonian University participated in the free word association test as de - scribed in this article. A Polish version of the Kent-Rosanoff list of stimulus words, which was previously used by I. Kurcz was employed (Kurcz, 1967). After an initial analysis it was determined that we would employ as a stimulus, each word from the Kent-Rosanoff list, which grammatically speaking is a noun, as well as the five most frequent word associations for each of those nouns obtained in Kurcz’s experiment (Kurcz, 1967). If given associations appeared for various words, for example, white for doctor, cheese, sheep, that word as a stimulus appeared only once in our experi- ment. The resulting stimulus list contained 60 words from the Kent-Rosanoff list, in its Polish version, as well as 260 words representing those associations (responses) which most frequently appeared in Kurcz’s research. It therefore, is not an exact repe- tition of the experiment conducted 45 years ago.
The conditions of the experiment conducted, as well as the method of analyzing the results, have been modified. The experiment was conducted in a computer lab, with the aid of a computer system, which has been created specifically for the requirements of this experiment. This system presents a list of stimuli and then writes down associa- tions in a data base. Instructions appeared on the computer screens of each partici - pant, which in addition were read aloud by the person conducting the experiment. Af- ter the instructions were read, the experiment commenced, whereby a stimulus word appeared on the computer screen of each participant, and he wrote the first free asso - ciation word which came to his mind - only one response was possible. When the participant wrote down his association, (or the time ran out for him to write down his association), the next stimulus word appeared on his screen, until the experiment was concluded. The number of stimulus-words as well as their order, was the same for all participants.
As a result we obtained 260 association lists, which consist of more than 16,000 as- sociated words.
Association list derived from the experiment will be used to evaluate algorithm de- rived association lists. But first, we have to show how the human associations are comparable.
2.3 Comparison of Human Association Lists
We shall compare a Polish list derived from our experiment to a semantically equivalent English list derived from the Edinburgh Associative Thesaurus. To illus- trate the problem we selected an ambiguous Polish word dom, which refers to the
English words home and house. Those lists will present words associated with their
basic stimulus, and ordered in accordance to their strength of association. Due to the varied number responses (95 for home and house and 540 for dom) we will be using a more qualitative measure of similarity based on the rank of occurring words on them, rather than on a direct comparison of association strength. That list measure
LMw(l1,l2), given two word lists l1 and l2 and a comparison window, which will be equivalent to the amount of words matching in l1 and l2 in a window of w words taken from the beginning of the lists.
In order to establish some basic expected levels of similarity, we will compare the list obtained in our experiment for the stimulus word dom, which meaning covers both English word home and house. First, each Polish association-word was carefully trans-
lated into English, then the lists automatically looked for identical words. Because words may differ in rank on the compared lists, the table includes the window size needed to match a word on both lists.
Table 1. Top 10 elements of the experiment lists for dom (author's experiment) and the EAT lists for home and house
dom home house
rodzinny (adv. family) house home
mieszkanie (flat) family garden
rodzina (n.family) mother door
spokój (peace) away boat
ciep o (ł warmth) life chimney
ogród (garden) parents roof
mój (my) help flat
bezpiecze stwo ń (security) range brick
mama (mother) rest building
pokój (room) stead bungalow
Those lists can be compared separately, but considering the ambiguity of dom, we can compare the list of association of dom with a list of intersparsed (i.e. a list com- posed of the 1st word related to home, next to the 1st word associated with house, then the 2nd word related to home etc.) associations of both home and house lists coming from EAT.
Table 2. Comparison of the experiment list and the EAT lists. Matching words are shown for their corresponding window sizes w for the LMw(l1,l2) measure
w home+house vs dom w home vs dom w House vs dom
3 family 3 family 3 family
6 garden 9 mother 6 flat
w home+house vs dom w home vs dom w House vs dom
12 roof 24 garden 11 roof
14 flat 26 parents 14 room
18 building 35 peace 15 building
19 chimney 41 security 19 chimney
26 parents 21 cottage 30 room 30 mother 32 brick 32 brick 35 cottage 34 security 64 security 40 warm 65 peace 41 warmth 74 warm 75 warmth
The original, i.e. used for comparison human association list, is a list of words as- sociated to a stimulus-word ordered by frequency of responses. Unfortunately, we can not distinguish automatically that words, which enter into semantic relation to the stimulus-word by frequency or by computed association strength, for example in the list associated to the word table a semantically unrelated cloth is substantially more frequent than legs and leg, which enter into ‘part of’ relation to the table. (Palermo, Jenkins, 1964). The described observation is language independent. The proposed method of comparison truncates from the resulting list language specific semantic as- sociations, e.g. home – house and house – home the most frequent on EAT as well as all non-semantic associations, e.g. home – office or house – Jack. Each resulting list consists of words, each of which is semantically related to a stimulus-word. In other words, the comparison of the human association list will automatically extract a sub- list of semantic associations.