You are on page 1of 5

Something Interesting About First and Second Order Similarity

Abstract vs. word pairs that are associates but not collocates. Despite
the lack of evidence that HAL and LSA depict paradigmatic
Brief description of what we’ve done.
and syntagmatic relationships, respectively, the authors
Keywords: semantic similarity; priming; ???. proceed to claim that HAL is important in accounting for
semantic (i.e. paradigmatic) priming while LSA is important
Introduction in accounting for associative priming. By showing that our
FS: Point 1 (Corresponding to results from BLP experi- semantic space measurements really represent paradigmatic
ment)): Buchanan et al. (2001) showed that the density of similarity and syntagmatic similarity, we can really see how
the semantic neighbourhood, as defined by a semantic space important paradigmatic and syntagmatic similarity are for
(HAL) measurement, is a significant factor in predicting both priming of paradigmatically-related words and priming
performance in the visual word recognition tasks of LDT and of syntagmatically associated words.
Naming. It is not clear what measure of semantic similarity
is important for the effect. That is, how should the semantic FS: Point 3, two related subpoints:
neighbourhood be defined in order to predict RT? If we • We are the only ones to control for the effect of the
define it in terms of a word’s most paradigmatically related semantic neighbourhood for item-level priming analyses.
words, we will get a measure of paradigmatic semantic This was a methodological issue missed by Gunther et al.
density. Conversely, if we define the neighbourhood in terms (2016) and Mandera et al. (2016). Without accounting for
of a word’s most syntagmatically related words, we will get this covariate, their analysis is
a measure of syntagmatic semantic density. These measures • We show that both FOS and SOS are important factors in
are substantially different, and it is not clear if both or only accounting for priming of paradigmatically related words
one of the measurements are important factors in predicting (which is surprising, and against findings of Jones et al.
variance in RT. We show that syntagmatic semantic density, (2006) and Gunther et al. (2016)). We show that both FOS
and not paradigmatic semantic density, is an important and SOS are important, overall, in accounting for item-
semantic-space factor in predicting variance in RT. level RT. This is strengthened by Point 2, where we showed
that FOS is syntagmatic similarity and SOS is paradigmatic
FS: Point 2 (Corresponding to results of Experiment 1, similarity, in that we are the first to show that both paradig-
see experiment 1 mean neighbour ranking.png): Within the matic and syntagmatic similarity are important in account-
priming dataset (SPP), we show that paradigmatically related ing for priming.
pairs are represented as more related with the FOS than the
SOS. Similarly, we show that syntagmatically related pairs
are represented as more related with the FOS than the SOS. What are the main points we’ll make here:
This establishes SOS as a measurement of paradigmatic • 1-2 paragraphs: Intro to the general issues
similarity and FOS as a measurement of syntagmatic simi- – The study of semantic similarity and why it’s impor-
larity. Establishing this is important because it enables us to tant...
evaluate the effect of paradigmatic and syntagmatic similarity – Different types of semantic similarity...
on paradigmatic categories and syntagmatic categories. It
• subsequent paragraphs: what others have done
has been claimed (Jones et al., 2006; Gunther et al., 2016)
that HAL and LSA also represent such relations. However, • final paragraph(s): Our take – what’s new and how do we
this doesn’t seem to be empirically supported. Gunther et show it.
al. (2016) report a very strong correlation between HAL
and LSA similarities in both of their item-level regression
Related Work
analyses (r = 0.89 and r = 0.91). Similarly, Table 4 from Note: This discussion (or parts of it) may be merged
Jones et al. (2006) paper shows that there is no statistically with/included in the intro.
signficant difference between word pairs that are collocates Semantic similarity literature:
performance in the task. After partialing out the effect of
FOS, there is no partial correlation between the SOS seman-
tic neighbourhood and RT. Conversely, there is a significant
partial correlation between the FOS semantic neighbourhood
and RT after partialing out the effect of SOS.
Discussion. From this analysis, we replicate the finding from
Buchanan et al. (2001) that semantic processing is indeed
a significant factor in the LDT. However, it is apparent that
the syntagmatic semantic neighbourhood density is facilita-
tory for lexical processing while there is no such evidence for
the paradigmatic semantic neighbourhood density. This re-
sult suggests that, that the single-word lexical decision task,
a word that has a relatively high level of syntagmatic similar-
ity with its closest neighbours, such as ‘school’ or ‘cheese’,
will be recognized more quickly than words that do not (e.g.,
Figure 1: Partial Correlations of First Order Semantic Dis- ‘theorem’). This contrasts work by Buchanan et al. (2001)
tance and Second Order Semantic Distance with LDT Reac- where elements of the syntagmatic neighbourhood as well as
tion Time. Error bars show 95% confidence intervals. Second the paradigmatic neighbourhood are conflated (cf. Figure 1
Order correlation n.s. in Buchanan et al. (2001)).
FS: this was a point that we thought might be brought
up by reviewers in our meeting on January 13th. One
• paper 1 and what’s important about it possible explanation for this result is that the FOS metric
LDT and semantic priming literature: encapsulates both paradigmatic and syntagmatic similarity,
• paper 1 and what’s important about it rendering the SOS metric extraneous. Table 1 clearly shows
Distributional semantics literature: that the two metrics are representing different relations. FS:
I haven’t created this table yet. It will just be a table or
• paper 1 and what’s important about it
figure of the different nearest neighbours for words by FOS
Materials and Methods metric vs. SOS metric Roller et al. (2016) and Melamud et
al. (2015) showed that leveraging both FOS, to represent
• distributional similarity method; corpus for training syntagmatic similarity, and SOS, to represent paradigmatic
• BLP; SPP similarity resulted in significant increases in performance.
• Multiple regression analyses; independent and dependent If FOS encapsulated both types of similarity, leveraging
variables solely FOS would have been sufficient. FS: I could also plot
the Jaccard similarity between the FOS nearest-neighbours
Results and SOS-nearest-neighbours as a function of the number of
Characterization of the Semantic neighbours we consider, if the argument in this paragraph
seems insufficient.
Neighbourhood
Using simple and effective distributional similarity metrics, FS: This isn’t a good transition Considering that syntag-
FOS and SOS, our goal is to determine whether increased matic neighbourhood density had a significant partial corre-
densities of both paradigmatic and syntagmatic neighbour- lation in single-word LDT, we hypothesized that syntagmatic
hoods are facilitatory for lexical processing. We hypothesize similarity should also be a significant factor in predicting RT
that both paradigmatic density and syntagmatic density would for the primed LDT.
have significant effects on the LDT RT.
Setup. We evaluate the partial correlations of syntagmatic Paradigmatic and Syntagmatic Similarity in
density and paradigmatic density with RT results from the
BLP. We compute these correlations on the 3,361 that appear
Semantic Priming
in all three of the BLP, the SPP, and the vocabulary of the Specifically, our hypothesis is that both syntagmatic similar-
word2vec model. We partial out the effects of length, log ity should be significant in predicting priming performance
frequency (calculated from the SUBTL corpus), and ortho- at the item level for words that are paradigmatically-related
graphic neighbourhood. and words that are syntagmatically-related. For instance, the
Results. Upon partialling out the effect of the syntagmatic word fork primed by the paradigmatically-related word spoon
neighbourhood, Figure 1 shows that there is no significant should be facilitated by the paradigmatic similarity and the
correlation between the paradigmatic neighbourhood and the syntagmatic similarity between the two words (i.e. the ten-
RT in the LDT task. Further, we see that the semantic neigh- dency for the two words to co-occur in the same sentence).
bourhood as defined by the FOS drives the facilitation of Conversely, a (prime, target) pair that only has a paradig-
matic relation (e.g. coat and gown) should obtain less fa-
cilitation. There is support for this intuition in the seman-
tic priming literature (Chiarello et al., 1990). Since the dis-
tributional similarity methods we employ, FOS and SOS,
are representative of the two types of similarity, both met-
rics should be significant factors in predicting primed RT for
paradigmatically-related words. So, we delineate the effects
of paradigmatic and syntagmatic similarity on (prime, target)
pairs, within the SPP, that are either syntagmatically-related
or paradigmatically-related.
Paradigmatic and Syntagmatic Relations in the SPP
We first show that FOS and SOS metrics are representative of
paradigmatic and syntagmatic relations, respectively, within
the SPP.
While the interpretations of these metrics are readily ac- Figure 2: Target words that are paradigmatically related to
cepted in the NLP literature (Roller & Erk, 2016; Melamud their respective prime are ranked higher by the SOS met-
et al., 2015), it is not apparent that this difference would be ric than the FOS metric. Conversely, target words that are
reflected in the smaller SPP dataset. syntagmatically related to their respective primes are ranked
Setup. We show that there is a statistically significant dif- higher by the FOS metric. One-sided paired t-tests show sig-
ference between FOS and SOS within the SPP. To show this nificant differences in all categories. FS: Maybe I should add
difference, we selected categories of word pairs in the SPP sample sizes for all of the categories in this figure
that are representative of paradigmatic and syntagmatic rela-
tionships, respectively.
We chose synonyms (e.g., frigid, cold), antonyms (e.g. hot, of paradigmatic similarity for pairs of words in the SPP.
cold), superordinates (e.g. sculpture, artwork), and categories This result enables us to evaluate whether both paradig-
(e.g. actor, actress) as relationships that are representative of matic and syntagmatic similarity are significant factors in
paradigmatic relationships. We chose these categories to be accounting for performance in the primed LDT. Specifi-
reflective of paradigmatic similarity since it is clear that ei- cally, we can evaluate the effects of both types of similarity
ther word in each respective pair can replace the other in a on paradigmatically-related pairs and syntagmatically related
sentence. Meanwhile, FPA (e.g. help, wanted) and BPA (e.g. pairs.
wanted, help) are, by definition, representative of syntagmatic
relationships. The aforementioned categorizations of associ- Analysis on Related Pairs
ations are recorded in the SPP. Setup.
For each prime word in these categories, we rank a sub- Results. Figure 3 shows that syntagmatic similarity is
set of 1,594 target words from the SPP from 1 (most simi- a significant factor for both paradigmatically-related and
lar within the DSM) to 1,594 (least similar within the DSM). syntagmatically-related words. Syntagmatic similarity is im-
The subset of target words is the intersection of words within portant for words that don’t necessarily have a syntagmatic
our ‘word2vec‘ model vocabulary and all of the 1,661 target association. That is, both degree of substitutability and degree
words within the SPP. of co-occurrence are of (independent) importance to lexical
For the prime words in the paradigmatic categories, we ex- processing time. For instance, it is important (to processing
pect that the mean ranking of target words will be assessed time) that ‘forever’ and ‘always’ are words that are highly
higher by the FOS metric than by the SOS metric. Conversely substitutable and that they also have a greater than chance co-
for the syntagmatic categories, we expect that the mean rank- occurrence ‘forever and always’).
ing of target words will be assessed higher by the SOS metric Discussion. This result is significant since other analyses of
than by the FOS metric. E.g., given the word ‘frigid’, we ex- priming with distributional models have suggested that mea-
pect the synonym ‘cold’ to be ranked highly by the SOS met- sures of paradigmatic similarity are important in capturing
ric while the phrasal associate ‘temperature’ is given a (rela- priming effects of paradigmatic relations while measures of
tively) lower rank. We expect an opposite pattern of rankings syntagmatic similarity are not (Jones et al., 2006; Günther et
by the FOS metric. al., 2016). From the results that we present, it is evident that
Results. Figure 2 shows that this is indeed the case. This both measures of similarity are significant factors in reacting
result shows that the difference between FOS and SOS is sta- to target words that were primed with a paradigmatically re-
tistically significant within the SPP, and that the differences lated word.
are in the expected direction. As expected, paradigmatic similarity does not have a sig-
Discussion. This result shows that the FOS is representa- nificant correlation with reaction time for phrasal associates
tive of syntagmatic similarity while SOS is representative since phrasal associates are generally not substitutable (e.g.
Figure 4: Density plots for related words and unrelated words
(blue and orange curves). The reaction time for a related pair
in the SPP is, on average, lower than reaction time for an
unrelated pair (unrelated prime, target). Since both seman-
tic similarity measure relatedness, they should both be im-
portant in depicting the priming effect, after controlling for
covariates. Paired t-tests show significant differences in both
categories: FS: t-test isn’t the right measure for this since the
sample size is so large. I’ll use Cohen’s d instead.
Figure 3: Partial correlation of semantic similarity measures
(FOS and SOS) with RTs for paradigmatically and syntag-
matically related words. Lexical characteristics and single-
word LDT RT for both the prime and the target are partialed
out. In addition, the human measures of similarity of back-
ward and forward association strength are partialled out. FS:
I’ll reproduce this result so that the green marker represents
the effect of partialing out the target covariate instead of the
semantic density, as discussed on January 17th.

(‘honey’, ‘bee’)). Thus, the SOS measure would not be mean-


ingful between two phrasal associates, resulting in a null par-
tial correlation.
We partial out the effects of the single-word LDT since,
as shown in the previous section, the similarity measures are
strongly correlated with single-word reaction time for associ-
ated words Thus, these partial correlations depict the corre-
lation of distributional similarity measures with primed reac- that an unrelated pair would have a low semantic similarity,
tion time. Since previous analyses of priming (Günther et al., and consequently a higher reaction time. Figure 4(a) illus-
2016; Mandera et al., 2017) did not include this significant trates this intuition.
covariate in their analyses, this is the first reliable analysis Other models of priming that include distributional simi-
that shows distributional similarity measures factor into item- larity (Mandera et al., 2017; Günther et al.,2016) have shown
level regression analyses of priming. that distributional similarity is indeed a significant factor in
FS: Is this sufficient justification of the small effect size? predicting reaction times for the LDT. There are 3 unresolved
Note that while the effect sizes are small, they are commen- questions following previous analyses: (1) since such analy-
surate with effect sizes of forward and backward associative ses only include distributional measurements of paradigmatic
strength measures which are known to be a reliable factor of similarity, it is not clear that syntagmatic similarity would be
priming (Hutchison et al., 2008). a significant factor in predicting RT; (2) it is not clear whether
the significant effects of paradigmatic similarity were actually
Analysis on Related and Unrelated Pairs due to the unique variance captured by syntagmatic similar-
ity; (3) since previous analyses did not control for the effect of
The previous section showed that semantic space measures of
the semantic neighbourhood of the target word, it is unclear
similarity are significant predictors of primed RT for related
whether a significant effect of any distributional similarity be-
pairs of words. That is, as semantic similarity increases be-
tween prime and target would be found in the first place.
tween a target and a related prime, RT decreases. Following
Mandera et al. (2017), we conduct a regression analysis on
Conclusions
all of the pairs within the SPP to determine whether both FOS
and SOS were significant factors in predicting reaction time Main contributions:
for related pairs as well as unrelated pairs. The expectation is • contrib 1
• contrib 2
• ...
Discussion points:
• disc 1
• disc 2
• ...
Limitations of the wor:
• lim 1
• lim 2
• ...

References

You might also like