Akira Utsumi

Interpretive Diversity Explains Metaphor–Simile Distinction

A number of properties—aptness, topic-vehicle similarity, vehicle conventionality—have recently been used to explain a metaphorsi-mile distinction. This paper argues that interpretive diversity better explains a metaphor-simile distinction than these properties. Interpretive diversity refers to the semantic richness of the figurative interpretation of a topic-vehicle pair and is determined depending on both the number of features involved in the interpretation and the uniformity of salience distribution of those features. The interpretive diversity view predicts that interpretively more diverse pairs should be easier to comprehend via a categorization process, and thus the preference for and the relative comprehensibility of the metaphor form should be greater. Two experiments demonstrated that, as predicted, interpretive diversity was correlated positively with metaphor preference (Experiment 1) and with the relative comprehensibility of the metaphor form compared to the simile form (Experiment 2). Furthermore, interpretive diversity was found to be more important in explaining metaphor-simile distinction than aptness, similarity, and conventionality.

Computational Exploration of Metaphor Comprehension Processes Using a Semantic Space Model

Recent metaphor research has revealed that metaphor comprehension involves both categorization and comparison processes. This finding has triggered the following central question: Which property determines the choice between these two processes for metaphor comprehension? Three competing views have been proposed to answer this question: the conventionality view (Bowdle & Gentner, 2005), aptness view (Glucksberg & Haught, 2006b), and interpretive diversity view (Utsumi, 2007); these views, respectively, argue that vehicle conventionality, metaphor aptness, and interpretive diversity determine the choice between the categorization and comparison processes. This article attempts to answer the question regarding which views are plausible by using cognitive modeling and computer simulation based on a semantic space model. In the simulation experiment, categorization and comparison processes are modeled in a semantic space constructed by latent semantic analysis. These two models receive word vectors for the constituent words of a metaphor and compute a vector for the metaphorical meaning. The resulting vectors can be evaluated according to the degree to which they mimic the human interpretation of the same metaphor; the maximum likelihood estimation determines which of the two models better explains the human interpretation. The result of the model selection is then predicted by three metaphor properties (i.e., vehicle conventionality, aptness, and interpretive diversity) to test the three views. The simulation experiment for Japanese metaphors demonstrates that both interpretive diversity and vehicle conventionality affect the choice between the two processes. On the other hand, it is found that metaphor aptness does not affect this choice. This result can be treated as computational evidence supporting the interpretive diversity and conventionality views.

A Complex Network Approach to Distributional Semantic Models

A number of studies on network analysis have focused on language networks based on free word association, which reflects human lexical knowledge, and have demonstrated the small-world and scale-free properties in the word association network. Nevertheless, there have been very few attempts at applying network analysis to distributional semantic models, despite the fact that these models have been studied extensively as computational or cognitive models of human lexical knowledge. In this paper, we analyze three network properties, namely, small-world, scale-free, and hierarchical properties, of semantic networks created by distributional semantic models. We demonstrate that the created networks generally exhibit the same properties as word association networks. In particular, we show that the distribution of the number of connections in these networks follows the truncated power law, which is also observed in an association network. This indicates that distributional semantic models can provide a plausible model of lexical knowledge. Additionally, the observed differences in the network properties of various implementations of distributional semantic models are consistently explained or predicted by considering the intrinsic semantic features of a word-context matrix and the functions of matrix weighting and smoothing. Furthermore, to simulate a semantic network with the observed network properties, we propose a new growing network model based on the model of Steyvers and Tenenbaum. The idea underlying the proposed model is that both preferential and random attachments are required to reflect different types of semantic relations in network growth process. We demonstrate that this model provides a better explanation of network behaviors generated by distributional semantic models.

Evaluating the performance of nonnegative matrix factorization for constructing semantic spaces: Comparison to latent semantic analysis

This study examines the ability of nonnegative matrix factorization (NMF) as a method for constructing semantic spaces, in which the meaning of each word is represented by a high-dimensional vector. The performance of two tests (i.e., a multiple-choice synonym test and a word association test) is compared between NMF and latent semantic analysis (LSA), which is the most popular method for constructing semantic spaces. As a result, it was found that NMF did not outperform LSA in either test. This finding indicates that NMF is less effective in acquiring word meanings than expected in the literature; in other words, the finding provides evidence for the ability of LSA to represent semantic meanings. Some properties of NMF were also revealed with reference to its ability to represent word meanings; the random initialization was superior to the SVD-based initialization, and the Euclidean distance is more appropriate for the objective function of NMF than the KL-divergence. In addition, it was shown that the inner product was a more appropriate method for measuring the syntagmatic similarity in a semantic space model, while the cosine was a better method for computing the paradigmatic similarity.

The role of prosody and context in sarcasm comprehension: Behavioral and fMRI evidence.

A hearers perception of an utterance as sarcastic depends on integration of the heard statement, the discourse context, and the prosody of the utterance, as well as evaluation of the incongruity among these aspects. The effect of prosody in sarcasm comprehension is evident in everyday conversation, but little is known about its underlying mechanism or neural substrates. To elucidate the neural underpinnings of sarcasm comprehension in the auditory modality, we conducted a functional MRI experiment with 21 adult participants. The participants were provided with a short vignette in which a child had done either a good or bad deed, about which a parent made a positive comment. The participants were required to judge the degree of the sarcasm in the parents positive comment (praise), which was accompanied by either positive or negative affective prosody. The behavioral data revealed that an incongruent combination of utterance and the context (i.e., the parents positive comment on a bad deed by the child) induced perception of sarcasm. There was a significant interaction between context and prosody: sarcasm perception was enhanced when positive prosody was used in the context of a bad deed or, vice versa, when negative prosody was used in the context of a good deed. The corresponding interaction effect was observed in the rostro-ventral portion of the left inferior frontal gyrus corresponding to Brodmanns Area (BA) 47. Negative prosody incongruent with a positive utterance (praise) activated the bilateral insula extending to the right inferior frontal gyrus, anterior cingulate cortex, and brainstem. Our findings provide evidence that the left inferior frontal gyrus, particularly BA 47, is involved in integration of discourse context and utterance with affective prosody in the comprehension of sarcasm.

Word Vectors and Two Kinds of Similarity

This paper examines what kind of similarity between words can be represented by what kind of word vectors in the vector space model. Through two experiments, three methods for constructing word vectors, i.e., LSA-based, cooccurrence-based and dictionary-based methods, were compared in terms of the ability to represent two kinds of similarity, i.e., taxonomic similarity and associative similarity. The result of the comparison was that the dictionary-based word vectors better reflect taxonomic similarity, while the LSA-based and the cooccurrence-based word vectors better reflect associative similarity.

Ineluctable background checking on social networks: Linking job seeker's résumé and posts

A growing source of concern is that the privacy of individuals can be violated by linking information from multiple sources. For example, the linking of a persons anonymized information with other information about that person can lead to de-anonymization of the person. To investigate the social risks of such linking, we investigated the use of social networks for background checking, which is the process of evaluating the qualifications of job seekers, and evaluated the risk posed by the linking of information the employer already has with information on social networks. After clarifying the risk, we developed a system that links information from different sources: information extracted from a job seekers résumé and anonymous posts on social networks. The system automatically calculates the similarity between information in the résumé and in the posts, and identifies the job seekers social network accounts even though the profiles may have been anonymized. As a part of our system, we developed a novel method for quantifying the implications of terms in a résumé by using the posts on social networks. In an evaluation using the résumés of two job seekers and the tweets of 100 users, the system identified the accounts of both job seekers with reasonably good accuracy (true positive rate of 0.941 and true negative rate of 0.999). These findings reveal the real social threat of linking information from different sources. Our research should thus form the basis for further study of the relationship between privacy in social networks and the freedom to express opinions.

The role of the amygdala in incongruity resolution: the case of humor comprehension

ABSTRACT A dominant theory of humor comprehension suggests that people understand humor by first perceiving some incongruity in an expression and then resolving it. This is called “the incongruity-resolution theory.” Experimental studies have investigated the neural basis of humor comprehension, and multiple neural substrates have been proposed; however, the specific substrate for incongruity resolution is still unknown. The reason may be that the resolution phase, despite its importance in humor comprehension, has not been successfully distinguished from the perception phase because both phases occur almost simultaneously. To reveal the substrate, we conducted a functional magnetic resonance study using 51 healthy participants. We used a humor-producing frame of “Given A, I’d say B, because C” so as to focus on the resolution phase independently by suspending humor processing just after the perception phase. This frame allowed us to separate the two phases. Based on our results, incongruity resolution evoked positive emotion and activated the left amygdala, which is known to be related to positive emotion. On the basis of these findings, we argue that the amygdala plays an important role in humor comprehension, considering its functional role in emotional evaluation, particularly the relevance detection for incoming stimuli.

Extracting Causal Knowledge by Time Series Analysis of Events

Causal knowledge is important for decision-making and risk aversion. However, it takes much time and effort to extract causal knowledge manually from a large-scale corpus. Therefore, many studies have proposed several methods for automatically extracting causal knowledge. These methods use a variety of linguistic or textual cues indicating causality on the basis of the assumption that causally related events tend to co-occur within a document. However, because of this assumption, they cannot extract causal knowledge that is not explicitly described in a document. Therefore, in this paper, we propose a novel method for extracting causal knowledge not explicitly described in a document using time series analysis of events. In our method, event expressions, which are represented by a pair of a noun phrase and a verb phrase, are extracted from newspaper articles. These extracted event expressions are clustered into distinct events, and the burst of the appearance of these clustered events is detected. Finally, using the time series data with burst, it is judged whether any event pairs have a causal relationship by Granger Causality test. We demonstrate through an evaluation experiment that the proposed method successfully extracts valid causal knowledge, almost all of which cannot be extracted by existing cue-based methods.

A New Partitioning Method for the IDS Method

The ink drop spread (IDS) method is a modeling technique based on the idea of soft computing. This method divides a multi-input-single-output (MISO) target system into multiple single-input-single-output (SISO) systems, and models each SISO system by plotting the input/output data. The IDS method combines the modeling results of SISO systems to model the target. It is important for the IDS method to decide appropriate partitions of the target system in order to accurately model the target. Existing partitioning methods divide each input domain independently of the other inputs, and thus generate unnecessary SISO systems. In this article, we propose a new partitioning method for the IDS method, which divides the input domains by considering the relationship between inputs. We also show that our method can achieve better performance with less partitions than existing methods.