Compound interpretation is based on analogy

Tian Shen; Harald Baayen

cs.CLOct 1, 2026

Compound interpretation is based on analogy

Authors: Tian Shen, Harald Baayen

Organizations: School of Foreign Languages, Northwest University, Xi’an, China · Department of Quantitative Linguistics, University of Tübingen, Tübingen, Germany

Abstract

How compound meanings are best predicted from constituent meanings remains a central question in computational models of lexical semantics. Comparing different computational models provides a way to evaluate alternative accounts of how semantic information is combined during compound comprehension. We propose a new model, the Compound Analogy Model (CAM), that predicts a compound's embedding by adding its constituent embeddings together with the average shift vectors of the two constituents' compound families. The resulting model is parameter-free and exploits local analogical structure in the semantic space. We evaluated CAM against the CAOSS model on Mandarin Chinese compounds. CAM consistently achieved higher prediction accuracy than CAOSS on both training and held-out data, with the exception of three-character compounds, for which analogical generalization is constrained by both small constituent families and a pronounced imbalance in family size between the two constituents. The advantage of CAM remained when evaluation was based on frequency-defined train-test splits that better approximate generalization from familiar to novel compounds. To assess the cognitive plausibility of the two models, we further examined whether model-derived semantic measures predict visual lexical decision latencies for two-character compounds. Predictors derived from CAM provided improved prediction for response latencies compared to predictors derived from the CAOSS model. These findings indicate that compound meaning is better characterized as local analogical generalization than as the application of a learned global linear transformation, and demonstrate that analogical semantic structure provides a cognitively plausible basis for compound comprehension.

Figures & tables

relatively transparent			relatively opaque
书店店	shu1dian4 dian4	bookshop inn, shop, store	秘书秘	mi4shu1 mi4	secretary, assistant hidden, secret
读书读	du2shu1 du2	read, study, attend school read, attend school	草书草	cao3shu1 cao3	cursive script grass, straw, draft, careless, hasty
书架架	shu1jia4 jia4	bookshelf support, frame, rack	背书背	bei4shu1 bei4	recite a lesson for memory, endorse back, unlucky, recite, violate

Table 1 : Examples of relatively transparent (left) and opaque (right) compounds with character 书 ( shu1 , ‘book’).

Panel of Figure 2 — Figure 2 : Core concepts of CAM. (a) A shift vector for a given constituent, here 水 ( shui3 , ‘water’), is obtained by subtracting the embedding of the other constituent, here 洪 ( hong2 , ‘huge’), from the compound embedding, here 洪水 ( hong2shui3 , ‘flood water’). (b) Compounds with 水 (‘water’, in blue) cluster (using t-SNE) in the embedding space. The points in grey represent compounds with nine other high-frequency monomorphemic words: 火 (‘fire’), 金 (‘gold’), 木 (‘wood’), 土 (‘earth’), 天 (‘sky’), 道 (‘road’), 心 (‘heart’), 人 (‘man’), and 名 (‘name’). (c) The corresponding shift vectors of 水 likewise cluster in the corresponding shift space.

Panel of Figure 5 — Figure 5 : Geometric illustration of CAM. Constituent embeddings $\bm{s}_{c_{1}}$ and $\bm{s}_{c_{2}}$ act as anchor points, while centroid shift vectors $\bar{\bm{d}}_{c_{1},1}$ and $\bar{\bm{d}}_{c_{2},2}$ provide position-specific adjustments. The mean of the two transcended contituents $((\bm{s}_{c_{1}}+\bar{\bm{d}}_{c_{1},1})+(\bm{s}_{c_{2}}+\bar{\bm{d}}_{c_{2},2}))/2$ yields the predicted compound representation $\hat{\bm{s}}_{w}$ .

Figure 6 : Prediction accuracy for CAM and CAOSS models for training and test data. For each of 30 runs, data were randomly split into training (90%) and test (10%) sets.

Figure 7 : Prediction accuracy of CAM and CAOSS under a frequency-based split with the lowest-frequency 10% of compounds in the test set and the remaining 90% in the training set.

Figure 8 : Boxplots of accuracy for CAM (red) and CAOSS (blue) models trained on only the words with a given length (2, 3, and 4 characters). CAM outperformed CAOSS except for the held-out data with three-character compounds.

Figure 9 : Prediction accuracy of CAM and CAOSS for compounds of different word lengths under a frequency-based split, with the lowest-frequency 10% of compounds assigned to the test set and the remaining 90% to the training set. CAM significantly outperformed CAOSS only for the held-out data of two-character compounds.

Word length	Dataset	CAM	CAOSS	$\chi^{2}(1)$	$p$
2-char	Train	0.37	0.13	4455.7	$<.0001$
	Test	0.17	0.09	111.26	$<.0001$
3-char	Train	0.94	0.71	1007.8	$<.0001$
	Test	0.37	0.41	1.40	0.24
4-char	Train	0.99	0.94	52.90	$<.0001$
	Test	0.69	0.56	2.69	0.10

Table 2 : Comparison of CAM and CAOSS accuracy for training and test splits based on frequency, using proportions tests.

Figure 10 : Boxplots for CAM (red) and CAOSS (blue) accuracies nouns by word length, for 30 cross-validation runs.

Figure 11 : Distributions of the vector lengths (L2-norms) for CAOSS (left panel) and CAM (right panel), for transcended (blue) and observed (red) constituent embeddings, as well as for the predicted and observed compound vectors.

Figure 12 : Absolute-value heatmap of the global CAOSS transformation matrix $\bm{W}$ for two-character Mandarin compounds. The two blocks correspond to the constituent-specific sub-matrices $\bm{W}_{1}$ and $\bm{W}_{2}$ . Both matrices are diagonally dominant, with brighter diagonal entries indicating larger transformation weights. The diagonal elements account for approximately 72.8% of the Frobenius norm of $\bm{W}$ .

A. parametric coefficients	Estimate	Std. Error	t-value	p-value
Intercept	-1.1593	0.0333	-34.7889	$<$ 0.0001
${P}_{c1}^{CAOSS}$	-0.0727	0.0226	-3.2211	0.0013
${P}_{c2}^{CAOSS}$	-0.1870	0.0312	-5.9939	$<$ 0.0001
B. smooth terms	edf	Ref.df	F-value	p-value
$\mathrm{s}(\log f(w))$	5.3077	6.3023	933.7357	$<$ 0.0001
te ( $V(c_{2}),S_{C}^{CAOSS}$ )	13.1744	15.6219	5.1673	$<$ 0.0001

Table 4 : Summary of a GAM fitted to visual lexical decision latencies with predictors derived from the CAOSS model (AIC: -11570.1).

Figure 13 : Partial effects of the smooth terms in the GAM fitted to visual lexical decision latencies using predictors derived from the CAOSS model. The left panel shows the nonlinear effect of log-transformed word frequency, with the effect levelling off at higher frequency levels. The middle panel displays the observed distribution of the two predictors. The right panel shows the contour plot of the tensor smooth for the interaction between right constituent family size and cosine similarity. Colours indicate predicted response latencies, with darker blue regions representing shorter latencies and red regions representing longer latencies.

	df	AIC	$\Delta$ AIC
$f(w)$	30.4	-7075.4	4494.7
$V(c_{2})$	19.4	-11512.7	57.4
$P_{c2}^{CAOSS}$	23.3	-11537.2	32.9
$P_{c1}^{CAOSS},P_{c2}^{CAOSS}$	22.8	-11537.3	32.8
$S_{C}^{CAOSS}$	18.5	-11560.3	9.8
$P_{c1}^{CAOSS}$	27.6	-11561.2	8.9

Table 5 : Variable importance, estimated by the change in AIC when a predictor is removed from the GAM model that is informed by CAOSS-based predictors. ( 22 ).

Panel of Figure 14 — Figure 14 : The geometry of CAOSS proximity measures. Left panel: transcended vectors of equal length; right panel: transcended vectors of different lengths.

A. parametric coefficients	Estimate	Std. Error	t-value	p-value
Intercept	-1.3555	0.0013	-1018.1199	$<$ 0.0001
B. smooth terms	edf	Ref.df	F-value	p-value
$\text{s}(\log(f(w))$	5.4745	6.4724	852.9476	$<$ 0.0001
$\text{te}(V(c_{2}),S_{C}^{CAM}$	14.5436	17.2955	3.4391	$<$ 0.0001
$\text{te}(P_{c1}^{CAM},Pr_{c2}^{CAM})$	3.8423	4.5405	17.2632	$<$ 0.0001

Table 6 : Summary of a GAM fitted to visual lexical decision latencies with predictors grounded in CAM (AIC: -11597.45).

Figure 15 : Partial effects of the smooth terms in a GAM fitted to visual lexical decision latencies using predictors derived from the CAM model. The first panel in first row shows the nonlinear effect of log-transformed word frequency. The middle and right panels in the first row display the observed distributions of the predictors involved in the two tensor-product smooths: right constituent family size ( $V(c_{2})$ ) with cosine similarity ( $S_{C}^{CAM}$ ), and the proximity measures of the first and second constituents ( $P_{c_{1}}^{CAM}$ and $P_{c_{2}}^{CAM}$ ), respectively. The second row presents the corresponding contour plots of the tensor smooths. Colours indicate predicted response latencies, with bluer shades representing shorter response latencies and lighter/redder shades representing longer response latencies.

	df	AIC	$\Delta$ AIC
$f(w)$	25.6	-7333.5	4264.0
$P_{c1}^{CAM},P_{c2}^{CAM}$	22.7	-11527.9	69.6
$P_{c1}^{CAM}$	23.6	-11538.9	58.6
$V(c_{2})$	14.1	-11567.1	30.4
$S_{C}^{CAM}$	19.4	-11584.9	12.5
$P_{c2}^{CAM}$	29.1	-11589.8	7.7

Table 7 : Variable importance, estimated by the change in AIC when a predictor is removed from the GAM model based on CAM predictors ( 24 ).

Figure 16 : Comparison of variable importances for CAM-based (blue) and CAOSS-based (orange) predictors, as well as second constituent family size, assessed with the increase in AIC when a predictor is withheld from the GAM.

Explore similar work

Losing My Composure: Predicting Compositionality Over Time
Jul 13, 2026Chris Jenkins, Emma Raimundo Schulz, Filip Miletić +1Compositionality
University of Stuttgart
From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options
Aug 13, 2026Obed Junias, Maria Leonor PachecoLogic Answer
University of Colorado Boulder
Structural Ranking of the Cognitive Plausibility of Computational Models of Analogy and Metaphors with the Minimal Cognitive Grid
May 2, 2026Alessio Donvito, Antonio LietoCognitive Science Analogy
University of Bari, Italy · University of Salerno, CIIT Lab / ICAR-CNR, Italy

Jul 13, 2026, cs.CLJ/K move · Enter open · S save

Losing My Composure: Predicting Compositionality Over Time

Chris Jenkins, Emma Raimundo Schulz, Filip Miletić, Sabine Schulte im Walde

University of Stuttgart

Compositionality

Open paper AskPDF arXiv

We explore the phenomenon of semantic change of German and English noun compounds, with the objective of investigating and modeling gradual changes of meanings and degrees of compositionality in the past and over time. To do so, we introduce the Compositionality Trend Prediction task, which is evaluated against a novel dataset of in-context compositionality ratings sampled across several decades of diachronic corpora for 23 German and 26 English target compounds, uniquely providing per-decade ratings and corresponding trends over time. These per-decade compositionality ratings allow us to investigate empirically untested hypotheses of generalized trends in compositionality over time, such as the idea that compounds should become less compositional (less transparent) over time. Beyond our empirical observations from the diachronic compositionality annotations, we perform experiments with semantic vector representations of varying complexity, as well as several temporal granularities for training these representations on diachronic data, resulting in about 100 models of each representation type, each covering a different 1--5 decade slice of a diachronic corpus. Contrary to the decisive tendency posited in the literature, we find only a small negative trend in compositionality over time in our target compounds. In our computational experiments, we find that using models trained on narrow time slices of diachronic data (single decades, or incrementally expanding temporal windows) align better with the per-decade compositionality ratings than those trained on an entire half-century window, the latter setting being an analog for the prevalent modeling approach of training representations on an entire half of a corpus' data. Additionally, we find static representations to be competitive with contextual representations in the Compositionality Trend Prediction task.

Abstract

Explore similar work

Losing My Composure: Predicting Compositionality Over Time

From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options

Structural Ranking of the Cognitive Plausibility of Computational Models of Analogy and Metaphors with the Minimal Cognitive Grid