We present the contribution of Universitat Pompeu Fabra's NLP group to the SemEval Task 9.2 (AMR-to-English Generation). The proposed generation pipeline comprises: (i) a series of rule-based graphtransducers for the syntacticization of the input graphs and the resolution of morphological agreements, and (ii) an off-theshelf statistical linearization component.

State-of-the-art prosody modelling in content-to-speech (CTS) applications still uses the same methodology to predict intonation cues as text-to-speech (TTS) applications, namely the analysis of the generated surface sentences with respect to part of speech, syntactic dependency relations and word order. On the other side, several theoretical studies argue that morphology, syntax, and information (or communicative) structure that organizes a given content (semantic or deep-syntactic structure) with respect to the intention of the speaker show a strong correlation with intonation. However, little empirical work based on sufficiently large corpora has been carried out so far to buttress this argumentation. We present empirical evidence for the Information Structure-Prosody correlation using the Wall Street Journal Penn Treebank corpus recorded by native American English speakers. Our experiments reach a prosody prediction accuracy of 80% using the hierarchical information structure from the Meaning-Text Theory, compared to 59% of the baseline.

Patent search is recall-driven, which goes hand in hand with at least a partial sacrifice of precision. As a consequence, patent analysts have to regularly view and examine a large amount of patents. This implies a very high workload. Interactive analysis aids that help to minimize this workload are thus of high demand. Still, these aids do not reduce the amount of the material to be examined, they only facilitate its examination. Its reduction can be achieved working with patent summaries instead of full patent documents. So far, high quality patent summaries are produced mainly manually and only a few research works address the problem of automatic patent summarization. Most often, these works either replicate the summarization metrics known from general discourse summarization or focus on the claims of a patent. However, it can be observed that neither of the strategies is adequate: general discourse state-of-the-art summarization techniques are of limited use due to the idiosyncrasies of the patent genre, and techniques that focus on claims only miss in their summaries important details provided in the other sections on the components of the invention introduced in the claims. We propose a patent summarization technique that takes the idiosyncrasies of the patent genre (such as the unbalanced distribution of the content across the different sections of a patent, excessive length of the sentences in the claims, abstract vocabuly, etc.) into account to obtain a comprehensive summary of the invention. In particular, we make use of lexical chains in the claims and in the description of the invention and of aligned claim-description segments at the subsentential level to assess the relevance of the individual fragments of the document for the summary. The most relevant fragments are selected and merged using full-fledged natural language generation techniques.

No abstract

Conversational interfaces involving text-to-speech (TTS) applications have improved expressiveness and overall naturalness to a reasonable extent in the last decades. Conversational features, such as speech acts, affective states and information structure have been instrumental to derive more expressive prosodic contours. However, synthetic speech is still perceived as monotonous, when a text that lacks those conversational features is read aloud in the interface, i.e. it is fed directly to the TTS application. In this paper, we propose a methodology for pre-processing raw texts before they arrive to the TTS application. The aim is to analyze syntactic and information (or communicative) structure, and then use the high-level linguistic features derived from the analysis to generate more expressive prosody in the synthesized speech. The proposed methodology encompasses a pipeline of four modules: (1) a tokenizer, (2) a syntactic parser, (3) a communicative parser, and (3) an SSML prosody tag converter. The implementation has been tested in an experimental setting for German, using web-retrieved articles.