Its the Meaning That Counts: The State of the Art in NLP and Semantics KI Künstliche Intelligenz

To accomplish that, a human judgment task was set up and the judges were presented with a sentence and the entities in that sentence for which Lexis had predicted a CREATED, DESTROYED, or MOVED state change, along with the locus of state change. The results were compared against the ground truth of the ProPara test data. If a prediction was incorrectly counted as a false positive, i.e., if the human judges counted the Lexis prediction as correct but it was not labeled in ProPara, the data point was ignored in the evaluation in the relaxed setting. This also eliminates the need for the second-order logic of start(E), during(E), and end(E), allowing for more nuanced temporal relationships between subevents. The default assumption in this new schema is that e1 precedes e2, which precedes e3, and so on.

In this article, we describe new, hand-crafted semantic representations for the lexical resource VerbNet that draw heavily on the linguistic theories about subevent semantics in the Generative Lexicon (GL). VerbNet defines classes of verbs based on both their semantic and syntactic similarities, paying particular attention to shared diathesis alternations. For each class of verbs, VerbNet provides common semantic roles and typical syntactic patterns. For each syntactic pattern in a class, VerbNet defines a detailed semantic representation that traces the event participants from their initial states, through any changes and into their resulting states. We applied that model to VerbNet semantic representations, using a class’s semantic roles and a set of predicates defined across classes as components in each subevent.

Machine Translation and Attention

The data presented in Table 2 elucidates that the semantic congruence between sentence pairs primarily resides within the 80–90% range, totaling 5,507 such instances. Moreover, the pairs of sentences with a semantic similarity exceeding 80% (within the 80–100% range) are counted as 6,927 pairs, approximately constituting 78% of the total amount of sentence pairs. This forms the major component of all results in the semantic similarity calculations. Most of the semantic similarity between the sentences of the five translators is more than 80%, this demonstrates that the main body of the five translations captures the semantics of the original Analects quite well.

Figure 1 shows an example of a sentence with 4 targets, denoted by highlighted words and sequence of words. Each of these targets will correspond directly with a frame PERFORMERS_AND_ROLES, IMPORTANCE, THWARTING, BECOMING_DRY frames, annotated by categories with boxes. You will notice that sword is a “weapon” and her (which can be co-referenced to Cyra) is a “wielder”.

Stemming

These representations show the relationships between arguments in a sentence, including peripheral roles like Time and Location, but do not make explicit any sequence of subevents or changes in participants across the timespan of the event. VerbNet’s explicit subevent sequences allow the extraction of preconditions and postconditions for many of the verbs in the resource and the tracking of any changes to participants. In addition, VerbNet allow users to abstract away from individual verbs to more general categories of eventualities. We believe VerbNet is unique in its integration of semantic roles, syntactic patterns, and first-order-logic representations for wide-coverage classes of verbs. Natural language processing and Semantic Web technologies have different, but complementary roles in data management.

These recurrent words in The Analects include key cultural concepts such as “君子 Jun Zi, 小人 Xiao Ren, 仁 Ren, 道 Dao, 礼 Li,” and others (Li et al., 2022). A comparison of sentence pairs with a semantic similarity of ≤ 80% reveals that these core conceptual words significantly influence the semantic variations among the translations of The Analects. The second category includes various personal names mentioned in The Analects. Our analysis suggests that the distinct translation methods of the five translators for these names significantly contribute to the observed semantic differences, likely stemming from different interpretation or localization strategies. Out of the entire corpus, 1,940 sentence pairs exhibit a semantic similarity of ≤ 80%, comprising 21.8% of the total sentence pairs. These low-similarity sentence pairs play a significant role in determining the overall similarity between the different translations.

• Verb-specific features incorporated in the semantic representations where possible. Have you ever misunderstood a sentence you’ve read and had to read it all over again? Have you ever heard a jargon term or slang phrase and had no idea what it meant? Understanding what people are saying can be difficult even for us homo sapiens. Clearly, making sense of human language is a legitimately hard problem for computers. Natural language processing (NLP) and Semantic Web technologies are both Semantic Technologies, but with different and complementary roles in data management.

Understanding that the statement ‘John dried the clothes’ entailed that the clothes began in a wet state would require that systems infer the initial state of the clothes from our representation. By including that initial state in the representation explicitly, we eliminate the need for real-world knowledge or inference, an NLU task that is notoriously difficult. In order to accommodate such inferences, the event itself needs to have substructure, a topic we now turn to in the next section. In the rest of this article, we review the relevant background on Generative Lexicon (GL) and VerbNet, and explain our method for using GL’s theory of subevent structure to improve VerbNet’s semantic representations. We show examples of the resulting representations and explain the expressiveness of their components. Finally, we describe some recent studies that made use of the new representations to accomplish tasks in the area of computational semantics.

However, it is crucial to note that these subdivisions were not exclusively reliant on punctuation marks. Instead, this study followed the principle of dividing the text into lines to make sure that each segment fully expresses the original meaning. Finally, each nlp semantic translated English text was aligned with its corresponding original text. Deep-learning models take as input a word embedding and, at each time state, return the probability distribution of the next word as the probability for every word in the dictionary.

They further provide valuable insights into the characteristics of different translations and aid in identifying potential errors. By delving deeper into the reasons behind this substantial difference in semantic similarity, this study can enable readers to gain a better understanding of the text of The Analects. Furthermore, this analysis can guide translators in selecting words more judiciously for crucial core conceptual words during the translation process. The Escape-51.1 class is a typical change of location class, with member verbs like depart, arrive and flee. The most basic change of location semantic representation (12) begins with a state predicate has_location, with a subevent argument e1, a Theme argument for the object in motion, and an Initial_location argument. The motion predicate (subevent argument e2) is underspecified as to the manner of motion in order to be applicable to all 40 verbs in the class, although it always indicates translocative motion.

How is Semantic Analysis different from Lexical Analysis?

We also defined our event variable e and the variations that expressed aspect and temporal sequencing. At this point, we only worked with the most prototypical examples of changes of location, state and possession and that involved a minimum of participants, usually Agents, Patients, and Themes. For readers, the core concepts in The Analects transcend the meaning of single words or phrases; they encapsulate profound cultural connotations that demand thorough and precise explanations. For instance, whether “君子 Jun Zi” is translated as “superior man,” “gentleman,” or otherwise. It is nearly impossible to study Confucius’s thought without becoming familiar with a few core concepts (LaFleur, 2016), comprehending the meaning is a prerequisite for readers. Various forms of names, such as “formal name,” “style name,” “nicknames,” and “aliases,” have deep roots in traditional Chinese culture.

• Within the similarity score intervals of 80–85% and 85–90%, the distributions of sentences across all five translators is more balanced, each accounting for about 20%.
• Other classes, such as Other Change of State-45.4, contain widely diverse member verbs (e.g., dry, gentrify, renew, whiten).
• A higher value on the y-axis indicates a higher degree of semantic similarity between sentence pairs.
• These low-similarity sentence pairs play a significant role in determining the overall similarity between the different translations.
• Both resources define semantic roles for these verb groupings, with VerbNet roles being fewer, more coarse-grained, and restricted to central participants in the events.

We cover how to build state-of-the-art language models covering semantic similarity, multilingual embeddings, unsupervised training, and more. Learn how to apply these in the real world, where we often lack suitable datasets or masses of computing power. Therefore, in semantic analysis with machine learning, computers use Word Sense Disambiguation to determine which meaning is correct in the given context. Consider the task of text summarization which is used to create digestible chunks of information from large quantities of text. Text summarization extracts words, phrases, and sentences to form a text summary that can be more easily consumed. The accuracy of the summary depends on a machine’s ability to understand language data.

Whether translations adopt a simplified or literal approach, readers stand to benefit from understanding the structure and significance of ancient Chinese names prior to engaging with the text. Most proficient translators typically include detailed explanations of these core concepts and personal names either in the introductory or supplementary sections of their translations. If feasible, readers should consult multiple translations for cross-reference, especially when interpreting key conceptual terms and names. However, given the abundance of online resources, sourcing accurate and relevant information is convenient. Readers can refer to online resources like Wikipedia or academic databases such as the Web of Science. While this process may be time-consuming, it is an essential step towards improving comprehension of The Analects.

Table 8a, b display the high-frequency words and phrases observed in sentence pairs with semantic similarity scores below 80%, after comparing the results from the five translations. This set of words, such as “gentleman” and “virtue,” can convey specific meanings independently. An error analysis of the results indicated that world knowledge and common sense reasoning were the main sources of error, where Lexis failed to predict entity state changes.

The first major change to this representation was that path_rel was replaced by a series of more specific predicates depending on what kind of change was underway. These slots are invariable across classes and the two participant arguments are now able to take any thematic role that appears in the syntactic representation or is implicitly understood, which makes the equals predicate redundant. It is now much easier to track the progress of a single entity across subevents and to understand who is initiating change in a change predicate, especially in cases where the entity called Agent is not listed first.

• Now, we have a brief idea of meaning representation that shows how to put together the building blocks of semantic systems.
• We are encouraged by the efficacy of the semantic representations in tracking entity changes in state and location.
• Thus, machines tend to represent the text in specific formats in order to interpret its meaning.

As we worked toward a better and more consistent distribution of predicates across classes, we found that new predicate additions increased the potential for expressiveness and connectivity between classes. We also replaced many predicates that had only been used in a single class. In this section, we demonstrate how the new predicates are structured and how they combine into a better, more nuanced, and more useful resource. For a complete list of predicates, their arguments, and their definitions (see Appendix A). Early rule-based systems that depended on linguistic knowledge showed promise in highly constrained domains and tasks. Machine learning side-stepped the rules and made great progress on foundational NLP tasks such as syntactic parsing.