Word Similarity Perception: an Explorative Analysis Alice Ruggeri (ruggeri@di.unito.it) Centre for Cognitive Science, University of Turin Loredana Cupi (loredana.cupi@unito.it) Department of Foreign Languages and Literatures and Modern Cultures, University of Turin Luigi Di Caro (dicaro@di.unito.it) Department of Computer Science, University of Turin Abstract From a computational perspective, being words ambigu- ous by nature, the disambiguation process (i.e., Word Sense Natural language is a medium for expressing things belonging to conceptual and cognitive levels, made of words and gram- Disambiguation) is one of the most studied tasks in Compu- mar rules used to carry semantics. However, its natural am- tational Linguistics. To make an example, the term count can biguity is the main critical issue that computational systems mean many things like nobleman or sum. Using contextual are generally asked to solve. In this paper, we propose to go beyond the current conceptualization of word similarity, i.e., information, it is often possible to make a choice. Again, the building block of disambiguation at computational level. this choice is done by means of comparisons among contexts, First, we analyze the origin of the perceived similarity, study- that are still made of words. In other terms, we may state that ing how conceptual, functional, and syntactic aspects influence its strength. We report the results of a two-stages experiment the computational part of almost all computational linguistics showing clear similarity perception patterns. Then, based on research is about the calculus of matching scores between lin- the insights gained in the cognitive tests, we developed a com- guistic items, i.e., words similarity. But what’s behind words putational system that automatically predicts word similarity reaching high levels of accuracy. similarity? There exist many annotated data related to similarity and Introduction relatedness between words, like wordsim-353 (Finkelstein Words are symbolic entities referring to something which fills et al.,2001) and SimLex-999 (Hill, Reichart, & Korho- a portion of an autopoietic space made of conceptual, cog- nen,2014). A large part of the proposed computational sys- nitive and contextual information. These three aspects are tems aims at finding relatedness between words, instead of fundamental to understand the meaning ascribed to linguistic similarity. Relatedness is more general than similarity, since expressions. it refers to a generic correlation (like cradle and baby, that One of the most important building block in almost all are words representing dissimilar concepts which, however, Computational Linguistics tasks is the computation of sim- share similar contexts). ilarity scores between texts at different levels: words, sen- One problem of similarity, as often faced in literature or tences and discourses. Since manual numeric annotations annotated in datasets, is that it cannot be a static value. In- of word-word similarity revealed a low agreement between deed, as the authors of these resources state in their works, the the annotators, cognitive studies can help improve computa- agreement between the annotators is usually not high (around tional systems by discovering what lies behind the perception 50-70%). The reason is trivial, however: people can give dif- of similairty between words and their referenced concepts. ferent degrees of importance with respect to the specific char- The concept of similarity has been extensively studied in the acteristics of the concepts to compare. If we ask one to say Cognitive Science community, since it is fundamental in the how much dog is similar to cat, the right answer can only be human cognition. We tend to rely on similarity to generate “it depends”. While we can all agree about the fact that the inferences and categorize objects into kinds when we do not concept dog is quite similar to cat, we cannot say 0.7 rather know exactly what properties are relevant, or when we can- than 0.9 (in the range [0,1]) with certainty. Different aspects not easily separate an object into separate properties. when can be taken into account: are we measuring the form of the specific knowledge is available, then a generic assessment of animal, or its behaviour? In both cases, it depends on which similarity is less relevant (G. L. Murphy & Medin,1985). part of the animal and which actions we are considering to Since words co-occur in textual representations (mutually make a choice. For instance, dogs use to return thrown ob- influencing one each other) it is possible to make experiments jects. From this point of view, dogs and cats are dissimilar. on contextual information to analyze what and how influence the perception of similarity. Let us consider two words as two In the light of this, our contribution provides the basis for mental representations. The intersection between them can understanding what lies behind a similarity between words be seen as the context that may also help define the correct and their referenced concepts. First, we analyze syntactic, similarity of the words in a text. For example, sugar and salt conceptual and functional aspects of the similarity percep- can be easily associated to the context kitchen, whereas salt tion; then, we develop a computational system which is able and sea intersect in another part of the mental representations. to predict similarity by leveraging contextual information. 482 The Cognitive Experiment ception of similarity, we decided to keep concrete words In this work, we present two tests to analyze how linguistic only. constructions are perceived by humans in terms of strength of Semantic coherence Another criterion used for the selection semantic similarity and if there exists a functionality-based of the words was the level of semantic similarity between connection that has an influence on its perception. The ex- the word pairs to compare. To better analyze whether the periment was presented to 96 users, having different ages and functional aspect plays a significant role in the similarity professions, without any particular cognitive or linguistic dis- perception, we extracted conceptual and functional pairs order. of words which had similar semantic closeness according Test on single words to a standard semantic similarity calculation. In the light of this, we used a Latent Semantic Space calculated over The first test of the experiment regards the perception of the almost 1 million of documents coming from the collection similarity between single words1 . In particular, the goal was of literary text contained in the project Gutenberg page2 . to analyze how the users focus on the functional links be- The selected conceptual and functional word pairs had the tween the words, and more importantly if such functional- property of having a very close semantic similarity (the based similarity is a preferential perception channel com- score differences were less than 0,01 in a [0,1] range). pared to the conceptual-based one. Words are ambigous, and many resources have been re- The test was composed by three word pairs, to leave it sim- leased with the goal of defining all the possible senses ple and to be not affected by users tiredness. Then, instead of of a word (i.e., WordNet). Word Sense Disambiguation randomly selecting three different word pairs, we wanted to (Bhattacharyya & Khapra,2012) is the task of resolving the consider three cases in which the functional links between ambiguity of a word in a given context. Notice that, in our the words have distinct levels of importance. Our assumption experiment, we do not need any disambiguation of the words, was that the more the importance of the functional link be- since this process is embodied in the human cognition, thus tween two words in a pair, the more its perceived similarity the users of the test will autonomously represent their subjec- (and thus the user preferences with respect to the conceptual tive sense to associate to the words under comparison. word pair). For this reason we added a final criterion: Then, since we wanted to compare conceptual with func- Increasing relevance of the functional aspect To estimate tional preferences, we designed the test as a comparison be- the importance of the functional aspect that relates two tween two word pairs, one involving conceptually-related words we analyzed the number of actions (or verbs) in words and one with words linked by direct functionalities. which they are usually involved with. In our test, the func- To generalize, let us consider the words a, b, and c with the tional word pairs salt-water, nail-polish, and ring-finger conceptual word pair a-b and the functional word pair a-c. have a functional link of 0.0033 , 0.01014 , and 0.06255 re- The user is asked to mark the most similar word (among b spectively (see Table 1). These values are calculated in the and c) to associate to a, and so the most correlated word pair. following way: given the total number of existing verbs The users were not aware of the goal of the test and of the NV(rw) for the root word rw and the number of effective difference between the word pairs. usages EU(sw) with the second word of the pair sw, we Since words and actions present a high variability in terms computed the functional link Fl(rw, sw) of the functional of conceptual range (or their mental representation), we put pair as EU(sw) / NV(rw). particular attention to the choice of the word pairs, according to the following principles: Conceptual granularity If we think at the words object and Table 1: The chosen word pairs in the first test. thing, we probably do not have enough information to make significant comparisons due to their large and unde- Root word Conceptual Pair Functional Pair [F. link] fined conceptual boundaries. The same happens in cases rw rw - sw rw - sw [Fl(rw,sw)] when two words represent very specific concepts such as salt salt - sugar salt - water [0.003] lactose and amino acid. The word pairs of the proposed nail nail - finger nail - polish [0.0101] test have been selected by considering this constraint (and ring ring - necklace ring - finger [0.0625] so they include words which are not too specific nor too general). We then considered a backup test set with random word Concreteness Words may have direct links with concrete ob- pairs matching with the same above-mentioned criteria, col- jects such as “table” and “dog”. In other cases, words lecting a total of 24 answers on 24 different cases of word such as “justice” and “thought” represent abstract con- 2 http://promo.net/pg/ cepts. Since it is not clear how this may affect the per- 3 For the verbs “to put”, “to add” and “to get” 1 Notice that “similarity between words” is intended as the simi- 4 For the verbs “to apply” and “to use” larity between the concepts they bring to mind. 5 For the verbs “to put” and “to wear” 483 pairs such as the main one of Table 1. This was done to prove Interpretation of the results the reliability of the test, seeing whether the results and the In this section, we give a preliminary interpretation of the analyses show a similar trend, being independent from the se- results on the collected answers. lection of the words. The results of the whole test is described In the first test on single words, we can state that, gener- in the final part of this section. ally, conceptual and functional word pairs are differently per- ceived according to the importance of the funcional link in Test on Phrases the functional word pair. This shows that words and their ref- The second test of the experiment concerns the perception of erenced concepts are mainly compared in terms of conceptual the similarity between phrases, o multi-word linguistic con- similarity, but when there exists importafnt functionalities be- structions6 . The goal was to analyze how the syntactic con- tween them, this influences the users preference towards the text of a target word influences the perception of similarity functional word pair ??. among entire phrases. More in detail, we wanted to discover For example, the similarity of the sugar-salt pair results to possible differences of such perception along different syn- be stronger compared to the water-salt one, since the action to tactic roles. We considered a simple syntactic structure of the add/put the salt in the water is “a needle in a haystack” with type subject-verb-object. respect to all the actions related to water and salt indepen- Given a root sentence such as “Mario sings the song”, we dently. This means that there is no exclusive action between created three variations by changing the subject, the verb, and water and salt (i.e., there are many actions that involve wa- the direct object. For example, by changing “Mario” with ter). An opposite example is represented by the word pair “The bird” we obtained “The bird sings the song”. The com- ring-finger, since the action to put/wear the ring on the fin- plete set of replacements are shown in Table 3. ger is much more exclusive than in the previous case. Such We presented to the 96 users a total of 4 sentences (see Ta- preference could be explained by stating that all word pairs, ble 2), that with its 4 variations produce a total of 16 pairs of especially with words that underlie actions, have a strong vi- sentences to be analyzed by the users in terms of perceived sual representation that makes them quickly perceivable. similarity, as in the first test. For each sentence, the users had to indicate the degree of similarity of the original sentence Table 4: Results showing the percentage of preferences in the with one of its variation using a value in the range [0,10]7 . choice of the most (perceived) correlated word pairs of the 0 means no semantic similarity between the two phrases and first test. 10 means total equality. The grammatical changes made on the original sentences were chosen maintaining the semantic Case Word pairs N. of preferences % validity (i.e., all the sentences represent valid mental repre- 1a. salt - sugar 75 78% sentations). 1b. salt - water 21 22% 2a. nail - finger 44 46% 2b. nail - polish 52 54% Table 2: The chosen phrases in the second test. 3a. ring - necklace 16 17% Phrase ID Phrase 3b. ring - finger 80 83% (a) Mario sings the song (b) Alan drives the car (c) Alice writes the book (d) Marco does the homeworks Table 5: Results on the backup of the first test (with 24 differ- ent cases including 8 low-FL cases, 8 medium-FL cases and 8 high-FL cases). The results are in line with the ones of the main test shown in Table 4. Funct. Pair Pref. w.r.t conceptual pair Table 3: The word replacements for subjects (SC), verbs (VC) Funct. Pair (low FL) 1 out of 8 (12.5%) and direct objects (OC). Funct. Pair (medium FL) 5 out of 8 (62.5%) Replacement (a) (b) (c) (d) Funct. Pair (high FL) 7 out of 8 (87.5%) SC bird robot computer software VC writes cleans cleans gives OC verse band sheet pasta This result is also in line with what stated by (Cohen et al.,2002), i.e., words that have a functionality-based relation- ship can have a more complex visual component that makes such correlation weaker. 6 Even in this case, “similarity” is intended as the similarity be- In Figure 1, we show the users preferences for the second tween the concepts related to the phrases. 7 We used a [0,10] range instead of a [0,1] range as in the previous test. In the case of verb replacement (VC) we can notice a test because it represents a more human-understandable and intuitive high meaning change in terms of similarity perception (sim- votation. ilarity values close to 0), so the verb represents the real root 484 Table 6: Some of the existing relations in ConceptNet, with example sentences in English. Relation Example sentence IsA NP is a kind of NP. LocatedNear You are likely to find NP near NP. UsedFor NP is used for VP. DefinedAs NP is defined as NP. HasA NP has NP. HasProperty NP is AP. CapableOf NP can VP. ReceivesAction NP can be VP. Figure 1: Results of the second test, showing the change HasPrerequisite NP—VP requires NP—VP. scores in terms of word similairty perception after subject, MotivatedByGoal You would VP because you want VP. verb and object replacements. SC stands subject change, VC MadeOf NP is made of NP. for verb change, and OC for object change. ... ... of the mental representations. The case of the subject change (Spearman correlation). We leveraged ConceptNet to retrieve (SC) shows a less important decrease of similarity perception, the semantic information associated to the words of each pair, while the object change (OC) resulted to be the less relevant then keeping the intersection. For example, considering the syntactic role influencing the meaning of the whole phrase. pair rice-bean, ConceptNet returns the following set of se- mantic information for the term rice: The Computational Analysis In the previous section we studied the role of the context (on [hasproperty-edible, isa-starch, memberof-oryza, different levels) within the process of word similarity percep- atlocation-refrigerator, usedfor-survival, atlocation- tion. Since the results indicated that both functional aspects atgrocerystore, isa-food, isa-domesticateplant, and syntactic roles have an impact on how people perceive relatedto-grain, madeof-sake, isa-grain, receivesaction- similarity, we experimented a computational approach for the cook, atlocation-pantry, atlocation-ricecrisp, atlocation- automatic estimation of the similarity based on functional and supermarket, ...] syntax-aware contextual information. In particular, we used the large and freely-available seman- Then, the semantic information for the word bean are: tic resource ConceptNet8 . A partial overview of the semantic [usedfor-fillbeanbagchair, atlocation-infield, knowledge contained in ConceptNet is illustrated in Table 6. atlocation-can, usedfor-nutrition, usedfor-cook, ConceptNet is a resource based on common-sense rather than atlocation-atgrocerystore, usedfor-grow, atlocation- linguistic knowledge since it contains much more function- foodstore, isa-legume, usedfor-count, isa- based information (e.g., all the actions an object can or can- domesticateplant, atlocation-cookpot, atlocation- not do) contained in even complex syntactic structures. The beansoup, atlocation-soup, isa-vegetable, ...] idea is also to exploit users perception of reality (the actual origin of ConceptNet) instead of the result of top-down ex- Finally, the intersection produces the following set: pert building of ontologies (e.g., WordNet). ConceptNet con- tains important semantic problems related to covarage, utility [atlocation-atgrocerystore, isa-domesticateplant, at- of semantic information and coherence, but we used it as a locationpantry] black box due to its largeness and common-sense nature. A At this point, for each non-empty intersection, we created deep analysis of this resource is out of the scope of this paper. one instance of the type: The experiment started from the transformation of a word- word-score similarity dataset into a context-based dataset in , which the words are replaced by sets of semantic information taken from ConceptNet. The aim was to figure out which and computed a standard term-document matrix, where the semantic facts make the similarity between two words per- term is a semantic term within the set of semantic informa- ceivable. tion retrieved from ConceptNet and the document dimension We used the dataset SimLex-999 (Hill et al.,2014) that con- represents the word pairs of the original dataset. After this tains one thousand word pairs that were manually annotated preprocessing phase, the score attribute is discretized into two with similarity scores. The inter-annotation agreement is 0.67 bins: 8 http://conceptnet5.media.mit.edu/ • non-similar class - range in the dataset [0, 5] 485 • similar class - range in the dataset [5.1, 10] Richards, Malinowski, & Crookshank,1946) introduced by different authors over time represents a first reference for The splitting of the data into two clusters allowed us to our study. People use symbols (our words) to communicate experiment a classic supervised classification system, where meanings (the effective content). The meaning is something a Machine Learning tool (a Support Vector Machine, in our untangible, which can be though even without any concrete case) has been used to learn a binary model for automatically presence. The last point is then the physical reference, i.e., classifying similar and non-similar word pairs. The result of the object in the reality9 . Note that there is no connection be- the experiment is shown in Table 7. Noticeably, the classi- tween symbols and references, since only imagined meanings fier has been able to reach a quite good accuracy (65.38% can allow the two to be linked. of correctly classified word pairs), considering that the inter- Interaction is another important aspect that has been inves- annotation agreement of the original data is only 0.67 (Spear- tigated in literature. Indeed, the actions change the type of man correlation). perception of an object, which models itself to fit with the context of use. Then, the Gestalt theory (Köhler,1929) con- Table 7: Classification results in terms of Precision, Recall, tains different notions about the perception of meaning ac- and F-measure. The total accuracy is 65.38%. cording to interaction and context. In particular, the core of Precision Recall F-Measure Class the model is the complementarity between the figure and the 0,697 0,475 0,565 non-similar ground. In our case, a word is the figure and the ground is 0,633 0,815 0,713 similar the context that lets emerge its specific sense. Finally, James 0,664 0,654 0,643 weighted total Gibson introduced the concept of affordances as the cognitive cues that an object exposes to the external world, indicating ways of use (Gibson,1977). In cognitive and computational Notice that similar word pairs are generally easier to iden- linguistics, this theory can be inherited to model words as ob- tify with respect to non-similar ones. jects and contexts as their interaction with the world. Related Work Computational Background This paper presents an idea which combines linguistic, cogni- In this section, we review the main works that are related to tive and computational perspectives. In this section, we men- our contribution from a computational perspective. Natural tion those theoretical and empirical methods that inspired our Language Processing represents an active research commu- motivational basis. nity whose focus is letting machines communicate by under- standing semantics within linguistic expressions. Ontology Linguistic Background Learning (Cimiano,2006) is the task of automatic extracting The difficulty of defining the meaning of meaning has to do structured semantic knowledge from texts, and it well fits the with some tricky issues like lexical ambiguity and polysemy, scope of this paper. Nevertheless, Word Sense Disambigua- vagueness, contextual variability of word meaning, etc. As a tion (WSD) (Stevenson & Wilks,2003) is maybe the most re- matter of fact, words are organized in lexicon as a complex lated NLP task, whose aim is to capture the correct mean- network of semantic relation which are basically subsumed ing of a word in a context. Generally speaking, many other within Saussure’s paradigmatic (the axis of combination) and tasks have the problem of comparing linguistic items in order syntagmatic (the axis of choice) axes (Saussure,1983). to make choices to pass from syntax to semantics. Named Some authors (Chaffin & Herrmann,1984) have already Entity Recognition (NER) (Nadeau & Sekine,2007;Marrero, suggested theoretical and empirical taxonomies of semantic Urbano, Sánchez-Cuadrado, Morato, & Gómez-Berbı́s,2013) relations consisting of some main families of relation (such is the task of identifying entities like people, organizations as contrast, similars, class inclusion, part-whole, etc.). As and locations in texts. This is often done by comparing words Murphy points out (M. L. Murphy,2003), lexicon has become in contexts to some learned patterns. In general, many other more central in linguistic theories and, even if there is no a NLP tasks are based on the evaluation of similarity scores widely accepted theory on its internal semantic structure and (Manning & Schütze,1999). how lexical information are represented in it, the semantic re- Nowadays, there exists a large set of available semantic lations among words are considered in scholarly literature as resources that can be used in Natural Language Processing relevant to the structure of both lexical and conceptual infor- techniques in order to understand the hidden meaning of per- mation and it is generally believed that relations among words ceived similarity between two words or concepts. For exam- determine meaning. ple, ConceptNet contains semantic information that are usu- ally associated with common terms (even if not correctly dis- Cognitive Background ambiguated). By analyzing the relationship betweeen anno- Although words perception could seem immediate, the input tated similarity scores and semantic information it is possi- we perceive is recognized and trasformed mediating back- 9 The existing terminology is quite varying: symbol- ground and contextual information, within a dynamic and co- thought/reference/referent (Aristotele); object-representation- operative process. The well-known semiotic triangle (Ogden, interpretant (Peirce); signified-sign-referent (De Saussure) 486 ble to create predictive models which automatically deduce similarity scores. words similarity by dynamically weighting words features based on their mutual interaction. References If we consider the objects / agents / actions to be terms in Baroni, M., & Lenci, A. (2010). Distributional memory: A text sentences, we can try to extract their meaning and se- general framework for corpus-based semantics. Com- mantic constraints by using the idea of affordances. For in- putational Linguistics, 36(4), 673–721. stance, let us think to the sentence “The squirrel climbs the Bhattacharyya, P., & Khapra, M. (2012). Word sense disam- tree”. In this case, we need to know what kind of subject biguation. Emerging Applications of Natural Language “squirrel” is to figure out (and visually imagine) how the ac- Processing: Concepts and New Research, 22. tion will be performed. According to this, no particular issues Chaffin, R., & Herrmann, D. J. (1984). The similarity and come out from the reading of this sentence. Let us now con- diversity of semantic relations. Memory & Cognition, sider the sentence “The elephant climbs the tree”. Even if the 12(2), 134–141. grammatical structure of the sentence is the same as before, Cimiano, P. (2006). Ontology learning from text. Springer. the agent of the action is different, and it obviously creates Cohen, L., Lehéricy, S., Chochon, F., Lemer, C., Rivaud, S., some semantic problems. In fact, from this case, some con- & Dehaene, S. (2002). Language-specific tuning of straints arise; in order to climb a tree, the subject needs to visual cortex? functional properties of the visual word fit to our mental model of something that can climb a tree. form area. Brain, 125(5), 1054–1069. In addition, this also depends on the mental model of “tree”. Finkelstein, L., Gabrilovich, E., Matias, Y., Rivlin, E., Solan, Moreover, different agents can be both correct subjects of an Z., Wolfman, G., et al. (2001). Placing search in con- action whilst they may produce different meanings in terms text: The concept revisited. In Proceedings of the 10th of how the action will be mentally performed. Consider the international conference on world wide web (pp. 406– sentences “The cat opens the door” and “The man opens the 414). door”. In both cases, some implicit knowledge suggests the Gibson, J. J. (1977). The Theory of Affordances. In R. Shaw manner the action is done: while in the second case we may & J. Bransford (Eds.), Perceiving, acting, and knowing: think at the cat that opens the door leaning to it, in the case Toward an ecological psychology. Lawrence Erlbaum. of the man we probably imagine the use of a door handle. A Hill, F., Reichart, R., & Korhonen, A. (2014). Simlex-999: study of these language dynamics can be of help for many Evaluating semantic models with (genuine) similarity NLP tasks like Part-Of-Speech tagging as well as more com- estimation. arXiv preprint arXiv:1408.3456. plex operations like dependency parsing and semantic rela- Köhler, W. (1929). Gestalt psychology. tions extraction. Some of these concepts are latently stud- Manning, C. D., & Schütze, H. (1999). Foundations of sta- ied in different disciplines related to statistics. Distributional tistical natural language processing. MIT press. Semantics (DS) (Baroni & Lenci,2010) represents a class of Marrero, M., Urbano, J., Sánchez-Cuadrado, S., Morato, J., statistical and linguistic analysis of text corpora that tries to & Gómez-Berbı́s, J. M. (2013). Named entity recog- estimate the validity of connections between subjects, verbs, nition: Fallacies, challenges and opportunities. Com- and objects by means of statistical sources of significance. puter Standards & Interfaces, 35(5), 482–489. Murphy, G. L., & Medin, D. L. (1985). The role of theories Conclusions and Future Work in conceptual coherence. Psychological review, 92(3), 289. In this paper, we proposed a combined analysis of linguis- Murphy, M. L. (2003). Semantic relations and the lexicon. tic, cognitive and computational aspects to assess the nature Cambridge University Press. of words similarity. First, we studied how word similarity Nadeau, D., & Sekine, S. (2007). A survey of named entity perception is influenced in terms of conceptual, functional recognition and classification. Lingvisticae Investiga- and syntactic roles. In future work, our aim is to extend the tiones, 30(1), 3–26. sample of users on more specific cases. Still, by changing Ogden, C. K., Richards, I. A., Malinowski, B., & Crook- the language of the users, we can have results that take into shank, F. G. (1946). The meaning of meaning. Har- account the cultural ground, understanding if and how word court, Brace & World New York. similarity depends on it. On the other side, we stressed the Proffitt, D. (2006). Embodied perception and the economy importance of computational understanding of similarity to of action. Perspectives on psychological science, 1(2), improve Computational Linguistics tasks which are based on 110–122. it, usually without any analysis of contextual information. In Saussure, F. d. (1983). Course in general linguistics, trans. particular, we used the large semantic knowledge contained R. Harris, London: Duckworth. in ConceptNet to create a Support Vector Machine classifier Stevenson, M., & Wilks, Y. (2003). Word-sense disambigua- to predict word similarity based on an annotated dataset. In tion. The Oxford Handbook of Comp. Linguistics, 249– future work, we will extend our experimental analysis to val- 265. idate existing similarity datasets and to produce predictive models for the automatic identification of human-readable 487