Natural language generators for task-oriented dialog should be able to vary the style of the output utterance while still effectively realizing the system dialog actions and their associated semantics. While the use of neural generation for training the response generation component of conversational agents promises to simplify the process of producing high quality responses in new domains, to our knowledge, there has been very little investigation of neural generators for task-oriented dialog that can vary their response style and we know of no experiments on models that can generate responses that are different in style from those seen during training, while still maintaining semantic fidelity to the input meaning representation. Here, we show that a model that is trained to achieve a single stylistic personality target can produce outputs that combine stylistic targets. We carefully evaluate the multivoice outputs for both semantic fidelity and for similarities to and differences from the linguistic features that characterize the original training style. We show that contrary to our predictions, the learned models do not always simply interpolate model parameters, but rather produce styles that are distinct and novel from the personalities they were trained on.
more »
« less
Controlling Personality-Based Stylistic Variation with Neural Natural Language Generators
Natural language generators for taskoriented
dialogue must effectively realize
system dialogue actions and their associated
semantics. In many applications,
it is also desirable for generators to control
the style of an utterance. To date,
work on task-oriented neural generation
has primarily focused on semantic fidelity
rather than achieving stylistic goals, while
work on style has been done in contexts
where it is difficult to measure content
preservation. Here we present three different
sequence-to-sequence models and
carefully test how well they disentangle
content and style. We use a statistical generator,
PERSONAGE, to synthesize a new
corpus of over 88,000 restaurant domain
utterances whose style varies according to
models of personality, giving us total control
over both the semantic content and the
stylistic variation in the training data. We
then vary the amount of explicit stylistic
supervision given to the three models. We
show that our most explicit model can simultaneously
achieve high fidelity to both
semantic and stylistic goals: this model
adds a context vector of 36 stylistic parameters
as input to the hidden state of the encoder
at each time step, showing the benefits
of explicit stylistic supervision, even
when the amount of training data is large.
more »
« less
- Award ID(s):
- 1748056
- NSF-PAR ID:
- 10079406
- Date Published:
- Journal Name:
- Proceedings of the annual SIGdial Meeting on Discourse and Dialogue (SIGDIAL)}
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Responses in task-oriented dialogue systems often realize multiple propositions whose ultimate form depends on the use of sentence planning and discourse structuring operations. For example a recommendation may consist of an explicitly evaluative utterance e.g. Chanpen Thai is the best option, along with content related by the justification discourse relation, e.g. It has great food and service, that combines multiple propositions into a single phrase. While neural generation methods integrate sentence planning and surface realization in one endto-end learning framework, previous work has not shown that neural generators can: (1) perform common sentence planning and discourse structuring operations; (2) make decisions as to whether to realize content in a single sentence or over multiple sentences; (3) generalize sentence planning and discourse relation operations beyond what was seen in training. We systematically create large training corpora that exhibit particular sentence planning operations and then test neural models to see what they learn. We compare models without explicit latent variables for sentence planning with ones that provide explicit supervision during training. We show that only the models with additional supervision can reproduce sentence planning and discourse operations and generalize to situations unseen in training.more » « less
-
null (Ed.)Neural natural language generation (NNLG) from structured meaning representations has become increasingly popular in recent years. While we have seen progress with generating syntactically correct utterances that preserve semantics, various shortcomings of NNLG systems are clear: new tasks require new training data which is not available or straightforward to acquire, and model outputs are simple and may be dull and repetitive. This paper addresses these two critical challenges in NNLG by: (1) scalably (and at no cost) creating training datasets of parallel meaning representations and reference texts with rich style markup by using data from freely available and naturally descriptive user reviews, and (2) systematically exploring how the style markup enables joint control of semantic and stylistic aspects of neural model output. We present YelpNLG, a corpus of 300,000 rich, parallel meaning representations and highly stylistically varied reference texts spanning different restaurant attributes, and describe a novel methodology that can be scalably reused to generate NLG datasets for other domains. The experiments show that the models control important aspects, including lexical choice of adjectives, output length, and sentiment, allowing the models to successfully hit multiple style targets without sacrificing semantics.more » « less
-
null (Ed.)The marine-based West Antarctic Ice Sheet (WAIS) is currently retreating due to shifting wind-driven oceanic currents that transport warm waters toward the ice margin, resulting in ice shelf thinning and accelerated mass loss of the WAIS. Previous results from geologic drilling on Antarctica’s continental margins show significant variability in marine-based ice sheet extent during the late Neogene and Quaternary. Numerical models indicate a fundamental role for oceanic heat in controlling this variability over at least the past 20 My. Although evidence for past ice sheet variability has been collected in marginal settings, sedimentologic sequences from the outer continental shelf are required to evaluate the extent of past ice sheet variability and the associated oceanic forcings and feedbacks. International Ocean Discovery Program Expedition 374 drilled a latitudinal and depth transect of five drill sites from the outer continental shelf to rise in the eastern Ross Sea to resolve the relationship between climatic and oceanic change and WAIS evolution through the Neogene and Quaternary. This location was selected because numerical ice sheet models indicate that this sector of Antarctica is highly sensitive to changes in ocean heat flux. The expedition was designed for optimal data-model integration and will enable an improved understanding of the sensitivity of Antarctic Ice Sheet (AIS) mass balance during warmer-than-present climates (e.g., the Pleistocene “super interglacials,” the mid-Pliocene, and the late early to middle Miocene). The principal goals of Expedition 374 were to • Evaluate the contribution of West Antarctica to far-field ice volume and sea level estimates; • Reconstruct ice-proximal atmospheric and oceanic temperatures to identify past polar amplification and assess its forcings and feedbacks; • Assess the role of oceanic forcing (e.g., sea level and temperature) on AIS stability/instability; • Identify the sensitivity of the AIS to Earth’s orbital configuration under a variety of climate boundary conditions; and • Reconstruct eastern Ross Sea paleobathymetry to examine relationships between seafloor geometry, ice sheet stability/instability, and global climate. To achieve these objectives, we will • Use data and models to reconcile intervals of maximum Neogene and Quaternary Antarctic ice advance with far-field records of eustatic sea level change; • Reconstruct past changes in oceanic and atmospheric temperatures using a multiproxy approach; • Reconstruct Neogene and Quaternary sea ice margin fluctuations in datable marine continental slope and rise records and correlate these records to existing inner continental shelf records; • Examine relationships among WAIS stability/instability, Earth’s orbital configuration, oceanic temperature and circulation, and atmospheric pCO2; and • Constrain the timing of Ross Sea continental shelf overdeepening and assess its impact on Neogene and Quaternary ice dynamics. Expedition 374 was carried out from January to March 2018, departing from Lyttelton, New Zealand. We recovered 1292.70 m of high-quality cores from five sites spanning the early Miocene to late Quaternary. Three sites were cored on the continental shelf (Sites U1521, U1522, and U1523). At Site U1521, we cored a 650 m thick sequence of interbedded diamictite, mudstone, and diatomite, penetrating the Ross Sea seismic Unconformity RSU4. The depositional reconstructions of past glacial and open-marine conditions at this site will provide unprecedented insight into environmental change on the Antarctic continental shelf during the early and middle Miocene. At Site U1522, we cored a discontinuous upper Miocene to Pleistocene sequence of glacial and glaciomarine strata from the outer shelf, with the primary objective to penetrate and date seismic Unconformity RSU3, which is interpreted to represent the first major continental shelf–wide expansion and coalescing of marine-based ice streams from both East and West Antarctica. At Site U1523, we cored a sediment drift located beneath the westerly flowing Antarctic Slope Current (ASC). Cores from this site will provide a record of the changing vigor of the ASC through time. Such a reconstruction will enable testing of the hypothesis that changes in the vigor of the ASC represent a key control on regulating heat flux onto the continental shelf, resulting in the ASC playing a fundamental role in ice sheet mass balance. We also cored two sites on the continental slope and rise. At Site U1524, we cored a Plio–Pleistocene sedimentary sequence on the continental rise on the levee of the Hillary Canyon, which is one of the largest conduits of Antarctic Bottom Water delivery from the Antarctic continental shelf into the abyssal ocean. Drilling at Site U1524 was intended to penetrate into middle Miocene and older strata but was initially interrupted by drifting sea ice that forced us to abandon coring in Hole U1524A at 399.5 m drilling depth below seafloor (DSF). We moved to a nearby alternate site on the continental slope (U1525) to core a single hole with a record complementary to the upper part of the section recovered at Site U1524. We returned to Site U1524 3 days later, after the sea ice cleared. We then cored Hole U1524C with the rotary core barrel with the intention of reaching the target depth of 1000 m DSF. However, we were forced to terminate Hole U1524C at 441.9 m DSF due to a mechanical failure with the vessel that resulted in termination of all drilling operations and a return to Lyttelton 16 days earlier than scheduled. The loss of 39% of our operational days significantly impacted our ability to achieve all Expedition 374 objectives as originally planned. In particular, we were not able to obtain the deeper time record of the middle Miocene on the continental rise or abyssal sequences that would have provided a continuous and contemporaneous archive to the high-quality (but discontinuous) record from Site U1521 on the continental shelf. The mechanical failure also meant we could not recover sediment cores from proposed Site RSCR-19A, which was targeted to obtain a high-fidelity, continuous record of upper Neogene and Quaternary pelagic/hemipelagic sedimentation. Despite our failure to recover a shelf-to-rise transect for the Miocene, a continental shelf-to-rise transect for the Pliocene to Pleistocene interval is possible through comparison of the high-quality records from Site U1522 with those from Site U1525 and legacy cores from the Antarctic Geological Drilling Project (ANDRILL).more » « less
-
Example-guided image synthesis has been recently attempted to synthesize an image from a semantic label map and an exemplary image. In the task, the additional exemplary image serves to provide style guidance that controls the appearance of the synthesized output. Despite the controllability advantage, the previous models are designed on datasets with specific and roughly aligned objects. In this paper, we tackle a more challenging and general task, where the exemplar is an arbitrary scene image that is semantically unaligned to the given label map. To this end, we first propose a new Masked Spatial-Channel Attention (MSCA) module which models the correspondence between two unstructured scenes via cross-attention. Next, we propose an end-to-end network for joint global and local feature alignment and synthesis. In addition, we propose a novel patch-based self-supervision scheme to enable training. Experiments on the large-scale CCOO-stuff dataset show significant improvements over existing methods. Moreover, our approach provides interpretability and can be readily extended to other tasks including style and spatial interpolation or extrapolation, as well as other content manipulation.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
